Understanding customization

Status and support

The following status and support information applies to customization:

• Customization is available for all languages.
• IBM Cloud You must have the Standard or Premium pricing plan to use customization. Users of the Lite plan cannot use the customization interface. For more information, see the Text to Speech service in the IBM Cloud® Catalog.
• IBM Cloud Premium customers can work with IBM to train a new custom voice for their specific application needs. To request a custom voice or for more information, complete and submit this IBM Request Form.

How customization works

The customization interface of the Text to Speech service creates a dictionary of words and their translations for a specific language. This dictionary is referred to as a custom model. Each custom entry in a custom model consists of a word/translation pair. A word's translation tells the service how to pronounce the word when it occurs in input text.

The customization interface provides methods to create and manage your custom models, which the service stores permanently. After you create a custom model, you can use it during synthesis with any version of the /v1/synthesize method. When the service synthesizes input text, it determines the pronunciation of words that appear in the custom model by applying their translations either directly or indirectly. Because you create a custom model for a specific language, a custom model can be used with any voice that is available in that language.

You specify the translation for a word in a custom model as a sounds-like translation or a phonetic translation. You can use both methods for entries in the same custom model, and you can mix the two methods within the same translation. A number of rules and guidelines apply to custom entries. For more information, see Rules for creating custom entries.

Sounds-like translation

Sounds-like translation uses the service's regular pronunciation rules to represent the pronunciation of a target word indirectly. A sounds-like translation is formed from the regular pronunciations of one or more other words. The service first substitutes the specified translation for any occurrence of the word that appears in the input text. It then applies its regular pronunciation rules to the translation, converting the translation to its phonetic representation to obtain the pronunciation.

For example, the service's regular pronunciation rules properly translate many common abbreviations and acronyms. The service pronounces the abbreviation cm as centimeter. It pronounces less common abbreviations letter by letter. For instance, the service pronounces the string Str (an abbreviation for street) as S T R, with each letter pronounced individually. You can use the sounds-like method to specify the translation street for the string Str.

Another example of an acronym is the word IEEE, which stands for Institute of Electrical and Electronic Engineers. By default, the service pronounces this acronym as I E E E. But the acronym is commonly pronounced I triple E, which you can easily define by using the simple sounds-like translation I triple E. If the word IEEE appears in your custom model with this translation, the service substitutes each occurrence of the word with the translation. It then applies its regular pronunciation rules to the individual words I, triple, and E to yield the common pronunciation.

You can apply the sounds-like method to more than just abbreviations and acronyms. It works equally well for complex or unusual words. For example, the following pair of sounds-like translations yields correct pronunciations for unusual words that are handled imperfectly by the service's regular pronunciation rules. Finding proper translations for such words can be more challenging than for simple abbreviations. The following examples use the regular pronunciation rules to alter the words' spelling for translation:

• Word: ayurvedic, Translation: aayervedic
• Word: gastroenteritis, Translation: gastro enteritis

As these examples show, developing sounds-like translations can be more trial-and-error than formulaic. You create a candidate translation based on your intuition and experience with the service. You then synthesize the word for the candidate translation as input text and listen to the resulting audio. If you are satisfied with the pronunciation, you can use the translation in your custom model; otherwise, you modify the translation and test it again.

Phonetic translation

The sounds-like method is a relatively simple and useful way of achieving a pronunciation. But it is not always possible to develop sounds-like translations. The direct alternative, the phonetic method, might appear to be more complicated and time-consuming, but it can achieve the pronunciation of any word.

Phonetic translation specifies a pronunciation in terms of phoneme symbols, syllable stress marks, and optional syllable boundaries that override the service's regular pronunciation rules. You specify a phonetic translation in one of the following formats:

• The standard International Phonetic Alphabet (IPA) representation
• The proprietary IBM Symbolic Phonetic Representation (SPR)

In either case, you specify a translation by using a specific phoneme format that is based on the Speech Synthesis Markup Language (SSML). SSML is an XML-based markup language that provides annotations of text for speech-synthesis applications. You specify the phonetic translation for a word by using the SSML <phoneme> element:

<phoneme alphabet="{ipa | ibm}" ph="{translation}"></phoneme>

The alphabet attribute specifies the phonetic representation type: ipa or ibm. The ph attribute specifies the phonetic translation string.

For example, consider the word trinitroglycerin. The service's regular pronunciation rules produce a pronunciation that differs from the one commonly used by chemists and physicians. The correct pronunciation can be achieved with a phonetic translation:

• IPA: tɹaɪnˈaɪtɹəglɪsəɹɨn
• SPR: trYn1YtrxglIsxrXn

In these examples, the phonetic translation string is composed of phoneme symbols and a single primary stress mark. The primary stress mark is represented by ˈ in IPA and by 1 in SPR. It is placed just before the symbol for the stressed vowel in both cases. Although the examples do not show it, you can also specify syllable boundaries and secondary stress positions in a phonetic translation. These elements are not required and normally are not needed to achieve a pronunciation. As with sounds-like translations, you can compose a phonetic translation from multiple strings that are delimited by spaces.

You can also specify IPA translations as IPA Unicode values. For more information, see Understanding phonetic symbols and the language-specific tables on the pages that are referred to from Phonetic symbols for supported languages. For an example translation that uses IPA Unicode values, see The phoneme element.

Mixed sounds-like and phonetic translation

You can mix the sounds-like and phonetic methods in the same translation. This feature can reduce the work that is involved in composing a translation.

For instance, assume that you used the sounds-like method to get part of a word pronounced satisfactorily. But you now need to fine-tune the remaining elements of the word. You can use the phonetic method to specify the difficult aspects of the word. The following example applies mixed translation to the word trinitroglycerin:

try<phoneme alphabet="ipa" ph="nˈaɪtɹəglɪsəɹɨn"></phoneme>