IBM Cloud Docs
Release notes for Speech to Text for IBM Cloud

Release notes for Speech to Text for IBM Cloud

IBM Cloud

The following features and changes were included for each release and update of managed instances of IBM Watson® Speech to Text that are hosted on IBM Cloud or for instances that are hosted on IBM Cloud Pak for Data as a Service. Unless otherwise noted, all changes are compatible with earlier releases and are automatically and transparently available to all new and existing applications.

For information about known limitations of the service, see Known limitations.

For information about releases and updates of the service for IBM Cloud Pak for Data, see Release notes for Speech to Text for IBM Cloud Pak for Data.

19 November 2024

New large speech model for German is now generally available

The large speech model for German is now generally available.

23 August 2024

All Large Speech Models are now generally available

The large speech models for all languages are now generally available (GA). They are supported for use in production environments and applications.

18 June 2024

New large speech models for Brazilian Portuguese and Spanish are now in open beta

The large speech models for Brazilian Portuguese and Spanish are now in open beta. Spanish includes the Castilian, Argentinian, Chilean, Colombian, Mexican, and Peruvian dialects.

15 May 2024

Large Speech Model for English is now generally available

The large speech model for English, which includes the United States, Australian, Indian, and United Kingdom dialects, is now generally available (GA). It is supported for use in production environments and applications.

07 March 2024

Large Speech Model for US English in Open Beta
The new Large speech model for US English is in open beta. See Large speech languages and models for more details with supported features (beta).

30 November 2023

Speech to Text parameter speech_begin_event

This parameter would enable the client application to know that some words or speech is detected and Speech to Text is in the process of decoding. For more details, see Using speech recognition parameters.

Parameter 'mapping_only' for custom words

By using the 'mapping_only' parameter, you can use custom words directly to map 'sounds_like' (or word) to 'display_as' value as post-processing instead of training. For more information, see The words resource.

See the guidelines for Non-Japanese and Japanese.

Support for Brazilian-Portuguese and French-Canadian on new improved next-generation language model customization

Language model customization for Brazilian-Portuguese and French-Canadian next-generation models is recently added. This service update includes further internal improvements.

New Smart Formatting Feature

A new smart formatting feature for next-generation models is supported in US English, Brazilian Portuguese, French and German languages. See Smart formatting Version for details.

Support for Castilian Spanish and LATAM Spanish on new improved next-generation language model customization

The language model customization for Castilian Spanish and LATAM Spanish next-generation models are added. This service update includes further internal improvements.

Large Speech Models for English, Japanese, and French - for early access

For early access feature, Large Speech Models are available for English, Japanese and French languages for you in IBM Watson Speech-to-Text and IBM watsonx Assistant. The feature set for these Large Speech Models is limited, but more accurate than Next-Generation models and are faster and cheaper to run due to smaller size and better streaming mode capability.

If you are interested in testing these base models, and sharing results and feedback, contact our Product Management team by filling out this form.

28 July 2023

Important: All previous-generation models are discontinued starting August 1, 2023
Important: All previous-generation models are now discontinued from the service. New clients must now only use the next-generation models. All existing clients must now migrate to the equivalent next-generation model. For more information about all next-generation models, see Next-generation languages and models. For more information on how to migrate to the next-generation models, see Migrating to next-generation models.

9 June 2023

Defect fix: Creating and training a custom Language Model is now optimal for both standard and low-latency Next-Generation models
Defect fix: When creating and training a custom Language Model with corpora text files and / or custom words using a Next-generation low-latency model, it is now performing the same way as with a standard model. Previously, it was not optimal only when using a Next-Generation low-latency model.
Defect fix: STT Websockets sessions no longer fail due to tensor error message
Defect fix: When using STT websockets, sessions no longer fail due to an error message “STT returns the error: Sizes of tensors must match except in dimension 0”.

18 May 2023

Updates to English next-generation Medical telephony model

The English next-generation Medical telephony model has been updated for improved speech recognition:

  • en-WW_Medical_Telephony
Added support for French and German on new improved next-generation language model customization

Language model customization for French and German next-generation models was recently added. This service update includes further internal improvements.

For more information about improved next-generation customization, see

Defect fix: Custom words containing half-width Katakana characters now return a clear error message with Japanese Telephony model

Defect fix: Per the documentation, only full-width Katakana characters are accepted in custom words and the next generation models now show an error message to explain that it's not supported. Previously, when creating custom words containing half-width Katakana characters, no error message was provided.

Defect fix: Japanese Telephony language model no longer fails due to long training time

Defect fix: When training a custom language model with Japanese Telephony, the service now effectively handles large numbers of custom words without failing.

2 May 2023

New procedure for upgrading a custom model that is based on an improved next-generation model

Two approaches are now available to upgrade a custom language model to an improved next-generation base model. You can still modify and then retrain the custom model, as already documented. But now, you can also upgrade the custom model by including the query parameter force=true with the POST /v1/customizations/{customization_id}/train request. The force parameter upgrades the custom model regardless of whether it contains changes (is in the ready or available state).

For more information, see Upgrading a custom language model based on an improved next-generation model.

Guidance for adding words to custom models that are based on improved next-generation models

The documentation now offers more guidance about adding words to custom models that are based on improved next-generation models. For performance reasons during training, the guidance encourages the use of corpora rather than the direct addition of custom words whenever possible.

For more information, see Guidelines for adding words to custom models based on improved next-generation models.

Japanese custom words for custom models that are based on improved next-generation models are handled differently

For Japanese custom models that are based on next-generation models, custom words are handled differently from other languages. For Japanese, you can add a custom word or sounds-like that does not exceed 25 characters in length. If your custom word or sounds-like exceeds that limit, the service adds the word to the custom model as if it were added by a corpus. The word does not appear as a custom word for the model.

For more information, see Guidelines for adding words to Japanese models based on improved next-generation models.

12 April 2023

Defect fix: The WebSocket interface now times out as expected when using next-generation models
Defect fix: When used for speech recognition with next-generation models, the WebSocket interface now times out as expected after long periods of silence. Previously, when used for speech recognition of short audio files, the WebSocket session could fail to time out. When the session failed to time out, the service did not return a final hypothesis to the waiting client application, and the client instead timed while waiting for the results.

6 April 2023

Defect fix: Limits to allow completion of training for next-generation Japanese custom models

Defect fix: Successful training of a next-generation Japanese custom language model requires that custom words and sounds-likes added to the model each contain no more than 25 characters. For the most effective training, it is recommended that custom words and sounds-likes contain no more than 20 characters. Training of Japanese custom models with longer custom words and sounds-likes fails to complete after multiple hours of training.

If you need to add the equivalent of a long word or sounds-like to a next-generation Japanese custom model, take these steps:

  1. Add a shorter word or sounds-like that captures the essence of the longer word or sounds-like to the custom model.
  2. Add one or more sentences that use the longer word or sounds-like to a corpus.
  3. Consider adding sentences to the corpus that provide more context for the word or sounds-like. Greater context gives the service more information with which to recognize the word and apply the correct sounds-like.
  4. Add the corpus to the custom model.
  5. Retrain the custom model on the combination of the shorter word or sounds-like and the corpus that contains the longer string.

The limits and steps just described allow next-generation Japanese custom models to complete training. Keep in mind that adding large numbers of new custom words to a custom language model increases the training time of the model. But the increased training time occurs only when the custom model is initially trained on the new words. Once the custom model has been trained on the new words, training time returns to normal.

For more information, see

Further improvements to updated next-generation language model customization

Language model customization for English and Japanese next-generation models was recently improved. This service update includes further internal improvements. For more information about improved next-generation customization, see

13 March 2023

Defect fix: Smart formatting for US English dates is now correct
Defect fix: Smart formatting now correctly includes days of the week and dates when both are present in the spoken audio, for example, Tuesday February 28. Previously, in some cases the day of the week was omitted and the date was presented incorrectly. Note that smart formatting is beta functionality.
Defect fix: Update documentation for speech hesitation words for next-generation models
Defect fix: Documentation for speech hesitation words for next-generation models has been updated. More details are provided about US English and Japanese hesitation words. Next-generation models include the actual hesitation words in transcription results, unlike previous-generation models, which include only hesitation markers. For more information, see Speech hesitations and hesitation markers.

27 February 2023

New Japanese next-generation telephony model

The service now offers a next-generation telephony model for Japanese: ja-JP_Telephony. The new model supports low latency and is generally available. It also supports language model customization and grammars. For more information about next-generation models and low latency, see

Improved language model customization for next-generation English and Japanese models

The service now provides improved language model customization for next-generation English and Japanese models:

  • en-AU_Multimedia
  • en-AU_Telephony
  • en-IN_Telephony
  • en-GB_Multimedia
  • en-GB_Telephony
  • en-US_Multimedia
  • en-US_Telephony
  • ja-JP_Multimedia
  • ja-JP_Telephony

Visible improvements to the models: The new technology improves the default behavior of the new English and Japanese models. Among other changes, the new technology optimizes the default behavior for the following parameters:

  • The default customization_weight for custom models that are based on the new versions of these models changes from 0.2 to 0.1.
  • The default character_insertion_bias for custom models that are based on the new versions of these models remains 0.0, but the models have changed in a manner that makes use of the parameter for speech recognition less necessary.

Upgrading to the new models: To take advantage of the improved technology, you must upgrade any custom language models that are based on the new models. To upgrade to the new version of one of these base models, do the following:

  1. Change your custom model by adding or modifying a custom word, corpus, or grammar that the model contains. Any change that you make moves the model to the ready state.

  2. Use the POST /v1/customizations/{customization_id}/train method to retrain the model. Retraining upgrades the custom model to the new technology and moves the model to the available state.

    Known issue: At this time, you cannot use the POST /v1/customizations/{customization_id}/upgrade_model method to upgrade a custom model to one of the new base models. This issue will be addressed in a future release.

Using the new models: Following the upgrade to the new base model, you are advised to evaluate the performance of the upgraded custom model by paying special attention to the customization_weight and character_insertion_bias parameters for speech recognition. When you retrain your custom model:

  • The custom model uses the new default customization_weight of 0.1 for your custom model. A non-default customization_weight that you had associated with your custom model is removed.
  • The custom model might no longer require use of the character_insertion_bias parameter for optimal speech recognition.

Improvements to language model customization render these parameters less important for high-quality speech recognition:

  • If you use the default values for these parameters, continue to do so after the upgrade. The default values will likely continue to offer the best results for speech recognition.
  • If you specify non-default values for these parameters, experiment with the default values following upgrade. Your custom model might work well for speech recognition with the default values.

If you feel that using different values for these parameters might improve speech recognition with your custom model, experiment with incremental changes to determine whether the parameters are needed to improve speech recognition.

Note: At this time, the improvements to language model customization apply only to custom models that are based on the next-generation English or Japanese base language models listed earlier. Over time, the improvements will be made available for other next-generation language models.

More information: For more information about upgrading and about speech recognition with these parameters, see

Defect fix: Grammar files now handle strings of digits correctly

Defect fix: When grammars are used, the service now handles longer strings of digits correctly. Previously, it was failing to complete recognition or returning incorrect results.

15 February 2023

Important: All previous-generation models are deprecated and will reach end of service on 31 July 2023

Important: All previous-generation models are deprecated and will reach end of service effective 31 July 2023. On that date, all previous-generation models will be removed from the service and the documentation. The previous deprecation date was 3 March 2023. The new date allows users more time to migrate to the appropriate next-generation models. But users must migrate to the equivalent next-generation model by 31 July 2023.

Most previous-generation models were deprecated on 15 March 2022. Previously, the Arabic and Japanese models were not deprecated. Deprecation now applies to all previous-generation models.

Note: When the previous-generation en-US_BroadbandModel is removed from service, the next-generation en-US_Multimedia model will become the default model for speech recognition requests.

Defect fix: Improved training time for next-generation custom language models

Defect fix: Training time for next-generation custom language models is now significantly improved. Previously, training time took much longer than necessary, as reported for training of Japanese custom language models. The problem was corrected by an internal fix.

Defect fix: Dynamically generated grammar files now work properly

Defect fix: Dynamically generated grammar files now work properly. Previously, dynamic grammar files could cause internal failures, as reported for integration of Speech to Text with IBM® watsonx™ Assistant. The problem was corrected by an internal fix.

20 January 2023

Deprecated Arabic and United Kingdom model names are no longer available

The following Arabic and United Kingdom model names are no longer accepted by the service:

  • ar-AR_BroadbandModel - Use ar-MS_BroadbandModel instead.
  • en-UK_NarrowbandModel - Use en-GB_NarrowbandModel instead.
  • en-UK_BroadbandModel - Use en-GB_BroadbandModel instead.

The Arabic model name was deprecated on 2 December 2020. The UK English model names were deprecated on 14 July 2017.

