Arjun4707/chatterbox-gujarati

Chatterbox — Gujarati fine-tune

A fine-tuned Chatterbox TTS model for Gujarati text-to-speech with voice cloning and emotion control.

Model details

Attribute	Value
Base model	ResembleAI/chatterbox (Multilingual)
Architecture	CosyVoice 2.0 — LLaMA-based T3 (0.5B params)
Language	Gujarati (gu)
Training hardware	NVIDIA L4 (24 GB VRAM)
Fine-tuning repo	gokhaneraslan/chatterbox-finetuning
Vocab extension	2454 → 2514 tokens (+60 Gujarati characters)

Training data

Fine-tuned on Gujarati clips from Arjun4707/gu-hi-tts (~33,851 clips after CPS + diarization filtering).

Data source: Audio clips scraped from publicly available YouTube videos. Preprocessed with speaker diarization (pyannote-audio) to keep only single-speaker clips, CPS filtered (4-25 chars/sec), duration filtered (2-20s).

Known limitations

• Very short utterances (1-3 words) produce poor quality — architecture needs minimum ~5 words
• Medium to long sentences (5-30 words) produce good quality with clear Gujarati pronunciation
• Not suitable for sub-2s audio generation

Training code

Full training pipeline and troubleshooting: BhammarArjun/TTS_4_training

License

CC-BY-NC-4.0 — Non-commercial use only.

The base Chatterbox model is MIT-licensed, but this fine-tuned version uses YouTube-sourced training data. To be transparent and responsible about data provenance, we apply CC-BY-NC-4.0 to this fine-tuned version.

Citation

@misc{arjun2026chatterboxgu,
  title   = {Chatterbox fine-tuned for Gujarati},
  author  = {Arjun Bhammar},
  year    = {2026},
  url     = {https://huggingface.co/Arjun4707/chatterbox-gujarati}
}

Acknowledgements

• Resemble AI for the Chatterbox TTS model
• gokhaneraslan for the fine-tuning framework