tontate/f5-tts-quebec-french-finetune

F5-TTS — Quebec French (Québécois) finetune

A zero-shot voice-cloning TTS checkpoint specialized for Quebec French. Give it a short reference clip of any voice and it clones that voice while applying Québécois pronunciation, prosody, and vocabulary.

Finetune of F5-TTS (F5TTS_Base architecture, vocos vocoder), warm-started from RASPIAUDIO/F5-French-MixedSpeakers-reduced and trained on a Québécois-accent-tagged subset of Mozilla Common Voice (fr).

License note: the warm-start base checkpoint is CC-BY-NC-4.0. This finetune inherits that restriction — non-commercial use only. The training data itself is CC0-1.0.

Quickstart

pip install f5-tts

f5-tts_infer-cli \
  --model F5TTS_Base \
  --ckpt_file model_quebec_v1.safetensors \
  --vocab_file vocab.txt \
  --ref_audio your_reference.wav \
  --ref_text "Transcript of the reference clip." \
  --gen_text "Je m'en vais au dépanneur chercher de la liqueur pis des chips."

Download the two required files with:

huggingface-cli download tontate/f5-tts-quebec-french-finetune \
  model_quebec_v1.safetensors vocab.txt

Tips: use a clean 3–10 s reference clip; leave --ref_text "" to auto-transcribe it. The checkpoint clones any voice — a Québécois-accented reference gives the most consistent results.

Audio samples

Generated with the final checkpoint (voice qc_a); more in samples/.

« Il faut que je stationne mon char devant le dépanneur. »

« Fait qu'on a fini par jaser toute la soirée sur la galerie. »

« Ça coûte quatre-vingt-dix-sept dollars et trente-cinq sous. »

Files

File	Size	Use
model_quebec_v1.safetensors	1.35 GB	Recommended for inference — EMA weights only
model_last.pt	5.39 GB	Full training checkpoint (model + EMA + optimizer) — for resuming training
checkpoints/model_9000.pt, checkpoints/model_12000.pt	5.39 GB each	Intermediate training checkpoints
vocab.txt	14 kB	Tokenizer vocab (unchanged from base / F5-TTS default)
scripts/prep_common_voice_qc.py	—	Dataset preparation script (reproducible)
dataset/processed_qc_dataset.tar.gz	—	The processed training set
samples/	—	Eval-sentence audio (2 voices) + per-checkpoint training-trajectory samples

model_quebec_v1.safetensors contains exactly the EMA weights of model_last.pt (bitwise-identical extraction, verified), repackaged for safe, fast loading.

Training

• Data: Common Voice fr (Scripted Speech 26.0), CC0-1.0, filtered to clips with Québécois/Canadien accent metadata. Final set: 11,542 clips, 281 speakers, 12.01 hours (after resample to 24 kHz mono, silence trim, dedup, 1–15 s length filter). Prep script included in scripts/.
• Run: 50 epochs / 67,750 updates with the F5-TTS finetuning trainer on a single RTX A5000 (24 GB), warm-started from the base checkpoint with its vocab unchanged.
• Training loss plateaued early (≈ epoch 2); trajectory samples for every 1,000 updates are in samples/training_trajectory/ if you want to compare checkpoints.

Evaluation

Evaluated on a 50-sentence Québécois test battery targeting dialect-specific phenomena (affrication, vowel laxing, /a/–/ɑ/ contrast & diphthongs, QC lexicon/elisions, numbers, question particles & prosody), synthesized with 3 reference voices, and compared against the warm-start base model (RASPIAUDIO/F5-French-MixedSpeakers-reduced, France-accented). Scripts and the sentence list are in scripts/eval/.

Metric	Base (France)	This model (Québec)
WER ↓ (faster-whisper large-v3)	0.067	0.077
CER ↓	0.041	0.043
Speaker similarity ↑ (ECAPA cosine)	0.479	0.675
Accent probe: mean P(Québec) ↑	0.17	0.40

• Intelligibility is preserved: WER/CER are comparable; the finetune's small gap concentrates in QC lexicon and elisions ("t'sé", "pis") that the ASR normalizes to standard French spellings.
• Voice cloning improves markedly on Québécois reference voices (+0.20 cosine), as expected from in-domain training.
• Accent probe: a logistic-regression classifier over mid-layer XLS-R-300m embeddings, trained to separate real Québec-tagged vs. European-tagged (France/Belgium/Switzerland) Common Voice fr clips — held-out accuracy 0.82 on speaker-disjoint test speakers. Applied to the generated audio, it rates the probability the audio sounds Québécois.
• Human check: informal native-listener review judged the output distinctly Québécois (satisfactory) rather than European French.

Caveats: reference clips are drawn from the training corpus (speaker-similarity numbers are in-domain); the accent probe is a proxy, not a listening test; number-heavy sentences are hardest for both models — write numbers out in words.

Limitations

• Non-commercial license (inherited from the base checkpoint).
• 12 hours of training data: rare words, heavy anglicisms, and long-form prosody may be less stable; write numbers out in words for best results.
• Common Voice accent tags are self-reported; the training accent mix is broadly Québécois/Canadian rather than a single regional variety.