Cloud Foundry deprecation and migration to resource groups

IBM announced the deprecation of IBM Cloud Foundry on 31 May 2022. As of 30 November 2022, new IBM Cloud Foundry applications cannot be created and only existing users are able to deploy applications. IBM Cloud Foundry reaches end of support on 1 June 2023. At that time, any IBM Cloud Foundry application runtime instances running IBM Cloud Foundry applications will be permanently disabled, deprovisioned, and deleted.

To continue to use your IBM Cloud applications beyond 1 June 2023, you must migrate to resource groups before that date. Resource groups are conceptually similar to Cloud Foundry spaces. They include several extra benefits, such as finer-grained access control by using IBM Cloud Identity and Access Management (IAM), the ability to connect service instances to apps and service across different regions, and an easy way to view usage per group.

The max_alternatives parameter is now available for use with next-generation models

The max_alternatives parameter is now available for use with all next-generation models. The parameter is generally available for all next-generation models. For more information, see Maximum alternatives.

Defect fix: Allow use of both max_alternatives and end_of_phrase_silence_time parameters with next-generation models

Defect fix: When you use both the max_alternatives and end_of_phrase_silence_time parameters in the same request with next-generation models, the service now returns multiple alternative transcripts while also respecting the indicated pause interval. Previously, use of the two parameters in a single request generated a failure. (Use of the max_alternatives parameter with next-generation models was previously available as an experimental feature to a limited number of customers.)

Defect fix: Update French Canadian next-generation telephony model (upgrade required)

Defect fix: The French Canadian next-generation telephony model, fr-CA_Telephony, was updated to address an internal inconsistency that could cause an error during speech recognition. You need to upgrade any custom models that are based on the fr-CA_Telephony model. For more information about upgrading custom models, see

Defect fix: Add documentation guidelines for creating Japanese sounds-likes based on next-generation models

Defect fix: In sounds-likes for Japanese custom language models that are based on next-generation models, the character-sequence ウー is ambiguous in some left contexts. Do not use characters (syllables) that end with the phoneme /o/, such as ロ and ト. In such cases, use ウウ or just ウ instead of ウー. For example, use ロウウマン or ロウマン instead of ロウーマン. For more information, see Guidelines for Japanese.

Adding words directly to custom models that are based on next-generation models increases the training time

Adding custom words directly to a custom model that is based on a next-generation model causes training of a model to take a few minutes longer than it otherwise would. If you are training a model with custom words that you added by using the POST /v1/customizations/{customization_id}/words or PUT /v1/customizations/{customization_id}/words/{word_name} method, allow for some minutes of extra training time for the model. For more information, see

Maximum hours of audio resources for custom acoustic models in the Tokyo location has been increased

The maximum hours of audio resources that you can add to custom acoustic models in the Tokyo location is again 200 hours. Previously, the maximum was reduced to 50 hours for the Tokyo region. That reduction has been rescinded and postponed until next year. For more more information, see Maximum hours of audio.

5 December 2022

New Netherlands Dutch next-generation multimedia model
The service now offers a next-generation multimedia model for Netherlands Dutch: nl-NL_Multimedia. The new model supports low latency and is generally available. It also supports language model customization and grammars. For more information about next-generation models and low latency, see
Defect fix: Correct custom word recognition in transcription results for next-generation models
Defect fix: For language model customization with next-generation models, custom words are now recognized and used in all transcripts. Previously, custom words sometimes failed to be recognized and used in transcription results.
Defect fix: Correct use of display_as field in transcription results for next-generation models
Defect fix: For language model customization with next-generation models, the value of the display_as field for a custom word now appears in all transcripts. Previously, the value of the word field sometimes appeared in transcription results.
Defect fix: Update custom model naming documentation
Defect fix: The documentation now provides detailed rules for naming custom language models and custom acoustic models. For more information, see

20 October 2022

Updates to English next-generation telephony models

The English next-generation telephony models have been updated for improved speech recognition:

  • en-AU_Telephony
  • en-GB_Telephony
  • en-IN_Telephony
  • en-US_Telephony

All of these models continue to support low latency. You do not need to upgrade custom models that are based on these models. For more information about all available next-generation models, see Next-generation languages and models.

Defect fix: Update Japanese next-generation multimedia model (upgrade required)

Defect fix: The Japanese next-generation multimedia model, ja-JP_Multimedia, was updated to address an internal inconsistency that could cause an error during speech recognition with low latency. You need to upgrade any custom models that are based on the ja-JP_Multimedia model. For more information about upgrading custom models, see

7 October 2022

New Swedish next-generation telephony model

The service now offers a next-generation telephony model for Swedish: sv-SE_Telephony. The new model supports low latency and is generally available. It also supports language model customization and grammars. For more information about next-generation models and low latency, see

Updates to English next-generation telephony models

The English next-generation telephony models have been updated for improved speech recognition:

  • en-AU_Telephony
  • en-GB_Telephony
  • en-IN_Telephony
  • en-US_Telephony

All of these models continue to support low latency. You do not need to upgrade custom models that are based on these models. For more information about all available next-generation models, see Next-generation languages and models.

21 September 2022

New Activity Tracker event for GDPR deletion of user information

The service now returns an Activity Tracker event when you use the DELETE /v1/user_data method to delete all information about a user. The event is named speech-to-text.gdpr-user-data.delete. For more information, see Activity Tracker events.

Defect fix: Update some next-generation models to improve low-latency response time

Defect fix: The following next-generation models were updated to improve their response time when the low_latency parameter is used:

  • en-IN_Telephony
  • hi-IN_Telephony
  • it-IT_Multimedia
  • nl-NL_Telephony

Previously, these models did not return recognition results as quickly as expected when the low_latency parameter was used. You do not need to upgrade custom models that are based on these models. For more information about all available next-generation models, see Next-generation languages and models.

19 August 2022

Important: Deprecation date for most previous-generation models is now 3 March 2023

Superseded: This deprecation notice is superseded by the 15 February 2023 service update. The end of service date for all previous-generation models is now 31 July 2023.

On 15 March 2022, the previous-generation models for all languages other than Arabic and Japanese were deprecated. At that time, the deprecated models were to remain available until 15 September 2022. To allow users more time to migrate to the appropriate next-generation models, the deprecated models will now remain available until 3 March 2023. As with the initial deprecation notice, the Arabic and Japanese previous-generation models are not deprecated. For a complete list of all deprecated models, see the 15 March 2022 service update.

On 3 March 2023, the deprecated models will be removed from the service and the documentation. If you use any of the deprecated models, you must migrate to the equivalent next-generation model by the 3 March 2023.

Note: When the previous-generation en-US_BroadbandModel is removed from service, the next-generation en-US_Multimedia model will become the default model for speech recognition requests.

15 August 2022

New French Canadian next-generation multimedia model

The service now offers a next-generation multimedia model for French Canadian: fr-CA_Multimedia. The new model supports low latency and is generally available. It also supports language model customization and grammars. For more information about next-generation models and low latency, see

Updates to English next-generation telephony models

The English next-generation telephony models have been updated for improved speech recognition:

  • en-AU_Telephony
  • en-GB_Telephony
  • en-IN_Telephony
  • en-US_Telephony

All of these models continue to support low latency. You do not need to upgrade custom models that are based on these models. For more information about all available next-generation models, see Next-generation languages and models.

Italian next-generation multimedia model now supports low latency

The Italian next-generation multimedia model, it-IT_Multimedia, now supports low latency. For more information about next-generation models and low latency, see

Important: Maximum hours of audio data being reduced for custom acoustic models

Important: The maximum amount of audio data that you can add to a custom acoustic model is being reduced from 200 hours to 50 hours. This change is being phased into different locations from August to September 2022. For information about the schedule for the limit reduction and what it means for existing custom acoustic models that contain more than 50 hours of audio, see Maximum hours of audio.

3 August 2022

Defect fix: Update speech hesitations and hesitation markers documentation

Defect fix: Documentation for speech hesitations and hesitation markers has been updated. Previous-generation models include hesitation markers in place of speech hesitations in transcription results for most languages; smart formatting removes hesitation markers from US English final transcripts. Next-generation models include the actual speech hesitations in transcription results; smart formatting has no effect on their inclusion in final transcription results.

For more information, see:

1 June 2022

Updates to multiple next-generation telephony models

The following next-generation telephony models have been updated for improved speech recognition:

  • en-AU_Telephony
  • en-GB_Telephony
  • en-IN_Telephony
  • en-US_Telephony
  • ko-KR_Telephony

You do not need to upgrade custom models that are based on these models. For more information about all available next-generation models, see Next-generation languages and models.

25 May 2022

New beta character_insertion_bias parameter for next-generation models

All next-generation models now support a new beta parameter, character_insertion_bias, which is available with all speech recognition interfaces. By default, the service is optimized for each individual model to balance its recognition of candidate strings of different lengths. The model-specific bias is equivalent to 0.0. Each model's default bias is sufficient for most speech recognition requests.

However, certain use cases might benefit from favoring hypotheses with shorter or longer strings of characters. The parameter accepts values between -1.0 and 1.0 that represent a change from a model's default. Negative values instruct the service to favor shorter strings of characters. Positive values direct the service to favor longer strings of characters. For more information, see Character insertion bias.

19 May 2022

New Italian it-IT_Multimedia next-generation model

The service now offers a next-generation multimedia model for Italian: it-IT_Multimedia. The new model is generally available. It does not support low latency, but it does support language model customization and grammars. For more information about all available next-generation models, see Next-generation languages and models.

Updated Korean telephony and multimedia next-generation models

The existing Korean next-generation models have been updated:

  • The ko-KR_Telephony model has been updated for improved low-latency support for speech recognition.
  • The ko-KR_Multimedia model has been updated for improved speech recognition. The model now also supports low latency.

Both models are generally available, and both support language model customization and grammars. You do not need to upgrade custom language models that are based on these models. For more information about all available next-generation models, see Next-generation languages and models.

Defect fix: Confidence scores are now reported for all transcription results

Defect fix: Confidence scores are now reported for all transcription results. Previously, when the service returned multiple transcripts for a single speech recognition request, confidence scores might not be returned for all transcripts.

11 April 2022

New Brazilian Portuguese pt-BR_Multimedia next-generation model

The service now offers a next-generation multimedia model for Brazilian Portuguese: pt-BR_Multimedia. The new model supports low latency and is generally available. It also supports language model customization and grammars. For more information about next-generation models and low latency, see

Update to German de-DE_Multimedia next-generation model to support low latency

The next-generation German model, de-DE_Multimedia, now supports low latency. You do not need to upgrade custom models that are based on the updated German base model. For more information about the next-generation models and low latency, see

Support for sounds-like is now documented for custom models based on next-generation models

For custom language models that are based on next-generation models, support is now documented for sounds-like specifications for custom words. Support for sounds-likes has been available since late 2021.

Differences exist between the use of the sounds_like field for custom models that are based on next-generation and previous-generation models. For more information about using the sounds_like field with custom models that are based on next-generation models, see Working with custom words for next-generation models.

Important: Deprecated customization_id parameter removed from the documentation

Important: On 9 October 2018, the customization_id parameter of all speech recognition requests was deprecated and replaced by the language_customization_id parameter. The customization_id parameter has now been removed from the documentation for the speech recognition methods:

  • /v1/recognize for WebSocket requests
  • POST /v1/recognize for synchronous HTTP requests (including multipart requests)
  • POST /v1/recognitions for asynchronous HTTP requests

Note: If you use the Watson SDKs, make sure that you have updated any application code to use the language_customization_id parameter instead of the customization_id parameter. The customization_id parameter will no longer be available from the equivalent methods of the SDKs as of their next major release. For more information about the speech recognition methods, see the API & SDK reference.

17 March 2022

Grammar support for next-generation models is now generally available

Grammar support is now generally available (GA) for next-general models that meet the following conditions:

  • The models are generally available.
  • The models support language model customization.

For more information, see the following topics:

New German next-generation multimedia model

The service now offers a next-generation multimedia model for German: de-DE_Multimedia. The new model is generally available. It does not support low latency. It does support language model customization (generally available) and grammars (beta).

For more information about all available next-generation models and their customization support, see

Beta next-generation en-WW_Medical_Telephony model now supports low latency

The beta next-generation en-WW_Medical_Telephony model now supports low latency. For more information about all next-generation models and low latency, see

15 March 2022

Important: Deprecation of most previous-generation models

Superseded: This deprecation notice is superseded by the 15 February 2023 service update. The end of service date for all previous-generation models is now 31 July 2023.

Effective 15 March 2022, previous-generation models for all languages other than Arabic and Japanese are deprecated. The deprecated models remain available until 15 September 2022, when they will be removed from the service and the documentation. The Arabic and Japanese previous-generation models are not deprecated.

