reenigne314

/

chatterbox-indic-lora

like 6

Hello reenigne314, really impressive work. The warm initialization of embedding layers from similar sounding tokens is very clever.

I don’t have much experience with LoRA adaptations yet , most of my experience has been with full finetuning. So I’ve been trying to wrap my head around this by reading your Substack post and the HF repo.

I’d really like to experiment with this further using larger multi speaker datasets like IndicVoices, and see how well the generalization side scales.

Would you be open to sharing the training script or even a rough training pipeline? Especially the T3 LoRA setup would be super helpful for reproducing and building on this work.