Request for Training Scripts & Indian Language Fine-tuning Support
Hello IndexTTS Team,
First, thank you for releasing IndexTTS-2 - it's truly a breakthrough in emotionally expressive TTS with impressive capabilities in duration control and zero-shot voice cloning.
1.Purpose of Request
I'm interested in fine-tuning IndexTTS-2 to support Indian languages, particularly Tamil, while preserving all the advanced features your model offers (emotion disentanglement, duration control, cross-language voice cloning, etc.).
- Fine-tuning Guide for New Languages
Could you provide documentation or guidance on:
How to extend the tokenizer for non-Latin scripts (Tamil uses Dravidian script with 247+ characters including conjuncts)
Recommended fine-tuning approach (full fine-tuning vs. LoRA/adapter-based methods)
Minimum dataset requirements for maintaining quality
How to preserve emotion disentanglement and duration control features when adapting to new languages
- Multilingual Support Roadmap
There's active community interest in multilingual support. Are there plans to:
Officially support Indian languages?
Release a multilingual checkpoint?
Provide parameter-efficient fine-tuning methods (LoRA adapters)?
with Technical Details
Please fine tune for Hindi and Urdu as well. Thank you.