The following previous-generation models are now deprecated:

  • Chinese (Mandarin): zh-CN_NarrowbandModel and zh-CN_BroadbandModel
  • Dutch (Netherlands): nl-NL_NarrowbandModel and nl-NL_BroadbandModel
  • English (Australian): en-AU_NarrowbandModel and en-AU_BroadbandModel
  • English (United Kingdom): en-GB_NarrowbandModel and en-GB_BroadbandModel
  • English (United States): en-US_NarrowbandModel, en-US_BroadbandModel, and en-US_ShortForm_NarrowbandModel
  • French (Canadian): fr-CA_NarrowbandModel and fr-CA_BroadbandModel
  • French (France): fr-FR_NarrowbandModel and fr-FR_BroadbandModel
  • German: de-DE_NarrowbandModel and de-DE_BroadbandModel
  • Italian: it-IT_NarrowbandModel and it_IT_BroadbandModel
  • Korean: ko-KR_NarrowbandModel and ko-KR_BroadbandModel
  • Portuguese (Brazilian): pt-BR_NarrowbandModel and pt-BR_BroadbandModel
  • Spanish (Argentinian): es-AR_NarrowbandModel and es-AR_BroadbandModel
  • Spanish (Castilian): es-ES_NarrowbandModel and es-ES_BroadbandModel
  • Spanish (Chilean): es-CL_NarrowbandModel and es-CL_BroadbandModel
  • Spanish (Colombian): es-CO_NarrowbandModel and es-CO_BroadbandModel
  • Spanish (Mexican): es-MX_NarrowbandModel and es-MX_BroadbandModel
  • Spanish (Peruvian): es-PE_NarrowbandModel and es-PE_BroadbandModel

If you use any of these deprecated models, you must migrate to the equivalent next-generation model by the end of service date.

Note: When the previous-generation en-US_BroadbandModel is removed from service on 15 September, the next-generation en-US_Multimedia model will become the default model for speech recognition requests.

Next-generation models now support audio-parsing parameters

All next-generation models now support the following audio-parsing parameters as generally available features:

  • end_of_phrase_silence_time specifies the duration of the pause interval at which the service splits a transcript into multiple final results. For more information, see End of phrase silence time.
  • split_transcript_at_phrase_end directs the service to split the transcript into multiple final results based on semantic features of the input. For more information, see Split transcript at phrase end.
Defect fix: Correct speaker labels documentation

Defect fix: Documentation of speaker labels included the following erroneous statement in multiple places: For next-generation models, speaker labels are not supported for use with interim results or low latency. Speaker labels are supported for use with interim results and low latency for next-generation models. For more information, see Speaker labels.

28 February 2022

Updates to English and French next-generation multimedia models to support low latency

The following multimedia models have been updated to support low latency:

  • Australian English: en-AU_Multimedia
  • UK English: en-GB_Multimedia
  • US English: en-US_Multimedia
  • French: fr-FR_Multimedia

You do not need to upgrade custom language models that are built on these base models. For more information about the next-generation models and low latency, see

New Castilian Spanish next-generation multimedia model

The service now offers a next-generation multimedia model for Castilian Spanish: es-ES_Multimedia. The new model supports low latency and is generally available. It also supports language model customization (generally available) and grammars (beta).

For more information about all available next-generation models and their customization support, see

11 February 2022

Defect fix: Correct custom model upgrade and base model version documentation

Defect fix: The documentation that describes the upgrade of custom models and the version strings that are used for different versions of base models has been updated. The documentation now states that upgrade for language model customization also applies to next-generation models. Also, the version strings that represent different versions of base models have been updated. And the base_model_version parameter can also be used with upgraded next-generation models.

For more information about custom model upgrade, when upgrade is necessary, and how to use older versions of custom models, see

Defect fix: Update capitalization documentation

Defect fix: The documentation that describes the service's automatic capitalization of transcripts has been updated. The service capitalizes appropriate nouns only for the following languages and models:

  • All previous-generation US English models
  • The next-generation German model

For more information, see Capitalization.

2 February 2022

New beta en-WW_Medical_Telephony model is now available

A new beta next-generation en-WW_Medical_Telephony is now available. The new model understands terms from the medical and pharmacological domains. Use the model in situations where you need to transcribe common medical terminology such as medicine names, product brands, medical procedures, illnesses, types of doctor, or COVID-19-related terminology. Common use cases include conversations between a patient and a medical provider (for example, a doctor, nurse, or pharmacist).

The new model is available for all supported English dialects: Australian, Indian, UK, and US. The new model supports language model customization and grammars as beta functionality. It supports most of the same parameters as the en-US_Telephony model, including smart_formatting for US English audio. It does not support the following parameters: low_latency, profanity_filter, redaction, and speaker_labels.

For more information, see The English medical telephony model.

Update to Chinese zh-CN_Telephony model

The next-generation Chinese model zh-CN_Telephony has been updated for improved speech recognition. The model continues to support low latency. By default, the service automatically uses the updated model for all speech recognition requests. For more information about all available next-generation models, see Next-generation languages and models.

If you have custom language models that are based on the updated model, you must upgrade your existing custom models to take advantage of the updates by using the POST /v1/customizations/{customization_id}/upgrade_model method. For more information, see Upgrading custom models.

Update to Japanese ja-JP_Multimedia next-generation model to support low latency

The next-generation Japanese model ja-JP_Multimedia now supports low latency. You can use the low_latency parameter with speech recognition requests that use the model. You do not need to upgrade custom models that are based on the updated Japanese base model. For more information about the next-generation models and low latency, see

3 December 2021

New Latin American Spanish next-generation telephony model

The service now offers a next-generation telephony model for Latin American Spanish: es-LA_Telephony. The new model supports low latency and is generally available.

The es-LA_Telephony model applies to all Latin American dialects. It is the equivalent of the previous-generation models that are available for the Argentinian, Chilean, Colombian, Mexican, and Peruvian dialects. If you used a previous-generation model for any of these specific dialects, use the es-LA_Telephony model to migrate to the equivalent next-generation model.

For more information about all available next-generation models, see Next-generation languages and models.

Important: Custom language models based on certain next-generation models must be re-created

Important: If you created custom language models based on certain next-generation models, you must re-create the custom models. Until you re-create the custom language models, speech recognition requests that attempt to use the custom models fail with HTTP error code 400.

You need to re-create custom language models that you created based on the following versions of next-generation models:

  • For the en-AU_Telephony model, custom models that you created from en-AU_Telephony.v2021-03-03 to en-AU_Telephony.v2021-10-04.
  • For the en-GB_Telephony model, custom models that you created from en-GB_Telephony.v2021-03-03 to en-GB_Telephony.v2021-10-04.
  • For the en-US_Telephony model, custom models that you created from en-US_Telephony.v2021-06-17 to en-US_Telephony.v2021-10-04.
  • For the en-US_Multimedia model, custom models that you created from en-US_Multimedia.v2021-03-03 to en-US_Multimedia.v2021-10-04.

To identify the version of a model on which a custom language model is based, use the GET /v1/customizations method to list all of your custom language models or the GET /v1/customizations/{customization_id} method to list a specific custom language model. The versions field of the output shows the base model for a custom language model. For more information, see Listing custom language models.

To re-create a custom language model, first create a new custom model. Then add all of the previous custom model's corpora and custom words to the new model. You can then delete the previous custom model. For more information, see Creating a custom language model.

28 October 2021

New Chinese next-generation telephony model

The service now offers a next-generation telephony model for Mandarin Chinese: zh-CN_Telephony. The new model supports low latency and is generally available. For more information about all available next-generation models, see Next-generation languages and models.

New Australian English and UK English next-generation multimedia models

The service now offers the following next-generation multimedia models. The new models are generally available, and neither model supports low latency.

  • Australian English: en-AU_Multimedia
  • UK English: en-GB_Multimedia

For more information about all available next-generation models, see Next-generation languages and models.

Updates to multiple next-generation models for improved speech recognition

The following next-generation models have been updated for improved speech recognition:

  • Australian English telephony model (en-AU_Telephony)
  • UK English telephony model (en-GB_Telephony)
  • US English multimedia model (en-US_Multimedia)
  • US English telephony model (en-US_Telephony)
  • Castilian Spanish telephony model (es-ES_Telephony)

For more information about all available next-generation models, see Next-generation languages and models.

Grammar support for previous-generation models is now generally available

Grammar support is now generally available (GA) for previous-general models that meet the following conditions:

  • The models are generally available.
  • The models support language model customization.

For more information, see the following topics:

New beta grammar support for next-generation models

Grammar support is now available as beta functionality for all next-generation models. All next-generation models are generally available (GA) and support language model customization. For more information, see the following topics:

Note: Beta support for grammars by next-generation models is available for the Speech to Text service on IBM Cloud only. Grammars are not yet supported for next-generation models on IBM Cloud Pak for Data.

New custom_acoustic_model field for supported features

The GET /v1/models and GET /v1/models/{model_id} methods now report whether a model supports acoustic model customization. The SupportedFeatures object now includes an additional field, custom_acoustic_model, a boolean that is true for a model that supports acoustic model customization and false otherwise. Currently, the field is true for all previous-generation models and false for all next-generation models.

22 October 2021

Defect fix: Address asynchronous HTTP failures
Defect fix: The asynchronous HTTP interface failed to transcribe some audio. In addition, the callback for the request returned status recognitions.completed_with_results instead of recognitions.failed. This error has been resolved.

6 October 2021

Updates to Czech and Dutch next-generation models

The following next-generation language models have changed as indicated:

  • The Czech telephony model, cs-CZ_Telephony, is now generally available (GA). The model continues to support low latency.
  • The Belgian Dutch telephony model, nl-BE_Telephony, has been updated for improved speech recognition. The model continues to support low latency.
  • The Netherlands Dutch telephony model, nl-NL_Telephony, is now GA. In addition, the model now supports low latency.

For more information about all available next-generation language models, see Next-generation languages and models.

New US HIPAA support for Premium plans in Dallas location

US Health Insurance Portability and Accountability Act (HIPAA) support is now available for Premium plans that are hosted in the Dallas (us-south) location. For more information, see Health Insurance Portability and Accountability Act (HIPAA).

16 September 2021

New beta Czech and Netherlands Dutch next-generation models

The service now supports the following new next-generation language models. Both new models are beta functionality.

  • Czech: cs-CZ_Telephony. The new model supports low latency.
  • Netherlands Dutch: nl-NL_Telephony. The new model does not support low latency.

For more information about all available next-generation language models, see Next-generation languages and models.

Updates to Korean and Brazilian Portuguese next-generation models

The following next-generation models have been updated:

  • The Korean model ko-KR_Telephony now supports low latency.
  • The Brazilian Portuguese model pt-BR_Telephony has been updated for improved speech recognition.
Defect fix: Correct interim results and low-latency documentation

Defect fix: Documentation that describes the interim results and low-latency features with next-generation models has been rewritten for clarity and correctness. For more information, see the following topics:

Defect fix: Improve speakers labels results

Defect fix: When you use speakers labels with next-generation models, the service now identifies the speaker for all words of the input audio, including very short words that have the same start and end timestamps.

31 August 2021

All next-generation models are now generally available

All existing next-generation language models are now generally available (GA). They are supported for use in production environments and applications.

Language model customization for next-generation models is now generally available

Language model customization is now generally available (GA) for all available next-generation languages and models. Language model customization for next-generation models is supported for use in production environments and applications.

You use the same commands to create, manage, and use custom language models, corpora, and custom words for next-generation models as you do for previous-generation models. But customization for next-generation models works differently from customization for previous-generation models. For custom models that are based on next-generation models:

  • The custom models have no concept of out-of-vocabulary (OOV) words.
  • Words from corpora are not added to the words resource.
  • You cannot currently use the sounds-like feature for custom words.
  • You do not need to upgrade custom models when base language models are updated.
  • Grammars are not currently supported.

For more information about using language model customization for next-generation models, see

Additional topics describe managing custom language models, corpora, and custom words. These operations are the same for custom models based on previous- and next-generation models.

16 August 2021

New beta Indian English, Indian Hindi, Japanese, and Korean next-generation models

The service now supports the following new next-generation language models. All of the new models are beta functionality.

  • Indian English: en-IN_Telephony. The model supports low latency.
  • Indian Hindi: hi-IN_Telephony. The model supports low latency.
  • Japanese: ja-JP_Multimedia. The model does not support low latency.
  • Korean: ko-KR_Multimedia and ko-KR_Telephony. The models do not support low latency.

For more information about the next-generation models and low latency, see Next-generation languages and models and Low latency.

16 July 2021

New beta French next-generation model
The French next-generation language model fr-FR_Multimedia is now available. The new model does not support low latency. The model is beta functionality.
Updates to beta US English next-generation model for improved speech recognition
The next-generation US English en-US_Telephony model has been updated for improved speech recognition. The updated model continues to be beta functionality.
Defect fix: Update documentation for hesitation markers
Defect fix: The documentation failed to state that next-generation models do not produce hesitation markers. The documentation has been updated to note that only previous-generation models produce hesitation markers. Next-generation models include the actual hesitations in transcription results. For more information, see Speech hesitations and hesitation markers.

15 June 2021

New beta Belgian Dutch next-generation model

The Belgian Dutch (Flemish) next-generation language model nl-BE_Telephony is now available. The new model supports low latency. The model is beta functionality. For more information about the next-generation models and about low latency, see Next-generation languages and models and Low latency.

