Incorrect Number Pronunciation in Piper VITS with eSpeak-NG

[Muzaffar-x]

When training a model on Piper VITS, some eSpeak-NG functions do not work correctly. Specifically, numbers 10 and above are pronounced as separate digits (e.g., "10" is read as "one zero" instead of "ten").

Interestingly, when testing a similar model VITS by Coqui with the same dataset, this issue does not occur.

Expected Behavior:
Numbers should be pronounced correctly rather than as separate digits.

Steps to Reproduce:

1. Train a model on Piper VITS using eSpeak-NG.
2. Try synthesizing numbers 10 and above.
3. Notice that instead of the expected pronunciation, numbers are read as separate digits.

Possible Causes:

• Issues with text tokenization before being fed into the model.
• Problems with eSpeak-NG integration in Piper VITS.
• Missing number processing rules in the current configuration.

Question:

How can this issue be resolved? Are there any ways to explicitly configure eSpeak-NG in Piper VITS to ensure correct number pronunciation?