Text to speech read out number in wrong language

Question

Hey team,

We've recently received complaints from users that the text-to-speech feature is reading numbers in Chinese, even though we specify the en-US voice. After some testing, we confirmed that for the following text input, the model incorrectly reads numbers in Chinese. Could someone help escalate this issue and investigate a fix?

Sample test input:

You can start by testing the example provided in the problem statement, such as nums = [2,7,11,15] and target = 9. Once you've confirmed it works, consider adding more diverse cases, like edge cases or different array sizes.

I’d like to escalate this bug to Azure support. Please let me know how to proceed.

Output audio example: https://storage.googleapis.com/leetgpt-audio/test/tts_test_en-US-NovaMultilingualNeural.wav

Answer 1

Hey LeetGPT

I have not found root cause of the issue, but I sensed that we might need a direct SSML input to get correct results.

Below SSML input works fine. Please let us know from your side

<speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis" xml:lang="en-US">
  <voice name="en-US-NovaMultilingualNeural">
    <p>You can start by testing the example provided in the problem statement, such as nums equals [2, 7, 11, 15] and target equals 9.</p>
    <p>Once you've confirmed it works, consider adding more diverse cases, like edge cases or different array sizes.</p>
  </voice>
</speak>

Thank you.