New beta low-latency support for Arabic, Canadian French, and Italian next-generation models

The following existing beta next-generation language models now support low latency:

  • Arabic ar-MS_Telephony model
  • Canadian French fr-CA_Telephony model
  • Italian it-IT_Telephony model

For more information about the next-generation models and about low latency, see Next-generation languages and models and Low latency.

Updates to beta Arabic and Brazilian Portuguese next-generation models for improved speech recognition

The following existing beta next-generation language models have been updated for improved speech recognition:

  • Arabic ar-MS_Telephony model
  • Brazilian Portuguese pt-BR_Telephony model

For more information about the next-generation models and about low latency, see Next-generation languages and models and Low latency.

26 May 2021

New beta support for audio_metrics parameter for next-generation models
The audio_metrics parameter is now supported as beta functionality for use with all next-generation languages and models. For more information, see Audio metrics.
New beta support for word_confidence parameter for next-generation models
The word_confidence parameter is now supported as beta functionality for use with all next-generation languages and models. For more information, see Word confidence.
Defect fix: Update documentation for next-generation models
Defect fix: The documentation has been updated to correct the following information:
  • When you use a next-generation model for speech recognition, final transcription results now include the confidence field. The field was always included in final transcription results when you use a previous-generation model. This fix addresses a limitation that was reported for the 12 April 2021 release of the next-generation models.
  • The documentation incorrectly stated that using the smart_formatting parameter causes the service to remove hesitation markers from final transcription results for Japanese. Smart formatting does not remove hesitation markers from final results for Japanese, only for US English. For more information, see What results does smart formatting affect?

27 April 2021

New beta Arabic and Brazilian Portuguese next-generation models

The service supports two new beta next-generation models:

  • The Brazilian Portuguese pt-BR_Telephony model, which supports low latency.
  • The Arabic (Modern Standard ) ar-MS_Telephony model, which does not support low latency.

For more information, see Next-generation languages and models.

Updates to beta Castilian Spanish next-generation model for improved speech recognition

The beta next-generation Castilian Spanish es-ES_Telephony model now supports the low_latency parameter. For more information, see Low latency.

New beta support for speaker labels with next-generation models

The speaker_labels parameter is now supported as beta functionality for use with the following next-generation models:

  • Australian English en-AU_Telephony model
  • UK English en-GB_Telephony model
  • US English en-US_Multimedia and en-US_Telephony models
  • German de-DE_Telephony model
  • Castilian Spanish es-ES_Telephony model

With the next generation models, the speaker_labels parameter is not supported for use with the interim_results or low_latency parameters at this time. For more information, see Speaker labels.

New HTTP error code for use of word_confidence with next-generation models

The word_confidence parameter is not supported for use with next-generation models. The service now returns the following 400 error code if you use the word_confidence parameter with a next-generation model for speech recognition:

{
  "error": "word_confidence is not a supported feature for model {model}",
  "code": 400,
  "code_description": "Bad Request"
}

12 April 2021

New beta next-generation language models and low_latency parameter

The service now supports a growing number of next-generation language models. The next-generation multimedia and telephony models improve upon the speech recognition capabilities of the service's previous generation of broadband and narrowband models. The new models leverage deep neural networks and bidirectional analysis to achieve both higher throughput and greater transcription accuracy. At this time, the next-generation models support only a limited number of languages and speech recognition features. The supported languages, models, and features will increase with future releases. The next-generation models are beta functionality.

Many of the next-generation models also support a new low_latency parameter that lets you request faster results at the possible expense of reduced transcription quality. When low latency is enabled, the service curtails its analysis of the audio, which can reduce the accuracy of the transcription. This trade-off might be acceptable if your application requires lower response time more than it does the highest possible accuracy.The low_latency parameter is beta functionality.

The low_latency parameter impacts your use of the interim_results parameter with the WebSocket interface. Interim results are available only for those next-generation models that support low latency, and only if both the interim_results and low_latency parameters are set to true.

17 March 2021

Defect fix: Fix limitation for asynchronous HTTP interface
Defect fix: The limitation that was reported with the asynchronous HTTP interface in the Dallas (us-south) location on 16 December 2020 has been addressed. Previously, a small percentage of jobs were entering infinite loops that prevented their execution. Asynchronous HTTP requests in the Dallas data center no longer experience this limitation.

2 December 2020

Arabic model renamed to ar-MS_BroadbandModel
The Arabic language broadband model is now named ar-MS_BroadbandModel. The former name, ar-AR_BroadbandModel, is deprecated. It will continue to function for at least one year but might be removed at a future date. You are encouraged to migrate to the new name at your earliest convenience.

2 November 2020

Canadian French models now generally available

The Canadian French models, fr-CA_BroadbandModel and fr-CA_NarrowbandModel, are now generally available (GA). They were previously beta. They also now support language model and acoustic model customization.

22 October 2020

Australian English models now generally available

The Australian English models, en-AU_BroadbandModel and en-AU_NarrowbandModel, are now generally available (GA). They were previously beta. They also now support language model and acoustic model customization.

Updates to Brazilian Portuguese models for improved speech recognition

The Brazilian Portuguese models, pt-BR_BroadbandModel and pt-BR_NarrowbandModel, have been updated for improved speech recognition. By default, the service automatically uses the updated models for all speech recognition requests. If you have custom language or custom acoustic models that are based on the models, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

The split_transcript_at_phrase_end parameter now generally available for all languages

The speech recognition parameter split_transcript_at_phrase_end is now generally available (GA) for all languages. Previously, it was generally available only for US and UK English. For more information, see Split transcript at phrase end.

7 October 2020

Updates to Japanese broadband model for improved speech recognition

The ja-JP_BroadbandModel model has been updated for improved speech recognition. By default, the service automatically uses the updated model for all speech recognition requests. If you have custom language or custom acoustic models that are based on this model, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

30 September 2020

Updates to pricing plans for the service

The pricing plans for the service have changed:

  • The service continues to offer a Lite plan that provides basic no-charge access to limited minutes of speech recognition per month.
  • The service offers a new Plus plan that provides a simple tiered pricing model and access to the service's customization capabilities.
  • The service offers a new Premium plan that provides significantly greater capacity and enhanced features.

The Plus plan replaces the Standard plan. The Standard plan continues to be available for purchase for a short time. It also continues to be available indefinitely to existing users of the plan with no change in their pricing. Existing users can upgrade to the Plus plan at any time.

For more information about the available pricing plans, see the following resources:

  • For general information about the pricing plans and answers to common questions, see the Pricing FAQs.
  • For more information about the pricing plans or to purchase a plan, see the Speech to Text service in the IBM Cloud® Catalog.

20 August 2020

New Canadian French models

The service now offers beta broadband and narrowband models for Canadian French:

  • fr-CA_BroadbandModel
  • fr-CA_NarrowbandModel

The new models do not support language model or acoustic model customization, speaker labels, or smart formatting. For more information about these and all supported models, see Supported previous-generation language models.

5 August 2020

New Australian English models

The service now offers beta broadband and narrowband models for Australian English:

  • en-AU_BroadbandModel
  • en-AU_NarrowbandModel

The new models do not support language model or acoustic model customization, or smart formatting. The new models do support speakers labels. For more information, see

Updates to multiple models for improved speech recognition

The following models have been updated for improved speech recognition:

  • French broadband model (fr-FR_BroadbandModel)
  • German broadband (de-DE_BroadbandModel) and narrowband (de-DE_NarrowbandModel) models
  • UK English broadband (en-GB_BroadbandModel) and narrowband (en-GB_NarrowbandModel) models
  • US English short-form narrowband (en-US_ShortForm_NarrowbandModel) model

By default, the service automatically uses the updated models for all speech recognition requests. If you have custom language or custom acoustic models that are based on these models, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

Hesitation marker for German changed

The hesitation marker that is used for the updated German broadband and narrowband models has changed from [hesitation] to %HESITATION. For more information, see Speech hesitations and hesitation markers.

4 June 2020

Defect fix: Improve latency for custom language models with many grammars
Defect fix: The latency issue for custom language models that contain a large number of grammars has been resolved. When initially used for speech recognition, such custom models could take multiple seconds to load. The custom models now load much faster, greatly reducing latency when they are used for recognition.

28 April 2020

Updates to Italian models for improved speech recognition

The Italian broadband (it-IT_BroadbandModel) and narrowband (it-IT_NarrowbandModel) models have been updated for improved speech recognition. By default, the service automatically uses the updated models for all speech recognition requests. If you have custom language or custom acoustic models that are based on these models, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

Dutch and Italian models now generally available

The Dutch and Italian language models are now generally available (GA) for speech recognition and for language model and acoustic model customization:

  • Dutch broadband model (nl-NL_BroadbandModel)
  • Dutch narrowband model (nl-NL_NarrowbandModel)
  • Italian broadband model (it-IT_BroadbandModel)
  • Italian narrowband model (it-IT_NarrowbandModel)

For more information about all available language models, see

1 April 2020

Acoustic model customization now generally available

Acoustic model customization is now generally available (GA) for all supported languages. As with custom language models, IBM does not charge for creating or hosting a custom acoustic model. You are charged only for using a custom model with a speech recognition request.

Using a custom language model, a custom acoustic model, or both types of model for transcription incurs an add-on charge of $0.03 (USD) per minute. This charge is in addition to the standard usage charge of $0.02 (USD) per minute, and it applies to all languages supported by the customization interface. So the total charge for using one or more custom models for speech recognition is $0.05 (USD) per minute.

16 March 2020

Speaker labels now supported for German and Korean
The service now supports speaker labels (the speaker_labels parameter) for German and Korean language models. Speaker labels identify which individuals spoke which words in a multi-participant exchange. For more information, see Speaker labels.
Activity Tracker now supported for asynchronous HTTP interface
The service now supports the use of Activity Tracker events for all operations of the asynchronous HTTP interface. IBM Cloud Activity Tracker records user-initiated activities that change the state of a service in IBM Cloud®. For more information, see Activity Tracker events.

24 February 2020

Updates to multiple models for improved speech recognition

The following models have been updated for improved speech recognition:

  • Dutch broadband model (nl-NL_BroadbandModel)
  • Dutch narrowband model (nl-NL_NarrowbandModel)
  • Italian broadband model (it-IT_BroadbandModel)
  • Italian narrowband model (it-IT_NarrowbandModel)
  • Japanese narrowband model (ja-JP_NarrowbandModel)
  • US English broadband model (en-US_BroadbandModel)

By default, the service automatically uses the updated models for all speech recognition requests. If you have custom language or custom acoustic models that are based on the models, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

Language model customization now available for Dutch and Italian

Language model customization is now supported for Dutch and Italian with the new versions of the following models:

  • Dutch broadband model (nl-NL_BroadbandModel)
  • Dutch narrowband model (nl-NL_NarrowbandModel)
  • Italian broadband model (it-IT_BroadbandModel)
  • Italian narrowband model (it-IT_NarrowbandModel)

For more information, see

Because the Dutch and Italian models are beta, their support for language model customization is also beta.

Japanese narrowband model now includes some multigram word units

The Japanese narrowband model (ja-JP_NarrowbandModel) now includes some multigram word units for digits and decimal fractions. The service returns these multigram units regardless of whether you enable smart formatting. The smart formatting feature understands and returns the multigram units that the model generates. If you apply your own post-processing to transcription results, you need to handle these units appropriately. For more information, see Japanese in the smart formatting documentation.

New speech activity detection and background audio suppression parameters for speech recognition

The service now offers two new optional parameters for controlling the level of speech activity detection. The parameters can help ensure that only relevant audio is processed for speech recognition.

  • The speech_detector_sensitivity parameter adjusts the sensitivity of speech activity detection. You can use the parameter to suppress word insertions from music, coughing, and other non-speech events.
  • The background_audio_suppression parameter suppresses background audio based on its volume to prevent it from being transcribed or otherwise interfering with speech recognition. You can use the parameter to suppress side conversations or background noise.

You can use the parameters individually or together. They are available for all interfaces and for most language models. For more information about the parameters, their allowable values, and their effect on the quality and latency of speech recognition, see Speech activity detection.

Activity Tracker now supported for customization interfaces

The service now supports the use of Activity Tracker events for all customization operations. IBM Cloud Activity Tracker records user-initiated activities that change the state of a service in IBM Cloud. You can use this service to investigate abnormal activity and critical actions and to comply with regulatory audit requirements. In addition, you can be alerted about actions as they happen. For more information, see Activity Tracker events.

Defect fix: Correct generation of processing metrics with WebSocket interface

Defect fix: The WebSocket interface now works seamlessly when generating processing metrics. Previously, processing metrics could continue to be delivered after the client sent a stop message to the service.

