deepdml

/

whisper-large-v3-turbo-ig-mix-norm

like 0

Hi! I have a question regarding the fine-tuning strategy used for Whisper models on specific languages.

✅ What I observed

• For languages like Igbo and Arabic, there are separate fine-tuned Whisper checkpoints available depending on the model size.
• Based on the training parameters, it seems like these were done using full fine-tuning, not LoRA.
  (This is just my assumption — not fully confirmed.)
• In many multilingual scenarios, parameter-efficient methods like LoRA are preferred instead of updating all weights.

💡 Questions

1. Does full fine-tuning offer a clear performance advantage over LoRA, especially for low-resource languages?
2. In terms of hallucination or overfitting,
   • Could full fine-tuning cause the model to lose its general multilingual capability or become too specialized in a single language?
3. Are there any known examples or references where LoRA was specifically used for single-language Whisper fine-tuning instead of full fine-tuning?
   • If so, I would love to check them out.