Piper Hebrew β noa (medium)
A Piper voice for Hebrew: noa, a natural female voice, plus an English speaker (ljspeech) in the same model. Fast, CPU-friendly, standard Piper ONNX.
Type plain Hebrew β phonemes follow renikud (stressed IPA).
Files
| File | |
|---|---|
he_IL-noa-medium.onnx |
Piper voice (ONNX, 22.05 kHz) |
he_IL-noa-medium.onnx.json |
Piper config (phoneme_type: hebrew, speakers) |
noa-ljspeech-step20322.pt |
PyTorch checkpoint (generator + discriminator) for further fine-tuning |
Speakers
| Speaker | Id | Language |
|---|---|---|
noa |
0 (default) | Hebrew |
ljspeech |
1 | English |
Usage
echo "ΧΧΧ Χ¦Χ€Χ ΧΧ‘Χ¨Χ ΧΧ¨ΧΧ ΧΧΧ Χ©Χ¦Χ€Χ ΧΧΧΧ." \
| uvx --from piper-tts piper -m he_IL-noa-medium.onnx -f noa.wav
phoneme_type is hebrew: Piper converts plain Hebrew to stressed IPA with
renikud, the phoneme convention this model is trained on.
Training
- Architecture β Piper VITS, medium.
- Init β
en_US-ljspeech-medium(trained from scratch on public-domain LJSpeech), converted to multi-speaker with the speaker conditioning zero-initialized. - Data
- noa β 24 h of Hebrew speech generated with Chatterbox, Knesset sentences, Whisper-filtered.
- ljspeech β LJSpeech (public domain), ~24 h.
- Steps β 20,322 (best held-out mel L1 0.405).
- Code β piper-tts-finetune.
License
CC BY 4.0.
- Downloads last month
- 181