18 December 2019

New beta Italian models available

The service now offers beta broadband and narrowband models for the Italian language:

  • it-IT_BroadbandModel
  • it-IT_NarrowbandModel

These language models support acoustic model customization. They do not support language model customization. Because they are beta, these models might not be ready for production use and are subject to change. They are initial offerings that are expected to improve in quality with time and usage.

For more information, see the following sections:

New end_of_phrase_silence_time parameter for speech recognition

For speech recognition, the service now supports the end_of_phrase_silence_time parameter. The parameter specifies the duration of the pause interval at which the service splits a transcript into multiple final results. Each final result indicates a pause or extended silence that exceeds the pause interval. For most languages, the default pause interval is 0.8 seconds; for Chinese the default interval is 0.6 seconds.

You can use the parameter to effect a trade-off between how often a final result is produced and the accuracy of the transcription. Increase the interval when accuracy is more important than latency. Decrease the interval when the speaker is expected to say short phrases or single words.

For more information, see End of phrase silence time.

New split_transcript_at_phrase_end parameter for speech recognition

For speech recognition, the service now supports the split_transcript_at_phrase_end parameter. The parameter directs the service to split the transcript into multiple final results based on semantic features of the input, such as at the conclusion of sentences. The service bases its understanding of semantic features on the base language model that you use with a request. Custom language models and grammars can also influence how and where the service splits a transcript.

The parameter causes the service to add an end_of_utterance field to each final result to indicate the motivation for the split: full_stop, silence, end_of_data, or reset.

For more information, see Split transcript at phrase end.

12 December 2019

Full support for IBM Cloud IAM

The Speech to Text service now supports the full implementation of IBM Cloud Identity and Access Management (IAM). API keys for IBM Watson® services are no longer limited to a single service instance. You can create access policies and API keys that apply to more than one service, and you can grant access between services. For more information about IAM, see Authenticating to Watson services.

To support this change, the API service endpoints use a different domain and include the service instance ID. The pattern is api.{location}.speech-to-text.watson.cloud.ibm.com/instances/{instance_id}.

  • Example HTTP URL for an instance hosted in the Dallas location:

    https://api.us-south.speech-to-text.watson.cloud.ibm.com/instances/6bbda3b3-d572-45e1-8c54-22d6ed9e52c2

  • Example WebSocket URL for an instance hosted in the Dallas location:

    wss://api.us-south.speech-to-text.watson.cloud.ibm.com/instances/6bbda3b3-d572-45e1-8c54-22d6ed9e52c2

For more information about the URLs, see the API & SDK reference.

These URLs do not constitute a breaking change. The new URLs work for both your existing service instances and for new instances. The original URLs continue to work on your existing service instances for at least one year, until December 2020.

New network and data security features available

Support for the following new network and data security features is now available:

  • Support for private network endpoints

    Users of Premium plans can create private network endpoints to connect to the Speech to Text service over a private network. Connections to private network endpoints do not require public internet access. For more information, see Public and private network endpoints.

  • Support for data encryption with customer-managed keys

    Users of new Premium and Dedicated instances can integrate IBM® Key Protect for IBM Cloud® with the Speech to Text service to encrypt your data and manage encryption keys. For more information, see Protecting sensitive information in your Watson service.

10 December 2019

New beta Netherlands Dutch models available

The service now offers beta broadband and narrowband models for the Netherlands Dutch language:

  • nl-NL_BroadbandModel
  • nl-NL_NarrowbandModel

These language models support acoustic model customization. They do not support language model customization. Because they are beta, these models might not be ready for production use and are subject to change. They are initial offerings that are expected to improve in quality with time and usage.

For more information, see the following sections:

25 November 2019

Updates to speaker labels for improved identification of individual speakers
Speaker labels are updated to improve the identification of individual speakers for further analysis of your audio sample. For more information about the speaker labels feature, see Speaker labels. For more information about the improvements to the feature, see IBM Research AI Advances Speaker Diarization in Real Use Cases.

12 November 2019

New Seoul location now available
The Speech to Text service is now available in the IBM Cloud Seoul location (kr-seo). As with other locations, the IBM Cloud location uses token-based IAM authentication. All new services instances that you create in this location use IAM authentication.

1 November 2019

New limits on maximum number of custom models
You can create no more than 1024 custom language models and no more than 1024 custom acoustic models per owning credentials. For more information, see Maximum number of custom models.

1 October 2019

New US HIPAA support for Premium plans in Washington, DC, location
US HIPAA support is available for Premium plans that are hosted in the Washington, DC (us-east), location and are created on or after 1 April 2019. For more information, see US Health Insurance Portability and Accountability Act (HIPAA).

22 August 2019

Defect fix: Multiple small improvements
The service was updated for small defect fixes and improvements.

30 July 2019

New models for Spanish dialects now available

The service now offers broadband and narrowband language models in six Spanish dialects:

  • Argentinian Spanish (es-AR_BroadbandModel and es-AR_NarrowbandModel)
  • Castilian Spanish (es-ES_BroadbandModel and es-ES_NarrowbandModel)
  • Chilean Spanish (es-CL_BroadbandModel and es-CL_NarrowbandModel)
  • Colombian Spanish (es-CO_BroadbandModel and es-CO_NarrowbandModel)
  • Mexican Spanish (es-MX_BroadbandModel and es-MX_NarrowbandModel)
  • Peruvian Spanish (es-PE_BroadbandModel and es-PE_NarrowbandModel)

The Castilian Spanish models are not new. They are generally available (GA) for speech recognition and language model customization, and beta for acoustic model customization.

The other five dialects are new and are beta for all uses. Because they are beta, these additional dialects might not be ready for production use and are subject to change. They are initial offerings that are expected to improve in quality with time and usage.

For more information, see the following sections:

24 June 2019

Updates to Brazilian Portuguese and US English models for improved speech recognition

The following narrowband models have been updated for improved speech recognition:

  • Brazilian Portuguese narrowband model (pt-BR_NarrowbandModel)
  • US English narrowband model (en-US_NarrowbandModel)

By default, the service automatically uses the updated models for all speech recognition requests. If you have custom language or custom acoustic models that are based on the models, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

New support for concurrent requests to update different custom acoustic models

The service now allows you to submit multiple simultaneous requests to add different audio resources to a custom acoustic model. Previously, the service allowed only one request at a time to add audio to a custom model.

New updated field for methods that list custom models

The output of the HTTP GET methods that list information about custom language and custom acoustic models now includes an updated field. The field indicates the date and time in Coordinated Universal Time (UTC) at which the custom model was last modified.

Change to schema for warnings associated with custom model training

The schema changed for a warning that is generated by a custom model training request when the strict parameter is set to false. The names of the fields changed from warning_id and description to code and message, respectively. For more information, see the API & SDK reference.

10 June 2019

Processing metrics not available with synchronous HTTP interface
Processing metrics are available only with the WebSocket and asynchronous HTTP interfaces. They are not supported with the synchronous HTTP interface. For more information, see Processing metrics.

17 May 2019

New processing metrics and audio metrics features for speech recognition

The service now offers two types of optional metrics with speech recognition requests:

  • Processing metrics provide detailed timing information about the service's analysis of the input audio. The service returns the metrics at specified intervals and with transcription events, such as interim and final results. Use the metrics to gauge the service's progress in transcribing the audio.
  • Audio metrics provide detailed information about the signal characteristics of the input audio. The results provide aggregated metrics for the entire input audio at the conclusion of speech processing. Use the metrics to determine the characteristics and quality of the audio.

You can request both types of metrics with any speech recognition request. By default, the service returns no metrics for a request.

Updates to Japanese broadband model for improved speech recognition

The Japanese broadband model (ja-JP_BroadbandModel) has been updated for improved speech recognition. By default, the service automatically uses the updated model for all speech recognition requests. If you have custom language or custom acoustic models that are based on the model, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

10 May 2019

Updates to Spanish models for improved speech recognition

The Spanish language models have been updated for improved speech recognition:

  • es-ES_BroadbandModel
  • es-ES_NarrowbandModel

By default, the service automatically uses the updated models for all speech recognition requests. If you have custom language or custom acoustic models that are based on the models, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

19 April 2019

New strict parameter for custom model training now available
The training methods of the customization interface now include a strict query parameter that indicates whether training is to proceed if a custom model contains a mix of valid and invalid resources. By default, training fails if a custom model contains one or more invalid resources. Set the parameter to false to allow training to proceed as long as the model contains at least one valid resource. The service excludes invalid resources from the training.
New limits on maximum number of out-of-vocabulary words for custom language models
You can now add a maximum of 90 thousand out-of-vocabulary (OOV) words to the words resource of a custom language model. The previous maximum was 30 thousand OOV words. This figure includes OOV words from all sources (corpora, grammars, and individual custom words that you add directly). You can add a maximum of 10 million total words to a custom model from all sources. For more information, see How much data do I need?.

3 April 2019

New limits on maximum amount of audio for custom acoustic models
Custom acoustic models now accept a maximum of 200 hours of audio. The previous maximum limit was 100 hours of audio.

21 March 2019

Visibility of service credentials now restricted by role

Users can now see only service credential information that is associated with the role that has been assigned to their IBM Cloud account. For example, if you are assigned a reader role, any writer or higher levels of service credentials are no longer visible.

This change does not affect API access for users or applications with existing service credentials. The change affects only the viewing of credentials within IBM Cloud.

15 March 2019

New support for A-law audio format
The service now supports audio in the A-law (audio/alaw) format. For more information, see audio/alaw format.

11 March 2019

Change to passing value of 0 for max_alternatives parameter
For the max_alternatives parameter, the service again accepts a value of 0. If you specify 0. the service automatically uses the default value, 1. A change made for the March 4 service update caused a value of 0 to return an error. (The service returns an error if you specify a negative value.)
Change to passing value of 0 for word_alternatives_threshold parameter
For the word_alternatives_threshold parameter, the service again accepts a value of 0 . A change made for the March 4 service update caused a value of 0 to return an error. (The service returns an error if you specify a negative value.)
New limit on maximum precision for confidence scores
The service now returns all confidence scores with a maximum precision of two decimal places. This change includes confidence scores for transcripts, word confidence, word alternatives, keyword results, and speaker labels.

4 March 2019

Updates to Brazilian Portuguese, French, and Spanish narrowband models for improved speech recognition

The following narrowband language models have been updated for improved speech recognition:

  • Brazilian Portuguese narrowband model (pt-BR_NarrowbandModel)
  • French French model (fr-FR_NarrowbandModel)
  • Spanish narrowband model (es-ES_NarrowbandModel)

By default, the service automatically uses the updated models for all speech recognition requests. If you have custom language or custom acoustic models that are based on the models, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

28 January 2019

New support for IBM Cloud IAM by WebSocket interface

The WebSocket interface now supports token-based Identity and Access Management (IAM) authentication from browser-based JavaScript code. The limitation to the contrary has been removed. To establish an authenticated connection with the WebSocket /v1/recognize method:

  • If you use IAM authentication, include the access_token query parameter.
  • If you use Cloud Foundry service credentials, include the watson-token query parameter.

For more information, see Open a connection.

20 December 2018

New beta grammars feature for custom language models now available

The service now supports grammars for speech recognition. Grammars are available as beta functionality for all languages that support language model customization. You can add grammars to a custom language model and use them to restrict the set of phrases that the service can recognize from audio. You can define a grammar in Augmented Backus-Naur Form (ABNF) or XML Form.

The following four methods are available for working with grammars:

  • POST /v1/customizations/{customization_id}/grammars/{grammar_name} adds a grammar file to a custom language model.
  • GET /v1/customizations/{customization_id}/grammars lists information about all grammars for a custom model.
  • GET /v1/customizations/{customization_id}/grammars/{grammar_name} returns information about a specified grammar for a custom model.
  • DELETE /v1/customizations/{customization_id}/grammars/{grammar_name} removes an existing grammar from a custom model.

You can use a grammar for speech recognition with the WebSocket and HTTP interfaces. Use the language_customization_id and grammar_name parameters to identify the custom model and the grammar that you want to use. Currently, you can use only a single grammar with a speech recognition request.

For more information about grammars, see the following documentation:

For information about all methods of the interface, see the API & SDK reference.

New numeric redaction feature for US English, Japanese, and Korean now available

A new numeric redaction feature is now available to mask numbers that have three or more consecutive digits. Redaction is intended to remove sensitive personal information, such as credit card numbers, from transcripts. You enable the feature by setting the redaction parameter to true on a recognition request. The feature is beta functionality that is available for US English, Japanese, and Korean only. For more information, see Numeric redaction.

New French and German narrowband models now available

The following new German and French language models are now available with the service:

  • French narrowband model (fr-FR_NarrowbandModel)
  • German narrowband model (de-DE_NarrowbandModel)

Both new models support language model customization (GA) and acoustic model customization (beta). For more information, see Language support for customization.

New US English en-US_ShortForm_NarrowbandModel now available

A new US English language model, en-US_ShortForm_NarrowbandModel, is now available. The new model is intended for use in Interactive Voice Response and Automated Customer Support solutions. The model supports language model customization (GA) and acoustic model customization (beta). For more information, see The US English short-form model.

