README.md

metadata

language:
  - gu
license: cc-by-nc-4.0
tags:
  - text-to-speech
  - tts
  - gujarati
  - f5-tts
  - flow-matching
  - indic
pipeline_tag: text-to-speech
base_model: SWivid/F5-TTS
datasets:
  - Arjun4707/gu-hi-tts

F5-TTS — Gujarati fine-tune

A fine-tuned F5-TTS model for Gujarati text-to-speech using flow matching.

Model details

Attribute	Value
Base model	SWivid/F5-TTS (F5TTS_v1_Base)
Architecture	Diffusion Transformer with ConvNeXt V2
Language	Gujarati (gu)
Training	150K steps (21 epochs) on NVIDIA L4 (24 GB)
Tokenizer	Custom (extended for Gujarati characters)

Training data

Fine-tuned on Gujarati clips from Arjun4707/gu-hi-tts (~36K clips after CPS + duration filtering).

Data source: Audio clips scraped from publicly available YouTube videos. Preprocessed to 24kHz mono PCM-16, silence-trimmed, peak-normalized to -3 dBFS.

Known limitations

• Single-speaker clips produce good quality; multi-speaker clips in training data caused stopping/blabbering artifacts
• Total generation length (prompt + generated) capped at ~30 seconds

Training code

Full training pipeline and troubleshooting: BhammarArjun/TTS_2_training

License

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

The base F5-TTS model is CC-BY-NC-4.0 (trained on the Emilia in-the-wild dataset). Our fine-tuning data was also sourced from YouTube audio.

Citation

@misc{arjun2026f5ttsgu,
  title   = {F5-TTS fine-tuned for Gujarati},
  author  = {Arjun Bhammar},
  year    = {2026},
  url     = {https://huggingface.co/Arjun4707/F5-TTS-Gujarati}
}

Acknowledgements

• SWivid / F5-TTS for the base model and training framework