rhasspy

/

piper-voices

like 639

Follow

Rhasspy 315

Environment

• OS: Android 15
• numThreads: 4
• debug: false
• provider: cpu
• speed: 1.0

Description

We are developing a Finnish TTS engine application on Android 15 using sherpa-onnx. However, the available Finnish models present critical issues that prevent them from being used in production—one due to pronunciation errors.

vits-piper-fi_FI-harri-medium.tar.bz2(latest)

• Issue: Fast inference speed, but suffers from severe pronunciation errors, phoneme stretching, and unstable letter/number sequence reading.
• Test Results:
  • Päijänne: Audio output is unnaturally stretched.
  • Köyliö: Mispronounced as kyöliä.
  • TG_1 (expected: teegee (yksi)): Occasionally correct, but frequently mispronounced as teetee, teedee, or teesee.
  • The level of pronunciation is not stable. E.g. 8 is mostly always pronounced correctly, other indices varies or is always poor.
  • The pronunciation is still stretched in most indices, both in the middle and at the end.
  • Pronunciation is too fast for 1 and sometimes for 5.
  • 6 and 7 always have extra content at the end of the pronunciation, and are always stretched.

The feedback regarding the pronunciation issues with the vits-piper model is based on testing by native Finnish speakers residing in Finland.

Although I am a non-native speaker and cannot evaluate all phonetic nuances directly, I was also able to notice the poor audio quality when listening to the generated audio (the same behavior is reproducible on Hugging Face Spaces as well).

Feature Requests / Suggestions

Fix/Improve VITS Piper Finnish Model: Resolve the G2P / pronunciation errors and audio stretching artifacts in vits-piper-fi_FI-harri-medium.

Thank you!