Updates to UK English and Spanish narrowband models for improved speech recognition

The following language models have been updated for improved speech recognition:

  • UK English narrowband model (en-GB_NarrowbandModel)
  • Spanish narrowband model (es-ES_NarrowbandModel)

By default, the service automatically uses the updated models for all speech recognition requests. If you have custom language or custom acoustic models that are based on the models, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

New support for G.279 audio format

The service now supports audio in the G.729 (audio/g729) format. The service supports only G.729 Annex D for narrowband audio. For more information, see audio/g729 format.

Speakers labels feature now available for UK English narrowband model

The speaker labels feature is now available for the narrowband model for UK English (en-GB_NarrowbandModel). The feature is beta functionality for all supported languages. For more information, see Speaker labels.

New limits on maximum amount of audio for custom acoustic models

The maximum amount of audio that you can add to a custom acoustic model has increased from 50 hours to 100 hours.

13 December 2018

New London location now available
The Speech to Text service is now available in the IBM Cloud London location (eu-gb). Like all locations, London uses token-based IAM authentication. All new services instances that you create in this location use IAM authentication.

12 November 2018

New support for smart formatting for Japanese speech recognition
The service now supports smart formatting for Japanese speech recognition. Previously, the service supported smart formatting for US English and Spanish only. The feature is beta functionality for all supported languages. For more information, see Smart formatting.

7 November 2018

New Tokyo location now available
The Speech to Text service is now available in the IBM Cloud Tokyo location (jp-tok). Like all locations, Tokyo uses token-based IAM authentication. All new services instances that you create in this location use IAM authentication.

30 October 2018

New support for token-based IBM Cloud IAM

The Speech to Text service has migrated to token-based IAM authentication for all locations. All IBM Cloud services now use IAM authentication. The Speech to Text service migrated in each location on the following dates:

  • Dallas (us-south): October 30, 2018
  • Frankfurt (eu-de): October 30, 2018
  • Washington, DC (us-east): June 12, 2018
  • Sydney (au-syd): May 15, 2018

The migration to IAM authentication affects new and existing service instances differently:

  • All new service instances that you create in any location now use IAM authentication to access the service. You can pass either a bearer token or an API key: Tokens support authenticated requests without embedding service credentials in every call; API keys use HTTP basic authentication. When you use any of the Watson SDKs, you can pass the API key and let the SDK manage the lifecycle of the tokens.
  • Existing service instances that you created in a location before the indicated migration date continue to use the {username} and {password} from their previous Cloud Foundry service credentials for authentication until you migrate them to use IAM authentication.

For more information, see the following documentation:

9 October 2018

Important updates to pricing charges for speech recognition requests

As of 1 October 2018, you are now charged for all audio that you pass to the service for speech recognition. The first one thousand minutes of audio that you send each month are no longer free. For more information about the pricing plans for the service, see the Speech to Text service in the IBM Cloud Catalog.

The Content-Type header now optional for most speech recognition requests

The Content-Type header is now optional for most speech recognition requests. The service now automatically detects the audio format (MIME type) of most audio. You must continue to specify the content type for the following formats:

  • audio/basic
  • audio/l16
  • audio/mulaw

Where indicated, the content type that you specify for these formats must include the sampling rate and can optionally include the number of channels and the endianness of the audio. For all other audio formats, you can omit the content type or specify a content type of application/octet-stream to have the service auto-detect the format.

When you use the curl command to make a speech recognition request with the HTTP interface, you must specify the audio format with the Content-Type header, specify "Content-Type: application/octet-stream", or specify "Content-Type:". If you omit the header entirely, curl uses a default value of application/x-www-form-urlencoded. Most of the examples in this documentation continue to specify the format for speech recognition requests regardless of whether it's required.

This change applies to the following methods:

  • /v1/recognize for WebSocket requests. The content-type field of the text message that you send to initiate a request over an open WebSocket connection is now optional.
  • POST /v1/recognize for synchronous HTTP requests. The Content-Type header is now optional. (For multipart requests, the part_content_type field of the JSON metadata is also now optional.)
  • POST /v1/recognitions for asynchronous HTTP requests. The Content-Type header is now optional.

For more information, see Audio formats.

Updates to Brazilian Portuguese broadband model for improved speech recognition

The Brazilian Portuguese broadband model, pt-BR_BroadbandModel, was updated for improved speech recognition. By default, the service automatically uses the updated model for all recognition requests. If you have custom language or custom acoustic models that are based on this model, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

The customization_id parameter renamed to language_customization_id

The customization_id parameter of the speech recognition methods is deprecated and will be removed in a future release. To specify a custom language model for a speech recognition request, use the language_customization_id parameter instead. This change applies to the following methods:

  • /v1/recognize for WebSocket requests
  • POST /v1/recognize for synchronous HTTP requests (including multipart requests)
  • POST /v1/recognitions for asynchronous HTTP requests

10 September 2018

New German broadband model

The service now supports a German broadband model, de-DE_BroadbandModel. The new German model supports language model customization (generally available) and acoustic model customization (beta).

Language model customization now available for Brazilian Portuguese

The existing Brazilian Portuguese models, pt-BR_BroadbandModel and pt-BR_NarrowbandModel, now support language model customization (generally available). The models were not updated to enable this support, so no upgrade of existing custom acoustic models is required.

Updates to US English and Japanese models for improved speech recognition

New versions of the US English and Japanese broadband and narrowband models are available:

  • US English broadband model (en-US_BroadbandModel)
  • US English narrowband model (en-US_NarrowbandModel)
  • Japanese broadband model (ja-JP_BroadbandModel)
  • Japanese narrowband model (ja-JP_NarrowbandModel)

The new models offer improved speech recognition. By default, the service automatically uses the updated models for all recognition requests. If you have custom language or custom acoustic models that are based on these models, you must upgrade your existing custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

Keyword spotting and word alternatives features now generally available

The keyword spotting and word alternatives features are now generally available (GA) rather than beta functionality for all languages. For more information, see

Defect fix: Improve documentation for customization interface

Defect fix: The following known issues that were associated with the customization interface have been resolved and are fixed in production. The following information is preserved for users who may have encountered the problems in the past.

  • If you add data to a custom language or custom acoustic model, you must retrain the model before using it for speech recognition. The problem shows up in the following scenario:

    1. The user creates a new custom model (language or acoustic) and trains the model.

    2. The user adds additional resources (words, corpora, or audio) to the custom model but does not retrain the model.

    3. The user cannot use the custom model for speech recognition. The service returns an error of the following form when used with a speech recognition request:

      {
  "code_description": "Bad Request",
  "code": 400,
  "error": "Requested custom language model is not available.
            Please make sure the custom model is trained."
}

    To work around this issue, the user must retrain the custom model on its latest data. The user can then use the custom model with speech recognition.

  • Before training an existing custom language or custom acoustic model, you must upgrade it to the latest version of its base model. The problem shows up in the following scenario:

    1. The user has an existing custom model (language or acoustic) that is based on a model that has been updated.
    2. The user trains the existing custom model against the old version of the base model without first upgrading to the latest version of the base model.
    3. The user cannot use the custom model for speech recognition.

    To work around this issue, the user must use the POST /v1/customizations/{customization_id}/upgrade_model or POST /v1/acoustic_customizations/{customization_id}/upgrade_model method to upgrade the custom model to the latest version of its base model. The user can then use the custom model with speech recognition.

7 September 2018

Session-based interface no longer available

The session-based HTTP REST interface is no longer supported. All information related to sessions is removed from the documentation. The following methods are no longer available:

  • POST /v1/sessions
  • POST /v1/sessions/{session_id}/recognize
  • GET /v1/sessions/{session_id}/recognize
  • GET /v1/sessions/{session_id}/observe_result
  • DELETE /v1/sessions/{session_id}

If your application uses the sessions interface, you must migrate to one of the remaining HTTP REST interfaces or to the WebSocket interface. For more information, see the service update for 8 August 2018.

8 August 2018

Deprecation notice for session-based speech recognition interface

The session-based HTTP REST interface is deprecated as of August 8, 2018. All methods of the sessions API will be removed from service as of September 7, 2018, after which you will no longer be able to use the session-based interface. This notice of immediate deprecation and 30-day removal applies to the following methods:

  • POST /v1/sessions
  • POST /v1/sessions/{session_id}/recognize
  • GET /v1/sessions/{session_id}/recognize
  • GET /v1/sessions/{session_id}/observe_result
  • DELETE /v1/sessions/{session_id}

If your application uses the sessions interface, you must migrate to one of the following interfaces by September 7:

  • For stream-based speech recognition (including live-use cases), use the WebSocket interface, which provides access to interim results and the lowest latency.
  • For file-based speech recognition, use one of the following interfaces:
    • For shorter files of up to a few minutes of audio, use either the synchronous HTTP interface (POST /v1/recognize) or the asynchronous HTTP interface (POST /v1/recognitions).
    • For longer files of more than a few minutes of audio, use the asynchronous HTTP interface. The asynchronous HTTP interface accepts as much as 1 GB of audio data with a single request.

The WebSocket and HTTP interfaces provide the same results as the sessions interface (only the WebSocket interface provides interim results). You can also use one of the Watson SDKs, which simplify application development with any of the interfaces. For more information, see the API & SDK reference.

13 July 2018

Updates to Spanish narrowband model for improved speech recognition

The Spanish narrowband model, es-ES_NarrowbandModel, was updated for improved speech recognition. By default, the service automatically uses the updated model for all recognition requests. If you have custom language or custom acoustic models that are based on this model, you must upgrade your custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

As of this update, the following two versions of the Spanish narrowband model are available:

  • es_ES.8kHz.general.lm20180522235959.am20180522235959 (current)
  • es_ES.8kHz.general.lm20180308235959.am20180308235959 (previous)

The following version of the model is no longer available:

  • es_ES.8kHz.general.lm20171031235959.am20171031235959

A recognition request that attempts to use a custom model that is based on the now unavailable base model uses the latest base model without any customization. The service returns the following warning message: Using non-customized default base model, because your custom {type} model has been built with a version of the base model that is no longer supported. To resume using a custom model that is based on the unavailable model, you must first upgrade the model by using the appropriate upgrade_model method described previously.

12 June 2018

New features for applications hosted in Washington, DC, location

The following features are enabled for applications that are hosted in Washington, DC (us-east):

  • The service now supports a new API authentication process. For more information, see the 30 October 2018 service update.
  • The service now supports the X-Watson-Metadata header and the DELETE /v1/user_data method. For more information, see Information security.

15 May 2018

New features for applications hosted in Sydney location

The following features are enabled for applications that are hosted in Sydney (au-syd):

  • The service now supports a new API authentication process. For more information, see the 30 October 2018 service update.
  • The service now supports the X-Watson-Metadata header and the DELETE /v1/user_data method. For more information, see Information security.

26 March 2018

Language model customization now available for French broadband model

The service now supports language model customization for the French broadband language model, fr-FR_BroadbandModel. The French model is generally available (GA) for production use with language model customization.

Updates to French, Korean, and Spanish models for improved speech recognition

The following models were updated for improved speech recognition:

  • Korean narrowband model (ko-KR_NarrowbandModel)
  • Spanish narrowband model (es-ES_NarrowbandModel)
  • French broadband model (fr-FR_BroadbandModel)

By default, the service automatically uses the updated models for all recognition requests. If you have custom language or custom acoustic models that are based on either of these models, you must upgrade your custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models.

The version parameter renamed to base_model_version

The version parameter of the following methods is now named base_model_version:

  • /v1/recognize for WebSocket requests
  • POST /v1/recognize for sessionless HTTP requests
  • POST /v1/sessions for session-based HTTP requests
  • POST /v1/recognitions for asynchronous HTTP requests

The base_model_version parameter specifies the version of a base model that is to be used for speech recognition. For more information, see Using upgraded custom models for speech recognition and Making speech recognition requests with upgraded custom models.

New support for smart formatting for Spanish speech recognition

Smart formatting is now supported for Spanish as well as US English. For US English, the feature also now converts keyword strings into punctuation symbols for periods, commas, question marks, and exclamation points. For more information, see Smart formatting.

1 March 2018

Updates to French and Spanish broadband models for improved speech recognition

The French and Spanish broadband models, fr-FR_BroadbandModel and es-ES_BroadbandModel, have been updated for improved speech recognition. By default, the service automatically uses the updated models for all recognition requests. If you have custom language or custom acoustic models that are based on either of these models, you must upgrade your custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information, see Upgrading custom models. The section presents rules for upgrading custom models, the effects of upgrading, and approaches for using upgraded models.

1 February 2018

New Korean models

The service now offers language models for Korean: ko-KR_BroadbandModel for audio that is sampled at a minimum of 16 kHz, and ko-KR_NarrowbandModel for audio that is sampled at a minimum of 8 kHz. For more information, see Previous-generation languages and models.

