CODEWITHSUNDEEP
ibm-cloudtext-to-speechwatson
Leandro David
Leandro David

Reputation: 597

Watson Text To Speech Roman Numerals recognized only up to XXXIX (39) in pt-BR

Doing some tests to synthesize parts of a book in portuguese (pt-BR) I noted that chapter names containing Roman numerals are not recognized as numbers if the numeral is above XXIX (39).

Like number L (50) or LX (60) will not be properly synthesized as a number and it will read as the letters. Using "say-as" SSMLS tag causes no change.

Is this expected ? Is there any possible workaround ?

Upvotes: 0

Views: 118

Answers (2)

Leonardo Kenji Shikida
Leonardo Kenji Shikida

Reputation: 761

Another option you have is to customize your model, as described at https://console.bluemix.net/docs/services/text-to-speech/custom-intro.html#customIntro

Here goes a brief example

Use this command to create a custom model

curl -X POST -u ***:*** --header "Content-Type: application/json" --data "{\"name\":\"cURL Test\",\"language\":\"en-US\", \"description\":\"Customization test via cURL\"}" "https://stream.watsonplatform.net/text-to-speech/api/v1/customizations"

The command above will return the customization_id (in this case, e250e7ee-fbec-47e8-a1b8-59435c1b18e3)

then use this command to add a translation

curl -X PUT -u ***:*** --header "Content-Type: application/json" --data "{\"translation\":\"thirty\"}" "https://stream.watsonplatform.net/text-to-speech/api/v1/customizations/e250e7ee-fbec-47e8-a1b8-59435c1b18e3/words/XXX"

finally, try it

curl -X GET -u ***:*** --header "Accept: audio/wav" --output xxx.wav "https://stream.watsonplatform.net/text-to-speech/api/v1/synthesize?text=chapter%20XXX&customization_id=e250e7ee-fbec-47e8-a1b8-59435c1b18e3"

it will generate an audio file that says "chapter thirty"

ps. for PT-BR, see https://console.bluemix.net/docs/services/text-to-speech/custom-models.html#customModels so you can use the option "language" and translate XXX to "trinta"

Upvotes: 2

Nikolay Shmyrev
Nikolay Shmyrev

Reputation: 25220

Is this expected?

I would expect that, yes, developers are usually quite lazy to implement a generic solution.

Is there any possible workaround ?

Preprocess text yourself and convert numbers to words.

Upvotes: 2

Related Questions