tiny-voca-fr

A small standard French text-to-speech model: two voices, 4.5M parameters, 18 MB with the waveform generator included. It runs on a CPU or in the browser.

French port of tiny-voca. Code · Demo

Voices

Voice Description Samples
FR_F1 Female, mid-forties, warm and clear samples/FR_F1-*.wav
FR_M1 Male, mid-fifties, deep and calm samples/FR_M1-*.wav

Files

File
tiny-voca-fr.pt PyTorch inference checkpoint
config.json Model configuration
onnx/model.onnx ONNX export (phoneme ids in, 24 kHz audio out)
onnx/model.json Voices, phoneme symbols and ONNX inputs
samples/ Four sentences per voice

Usage

git clone https://github.com/DanielSuncost/tiny-voca-fr.git && cd tiny-voca-fr
pip install -e . && sudo apt install espeak-ng
hf download DanielSuncost/tiny-voca-fr --local-dir weights
TINY_VOCA_LANGUAGE=fr python -m tiny_voca.infer --checkpoint weights/tiny-voca-fr.pt \
  --voice FR_F1 --text "Bonjour, comment allez-vous ?" --output bonjour.wav

ONNX inputs: tokens and stress (int64 [1, n], from the eSpeak NG fr-fr front end in the code repository), voice (int64 [1], 0 = FR_F1, 1 = FR_M1), speed, pitch_shift, pitch_range (float32 [1]). Outputs: audio (float32, 24 kHz) and durations (frames per token).

Training data

Synthetic. Qwen3-TTS 1.7B Base (Apache-2.0) read 9,200 sentences in two voices designed with Qwen3-TTS VoiceDesign: 8,000 from the Common Voice sentence collector (CC0) and 1,200 templates for numbers, prices, times and dates. Clips were screened with Whisper large-v3-turbo and aligned with wav2vec2; 14.1 hours were used for training (100,000 steps).

Results

Whisper large-v3-turbo word error rate: 0% for both voices on unseen pronunciation test sentences (liaison, h aspiré, nasal vowels, tu/tout, œuf/œufs); 0% (FR_F1) and 4.5% (FR_M1, same as the teacher) on held-out test sentences.

Limits

  • The male voice is less natural than the female voice. Its vocoder now mixes noise into half of the pulse power above 1 kHz (mixed excitation, no extra parameters), which removes most of the earlier mechanical buzz; a faint trace can remain. Use the current code with these weights: older code loads them without the mix.
  • eSpeak NG misses some liaisons (comment‿allez), adds a few false ones, and reads fils (son) as [fil].
  • Two fixed voices. No voice cloning.

License

Apache-2.0.

Downloads last month
38
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for DanielSuncost/tiny-voca-fr

Base model

n4ze3m/tiny-voca
Quantized
(3)
this model