For language model customization, the Korean models are generally available (GA) for production use; for acoustic model customization, they are beta functionality. For more information, see Language support for customization.

  • For more information about how the service parses corpora for Korean, see Parsing of Korean.
  • For more information about creating sounds-like pronunciations for custom words in Korean, see Guidelines for Korean.

14 December 2017

Language model customization now generally available

Language model customization and all associated parameters are now generally available (GA) for all supported languages: Japanese, Spanish, UK English, and US English.

Beta acoustic model customization now available for all languages

The service now supports acoustic model customization as beta functionality for all available languages. You can create custom acoustic models for broadband or narrowband models for all languages. For an introduction to customization, including acoustic model customization, see Understanding customization.

New version parameter for speech recognition

The various methods for making recognition requests now include a new version parameter that you can use to initiate requests that use either the older or upgraded versions of base and custom models. Although it is intended primarily for use with custom models that have been upgraded, the version parameter can also be used without custom models. For more information, see Making speech recognition requests with upgraded custom models.

Updates to US English models for improved speech recognition

The US English models, en-US_BroadbandModel and en-US_NarrowbandModel, have been updated for improved speech recognition. By default, the service automatically uses the updated models for all recognition requests. If you have custom language or custom acoustic models that are based on the US English models, you must upgrade your custom models to take advantage of the updates by using the following methods:

  • POST /v1/customizations/{customization_id}/upgrade_model
  • POST /v1/acoustic_customizations/{customization_id}/upgrade_model

For more information about the procedure, see Upgrading custom models. The section presents rules for upgrading custom models, the effects of upgrading, and approaches for using upgraded models. Currently, the methods apply only to the new US English base models. But the same information will apply to upgrades of other base models as they become available.

Language model customization now available for UK English

The service now supports language model customization for the UK English models, en-GB_BroadbandModel and en-GB_NarrowbandModel. Although the service handles UK and US English corpora and custom words in a generally similar fashion, some important differences exist:

2 October 2017

New beta acoustic model customization interface for US English, Japanese, and Spanish

The customization interface now offers acoustic model customization. You can create custom acoustic models that adapt the service's base models to your environment and speakers. You populate and train a custom acoustic model on audio that more closely matches the acoustic signature of the audio that you want to transcribe. You then use the custom acoustic model with recognition requests to increase the accuracy of speech recognition.

Custom acoustic models complement custom language models. You can train a custom acoustic model with a custom language model, and you can use both types of model during speech recognition. Acoustic model customization is a beta interface that is available only for US English, Japanese, and Spanish.

New beta customization_weight parameter for custom language models

For language model customization, the service now includes a beta feature that sets an optional customization weight for a custom language model. A customization weight specifies the relative weight to be given to words from a custom language model versus words from the service's base vocabulary. You can set a customization weight during both training and speech recognition. For more information, see Using customization weight.

Updates to Japanese broadband model for improved speech recognition

The ja-JP_BroadbandModel language model was upgraded to capture improvements in the base model. The upgrade does not affect existing custom models that are based on the model.

New endianness parameter for audio/l16 audio format

The service now includes a parameter to specify the endianness of audio that is submitted in audio/l16 (Linear 16-bit Pulse-Code Modulation (PCM)) format. In addition to specifying rate and channels parameters with the format, you can now also specify big-endian or little-endian with the endianness parameter. For more information, see audio/l16 format.

14 July 2017

New support for MP3 (MPEG) audio format

The service now supports the transcription of audio in the MP3 or Motion Picture Experts Group (MPEG) format. For more information, see audio/mp3 and audio/mpeg formats.

Beta language model customization now available for Spanish

The language model customization interface now supports Spanish as beta functionality. You can create a custom model based on either of the base Spanish language models: es-ES_BroadbandModel or es-ES_NarrowbandModel; for more information, see Creating a custom language model. Pricing for recognition requests that use Spanish custom language models is the same as for requests that use US English and Japanese models.

New dialect field for method that creates a custom language model

The JSON CreateLanguageModel object that you pass to the POST /v1/customizations method to create a new custom language model now includes a dialect field. The field specifies the dialect of the language that is to be used with the custom model. By default, the dialect matches the language of the base model. The parameter is meaningful only for Spanish models, for which the service can create a custom model that is suited for speech in one of the following dialects:

  • es-ES for Castilian Spanish (the default)
  • es-LA for Latin-American Spanish
  • es-US for North-American (Mexican) Spanish

The GET /v1/customizations and GET /v1/customizations/{customization_id} methods of the customization interface include the dialect of a custom model in their output. For more information, see Creating a custom language model and Listing custom language models.

New names for UK English models

The names of the language models en-UK_BroadbandModel and en-UK_NarrowbandModel have been deprecated. The models are now available with the names en-GB_BroadbandModel and en-GB_NarrowbandModel.

The deprecated en-UK_{model} names continue to function, but the GET /v1/models method no longer returns the names in the list of available models. You can still query the names directly with the GET /v1/models/{model_id} method.

1 July 2017

Language model custom now generally available for US English and Japanese

The language model customization interface of the service is now generally available (GA) for both of its supported languages, US English and Japanese. IBM does not charge for creating, hosting, or managing custom language models. As described in the next bullet, IBM now charges an extra $0.03 (USD) per minute of audio for recognition requests that use custom models.

Updates to pricing plans for the service

IBM updated the pricing for the service by

  • Eliminating the add-on price for using narrowband models
  • Providing Graduated Tiered Pricing for high-volume customers
  • Charging an additional $0.03 (USD) per minute of audio for recognition requests that use US English or Japanese custom language models

For more information about the pricing updates, see

Empty body no longer required for HTTP POST requests

You no longer need to pass an empty data object as the body for the following POST requests:

  • POST /v1/sessions
  • POST /v1/register_callback
  • POST /v1/customizations/{customization_id}/train
  • POST /v1/customizations/{customization_id}/reset
  • POST /v1/customizations/{customization_id}/upgrade_model

For example, you now call the POST /v1/sessions method with curl as follows:

curl -X POST -u "{username}:{password}" \
--cookie-jar cookies.txt \
"{url}/v1/sessions"

You no longer need to pass the following curl option with the request: --data "{}". If you experience any problems with one of these POST requests, try passing an empty data object with the body of the request. Passing an empty object does not change the nature or meaning of the request in any way.

22 May 2017

The continuous parameter removed from all methods

The continuous parameter is removed from all methods that initiate recognition requests. The service now transcribes an entire audio stream until it ends or times out, whichever occurs first. This behavior is equivalent to setting the former continuous parameter to true. By default, the service previously stopped transcription at the first half-second of non-speech (typically silence) if the parameter was omitted or set to false.

Existing applications that set the parameter to true will see no change in behavior. Applications that set the parameter to false or that relied on the default behavior will likely see a change. If a request specifies the parameter, the service now responds by returning a warning message for the unknown parameter:

"warnings": [
  "Unknown arguments: continuous."
]

The request succeeds despite the warning, and an existing session or WebSocket connection is not affected.

IBM removed the parameter to respond to overwhelming feedback from the developer community that specifying continuous=false added little value and could reduce overall transcription accuracy.

Sending audio required to avoid session timeout

It is no longer possible to avoid a session timeout without sending audio:

  • When you use the WebSocket interface, the client can no longer keep a connection alive by sending a JSON text message with the action parameter set to no-op. Sending a no-op message does not generate an error, but it has no effect.
  • When you use sessions with the HTTP interface, the client can no longer extend the session by sending a GET /v1/sessions/{session_id}/recognize request. The method still returns the status of an active session, but it does not keep the session active.

You can now do the following to keep a session alive:

  • Set the inactivity_timeout parameter to -1 to avoid the 30-second inactivity timeout.
  • Send any audio data, including just silence, to the service to avoid the 30-second session timeout. You are charged for the duration of any data that you send to the service, including the silence that you send to extend a session.

For more information, see Timeouts. Ideally, you would establish a session just before you obtain audio for transcription and maintain the session by sending audio at a rate that is close to real time. Also, make sure your application recovers gracefully from closed sessions or connections.

IBM removed this functionality to ensure that it continues to offer all users a best-in-class, low-latency speech recognition service.

10 April 2017

Speaker labels now supported for US English, Spanish, and Japanese

The service now supports the speaker labels feature for the following broadband models:

  • US English broadband model (en-US-BroadbandModel)
  • Spanish broadband model (es-ES-BroadbandModel)
  • Japanese broadband model (ja-JP_BroadbandModel)

For more information, see Speaker labels.

New support for Web Media (WebM) audio format

The service now supports the Web Media (WebM) audio format with the Opus or Vorbis codec. The service now also supports the Ogg audio format with the Vorbis codec in addition to the Opus codec. For more information about supported audio formats, see audio/webm format.

New support for Cross-Origin Resource Sharing

The service now supports Cross-Origin Resource Sharing (CORS) to allow browser-based clients to call the service directly. For more information, see CORS support.

New method to unregister a callback URL with asynchronous HTTP interface

The asynchronous HTTP interface now offers a POST /v1/unregister_callback method that removes the registration for an allowlisted callback URL. For more information, see Unregistering a callback URL.

Defect fix: Eliminate timeouts for long audio with WebSocket interface

Defect fix: The WebSocket interface no longer times out for recognition requests for especially long audio files. You no longer need to request interim results with the JSON start message to avoid the timeout. (This issue was described in the update for 10 March 2016.)

New HTTP error codes

The following language model customization methods can now return the following HTTP error codes:

  • The DELETE /v1/customizations/{customization_id} method now returns HTTP response code 401 if you attempt to delete a nonexistent custom model.
  • The DELETE /v1/customizations/{customization_id}/corpora/{corpus_name} method now returns HTTP response code 400 if you attempt to delete a nonexistent corpus.

8 March 2017

Asynchronous HTTP interface is now generally available
The asynchronous HTTP interface is now generally available (GA). Prior to this date, it was beta functionality.

1 December 2016

New beta speaker labels feature

The service now offers a beta speaker labels feature for narrowband audio in US English, Spanish, or Japanese. The feature identifies which words were spoken by which speakers in a multi-person exchange. The sessionless, session-based, asynchronous, and WebSocket recognition methods each include a speaker_labels parameter that accepts a boolean value to indicate whether speaker labels are to be included in the response. For more information about the feature, see Speaker labels.

Beta language model customization now available for Japanese

The beta language model customization interface is now supported for Japanese in addition to US English. All methods of the interface support Japanese. For more information, see the following sections:

New method for listing information about a corpus

The language model customization interface now includes a GET /v1/customizations/{customization_id}/corpora/{corpus_name} method that lists information about a specified corpus. The method is useful for monitoring the status of a request to add a corpus to a custom model. For more information, see Listing corpora for a custom language model.

New count field for methods that list words for custom language models

The JSON response that is returned by the GET /v1/customizations/{customization_id}/words and GET /v1/customizations/{customization_id}/words/{word_name} methods now includes a count field for each word. The field indicates the number of times the word is found across all corpora. If you add a custom word to a model before it is added by any corpora, the count begins at 1. If the word is added from a corpus first and later modified, the count reflects only the number of times it is found in corpora. For more information, see Listing custom words from a custom language model.

For custom models that were created prior to the existence of the count field, the field always remains at 0. To update the field for such models, add the model's corpora again and include the allow_overwrite parameter with the POST /v1/customizations/{customization_id}/corpora/{corpus_name} method.

New sort parameter for methods that list words for custom language models

The GET /v1/customizations/{customization_id}/words method now includes a sort query parameter that controls the order in which the words are to be listed. The parameter accepts two arguments, alphabetical or count, to indicate how the words are to be sorted. You can prepend an optional + or - to an argument to indicate whether the results are to be sorted in ascending or descending order. By default, the method displays the words in ascending alphabetical order. For more information, see Listing custom words from a custom language model.

For custom models created prior to the introduction of the count field, use of the count argument with the sort parameter is meaningless. Use the default alphabetical argument with such models.

New error field format for methods that list words for custom language models

The error field that can be returned as part of the JSON response from the GET /v1/customizations/{customization_id}/words and GET /v1/customizations/{customization_id}/words/{word_name} methods is now an array. If the service discovered one or more problems with a custom word's definition, the field lists each problem element from the definition and provides a message that describes the problem. For more information, see Listing custom words from a custom language model.

The keywords_threshold and word_alternatives_threshold parameters no longer accept a null value

The keywords_threshold and word_alternatives_threshold parameters of the recognition methods no longer accept a null value. To omit keywords and word alternatives from the response, omit the parameters. A specified value must be a float.

22 September 2016

New beta language model customization interface
The service now offers a new beta language model customization interface for US English. You can use the interface to tailor the service's base vocabulary and language models via the creation of custom language models that include domain-specific terminology. You can add custom words individually or have the service extract them from corpora. To use your custom models with the speech recognition methods that are offered by any of the service's interfaces, pass the customization_id query parameter. For more information, see
New support for audio/mulaw audio format
The list of supported audio formats now includes audio/mulaw, which provides single-channel audio encoded using the u-law (or mu-law) data algorithm. When you use this format, you must also specify the sampling rate at which the audio is captured. For more information, see audio/mulaw format.
New supported_features identified when listing models
The GET /v1/models and GET /v1/models/{model_id} methods now return a supported_features field as part of their output for each language model. The additional information describes whether the model supports customization. For more information, see the API & SDK reference.

