Open Tejaswgupta opened 6 months ago
Just to get a better idea, does your dataset include transcriptions in Hinglish? If so, you can try fine-tuning/distilling directly on these labels, and the model should learn the semantics of Hinglish directly from your training data
When doing so, you can set the language in the tokenizer to “hindi”, I think this is the best option to get language transfer from Hindi -> Hinglish (set argument --language="hi"
)
could you please guide how to finetune the distilled model directly instead of training the model itself and then do distillation process?
Here's an overview of the training methods, with a link to direct fine-tuning: https://github.com/huggingface/distil-whisper/tree/main/training#overview-of-training-methods
Does that answer your question @A-Raafat?
@sanchit-gandhi Yes. That answers it. Thank you
@sanchit-gandhi
Here's an overview of the training methods, with a link to direct fine-tuning: https://github.com/huggingface/distil-whisper/tree/main/training#overview-of-training-methods
Hello, I have a question. In this link, it points to the fine-tuning guide at https://huggingface.co/blog/fine-tune-whisper. However, it seems that the fine-tuning code does not support timestamp training and conditioning on previous labels in distillation (as seen in run_distillation.py with timestamp_probability and condition_on_prev_probability). I believe these two features are crucial for ASR tasks and have a significant impact on the performance. So, I would like to ask if I can use similar options like timestamp_probability and condition_on_prev_probability if I only want to do fine-tuning. Thank you.
Is it possible to fine-tune Whisper/Distil-Whisper to achieve mixed speech transcription like Hindi+English in a single sentence which is common in casual conversations. Has anyone tried this before? Would training on a mixture of Hindi datasets and English datasets work? I recently used a fine-tuned Whisper for ASR and it ended up hallucinating and adding additional text which I haven't been able to fix yet.