F5-TTS Malayalam v2

A F5-TTS model fine-tuned on Malayalam speech, trained on transcripts and audio from my own YouTube channel — a Malayalam-language ML/Data Science education channel. The goal was to build a TTS model that speaks Malayalam in my own voice and speaking style.

Try it out in the demo Space.

Model Details

• Base model: F5-TTS
• Language: Malayalam (ml)
• Fine-tuning data: ~15.4 hours of paired audio and transcripts, sourced from my own YouTube videos
• Voice: Fine-tuned to approximate my own voice and speaking style
• License: cc-by-nc-4.0 (non-commercial — inherits F5-TTS's licensing terms; check the base model license before any commercial use)

Files

• model.pt — fine-tuned checkpoint
• vocab.txt — vocabulary file used during fine-tuning

Training

This was a personal, low-budget experiment:

• Fine-tuned entirely on free GPU compute and storage on Kaggle, often running into the weekly quota limits
• No large-scale infra or paid compute was used
• Training data was self-sourced: audio and transcripts from my own uploaded YouTube videos

Known Limitations

This is an early, imperfect model and a fun side project rather than a production-ready system. Known issues include (but aren't limited to):

• Occasional pronunciation and prosody errors
• Inconsistent output quality across different input text
• Artifacts/imperfections inherited from limited training data (15.4 hours) and constrained compute budget

These are being addressed in future iterations.

Usage

See the F5-TTS repository for inference instructions, or use the hosted demo Space to try it directly in the browser.

Acknowledgements

• F5-TTS authors for the base model and training framework
• Kaggle for free GPU compute

Citation

If you use this model, please consider linking back to this repository and the original F5-TTS work.