30 June 2016

Beta asynchronous HTTP interface now supports all available languages
The beta asynchronous HTTP interface now supports all languages that are supported by the service. The interface was previously available for US English only. For more information, see The asynchronous HTTP interface and the API & SDK reference.

23 June 2016

New beta asynchronous HTTP interface now available
A beta asynchronous HTTP interface is now available. The interface provides full recognition capabilities for US English transcription via non-blocking HTTP calls. You can register callback URLs and provide user-specified secret strings to achieve authentication and data integrity with digital signatures. For more information, see The asynchronous HTTP interface and the API & SDK reference.
New beta smart_formatting parameter for speech recognition
A beta smart formatting feature that converts dates, times, series of digits and numbers, phone numbers, currency values, and Internet addresses into more conventional representations in final transcripts. You enable the feature by setting the smart_formatting parameter to true on a recognition request. The feature is beta functionality that is available for US English only. For more information, see Smart formatting.
New French broadband model
The list of supported models for speech recognition now includes fr-FR_BroadbandModel for audio in the French language that is sampled at a minimum of 16 kHz. For more information, see Previous-generation languages and models.
New support for audio/basic audio format
The list of supported audio formats now includes audio/basic. The format provides single-channel audio that is encoded by using 8-bit u-law (or mu-law) data that is sampled at 8 kHz. For more information, see audio/basic format.
Speech recognition methods now return warnings for invalid parameters
The various recognition methods can return a warnings response that includes messages about invalid query parameters or JSON fields that are included with a request. The format of the warnings changed. For example, "warnings": "Unknown arguments: [u'{invalid_arg_1}', u'{invalid_arg_2}']." is now "warnings": "Unknown arguments: {invalid_arg_1}, {invalid_arg_2}."
Empty body required for HTTP POST methods that pass no data
For HTTP POST requests that do not otherwise pass data to the service, you must include an empty request body of the form {}. With the curl command, you use the --data option to pass the empty data.

10 March 2016

New maximum limits on audio transmitted for speech recognition
Both forms of data transmission (one-shot delivery and streaming) now impose a size limit of 100 MB on the audio data, as does the WebSocket interface. Formerly, the one-shot approach had a maximum limit of 4 MB of data. For more information, see Audio transmission (for all interfaces) and Send audio and receive recognition results (for the WebSocket interface). The WebSocket section also discusses the maximum frame or message size of 4 MB enforced by the WebSocket interface.
HTTP and WebSocket interfaces can now return warnings
The JSON response for a recognition request can now include an array of warning messages for invalid query parameters or JSON fields that are included with a request. Each element of the array is a string that describes the nature of the warning followed by an array of invalid argument strings. For example, "warnings": [ "Unknown arguments: [u'{invalid_arg_1}', u'{invalid_arg_2}']." ]. For more information, see the API & SDK reference.
Beta Apple iOS SDK is deprecated
The beta Watson Speech Software Development Kit (SDK) for the Apple® iOS operating system is deprecated. Use the Watson SDK for the Apple® iOS operating system instead. The new SDK is available from the ios-sdk repository in the watson-developer-cloud namespace on GitHub.
WebSocket interface can produce delayed results
The WebSocket interface can take minutes to produce final results for a recognition request for an especially long audio file. For the WebSocket interface, the underlying TCP connection remains idle while the service prepares the response. Therefore, the connection can close due to a timeout. To avoid the timeout with the WebSocket interface, request interim results (\"interim_results\": \"true\") in the JSON for the start message to initiate the request. You can discard the interim results if you do not need them. This issue will be resolved in a future update.

19 January 2016

New profanity filtering feature
The service was updated to include a new profanity filtering feature on January 19, 2016. By default, the service censors profanity from its transcription results for US English audio. For more information, see Profanity filtering.

17 December 2015

New keyword spotting feature
The service now offers a keyword spotting feature. You can specify an array of keyword strings that are to be matched in the input audio. You must also specify a user-defined confidence level that a word must meet to be considered a match for a keyword. For more information, see Keyword spotting. The keyword spotting feature is beta functionality.
New word alternatives feature
The service now offers a word alternatives feature. The feature returns alternative hypotheses for words in the input audio that meet a user-defined confidence level. For more information, see Word alternatives. The word alternatives feature is beta functionality.
New UK English and Arabic models
The service supports more languages with its transcription models: en-UK_BroadbandModel and en-UK_NarrowbandModel for UK English, and ar-AR_BroadbandModel for Modern Standard Arabic. For more information, see Previous-generation languages and models.
New session_closed field for session-based methods
In the JSON responses that it returns for errors with session-based methods, the service now also includes a new session_closed field. The field is set to true if the session is closed as a result of the error. For more information about possible return codes for any method, see the API & SDK reference.
HTTP platform timeout no longer applies
HTTP recognition requests are no longer subject to a 10-minute platform timeout. The service now keeps the connection alive by sending a space character in the response JSON object every 20 seconds while recognition is ongoing. For more information, see Timeouts.
Rate limiting with curl command is no longer needed
When you use the curl command to transcribe audio with the service, you no longer need to use the --limit-rate option to transfer data at a rate no faster than 40,000 bytes per second.
Changes to HTTP error codes
The service no longer returns HTTP status code 490 for the session-based HTTP methods GET /v1/sessions/{session_id}/observe_result and POST /v1/sessions/{session_id}/recognize. The service now responds with HTTP status code 400 instead.

21 September 2015

New mobile SDKs available

Two new beta mobile SDKs are available for the speech services. The SDKs enable mobile applications to interact with both the Speech to Text and Text to Speech services.

  • The Watson Speech SDK for the Google Android™ platform supports streaming audio to the Speech to Text service in real time and receiving a transcript of the audio as you speak. The project includes an example application that showcases interaction with both of the speech services. The SDK is available from the speech-android-sdk repository in the watson-developer-cloud namespace on GitHub.
  • The Watson Speech SDK for the Apple® iOS operating system supports streaming audio to the Speech to Text service and receiving a transcript of the audio in response. The SDK is available from the speech-ios-sdk repository in the watson-developer-cloud namespace on GitHub.

Both SDKs support authenticating with the speech services by using either your IBM Cloud service credentials or an authentication token. Because the SDKs are beta, they are subject to change in the future.

New Brazilian Portuguese and Mandarin Chinese models

The service supports two new languages, Brazilian Portuguese and Mandarin Chinese, with the following models:

  • Brazilian Portuguese broadband model (pt-BR_BroadbandModel)
  • Brazilian Portuguese narrowband model (pt-BR_NarrowbandModel)
  • Mandarin Chinese broadband model (zh-CN_BroadbandModel)
  • Mandarin Chinese narrowband model (zh-CN_NarrowbandModel)

For more information, see Previous-generation languages and models.

New support for audio/ogg;codecs=opus audio format

The HTTP POST requests /v1/sessions/{session_id}/recognize and /v1/recognize, as well as the WebSocket /v1/recognize request, support transcription of a new media type: audio/ogg;codecs=opus for Ogg format files that use the Opus codec. In addition, the audio/wav format for the methods now supports any encoding. The restriction about the use of linear PCM encoding is removed. For more information, see audio/ogg format.

New sequence_id parameter for long polling of sessions

The service now supports overcoming timeouts when you transcribe long audio files with the HTTP interface. When you use sessions, you can employ a long polling pattern by specifying sequence IDs with the GET /v1/sessions/{session_id}/observe_result and POST /v1/sessions/{session_id}/recognize methods for long-running recognition tasks. By using the new sequence_id parameter of these methods, you can request results before, during, or after you submit a recognition request.

New capitalization feature for US English transcription

For the US English language models, en_US_BroadbandModel and en_US_NarrowbandModel, the service now correctly capitalizes many proper nouns. For example, the service would return new text that reads "Barack Obama graduated from Columbia University" instead of "barack obama graduated from columbia university". This change might be of interest to you if your application is sensitive in any way to the case of proper nouns.

New HTTP error code

The HTTP DELETE /v1/sessions/{session_id} request does not return status code 415 "Unsupported Media Type". This return code is removed from the documentation for the method.

1 July 2015

The Speech to Text service is now generally available

The service moved from beta to general availability (GA) on July 1, 2015. The following differences exist between the beta and GA versions of the Speech to Text APIs. The GA release requires that users upgrade to the new version of the service.

The GA version of the HTTP API is compatible with the beta version. You need to change your existing application code only if you explicitly specified a model name. For example, the sample code available for the service from GitHub included the following line of code in the file demo.js:

model: 'WatsonModel'

This line specified the default model WatsonModel, for the beta version of the service. If your application also specified this model, you need to change it to use one of the new models that are supported by the GA version. For more information, see the next bullet.

New token-based programming model

The service now supports a new programming model for direct interaction between a client and the service over a WebSocket connection. By using this model, a client can obtain an authentication token for communicating directly with the service. The token bypasses the need for a server-side proxy application in IBM Cloud to call the service on the client's behalf. Tokens are the preferred means for clients to interact with the service.

The service continues to support the old programming model that relied on a server-side proxy to relay audio and messages between the client and the service. But the new model is more efficient and provides higher throughput.

New model parameter for speech recognition

The POST /v1/sessions and POST /v1/recognize methods, along with the WebSocket /v1/recognize method, now support a model query parameter. You use the parameter to specify information about the audio:

  • The language: English, Japanese, or Spanish
  • The minimum sampling rate: broadband (16 kHz) or narrowband (8 kHz)

For more information, see Previous-generation languages and models.

New inactivity_timeout parameter for speech recognition

The inactivity_timeout parameter sets the timeout value in seconds after which the service closes the connection if it detects silence (no speech) in streaming mode. By default, the service terminates the session after 30 seconds of silence. The POST /v1/recognize and WebSocket /v1/recognize methods support the parameter. For more information, see Timeouts.

New max_alternatives parameter for speech recognition

The max_alternatives parameter instructs the service to return the n-best alternative hypotheses for the audio transcription. The POST /v1/recognize and WebSocket /v1/recognize methods support the parameter. For more information, see Maximum alternatives.

New word_confidence parameter for speech recognition

The word_confidence parameter instructs the service to return a confidence score for each word of the transcription. The POST /v1/recognize and WebSocket /v1/recognize methods support the parameter. For more information, see Word confidence.

New timestamps parameter for speech recognition

The timestamps parameter instructs the service to return the beginning and ending time relative to the start of the audio for each word of the transcription. The POST /v1/recognize and WebSocket /v1/recognize methods support the parameter. For more information, see Word timestamps.

Renamed sessions method for observing results

The GET /v1/sessions/{session_id}/observeResult method is now named GET /v1/sessions/{session_id}/observe_result. The name observeResult is still supported for backward compatibility.

New support for Waveform Audio File (WAV) audio format

The Content-Type header of the recognize methods now supports audio/wav for Waveform Audio File (WAV) files, in addition to audio/flac and audio/l16. For more information, see audio/wav format.

Limits on maximum amount of audio for speech recognition

The service now has a limit of 100 MB of data per session in streaming mode. You can specify streaming mode by specifying the value chunked with the header Transfer-Encoding. One-shot delivery of an audio file still imposes a size limit of 4 MB on the data that is sent. For more information, see Audio transmission.

New header to opt out of contributing to service improvements

The GET /v1/sessions/{session_id}/observe_result, POST /v1/sessions/{session_id}/recognize, and POST /v1/recognize methods now include the header parameter X-WDC-PL-OPT-OUT to control whether the service uses the audio and transcription data from a request to improve future results. The WebSocket interface includes an equivalent query parameter. Specify a value of 1 to prevent the service from using the audio and transcription results. The parameter applies only to the current request. The new header replaces the X-logging header from the beta API. See Controlling request logging for Watson services.

Changes to HTTP error codes

The service can now respond with the following HTTP error codes:

  • For the /v1/models, /v1/models/{model_id}, /v1/sessions, /v1/sessions/{session_id}, /v1/sessions/{session_id}/observe_result, /v1/sessions/{session_id}/recognize, and /v1/recognize methods, error code 415 ("Unsupported Media Type") is added.
  • For POST and GET requests to the /v1/sessions/{session_id}/recognize method, the following error codes are modified:
    • Error code 404 ("Session_id not found") has a more descriptive message (POST and GET).
    • Error code 503 ("Session is already processing a request. Concurrent requests are not allowed on the same session. Session remains alive after this error.") has a more descriptive message (POST only).
    • For HTTP POST requests to the /v1/sessions and /v1/recognize methods, error code 503 ("Service Unavailable") can be returned. The error code can also be returned when you create a WebSocket connection with the /v1/recognize method.