Fine-tuning for Punjabi (Gurmukhi). Looking for guidance
#42
by jssaluja - opened
Congrats on the impressive release!
I'd like to fine-tune for Punjabi (Gurmukhi script), not in the 14 supported languages. I have:
~60K clean segments with verified transcriptions
~60 hours
~1 speaker
~ < 30 seconds each
I can expand this x 5 speakers
Questions:
Does adding a new language require vocabulary/tokenizer expansion, or can this be handled via fine-tuning alone?
Are there recommended data augmentation strategies to improve robustness beyond studio conditions?
Any published fine-tuning scripts, data augmentation strategies, or recommended hyperparameters?
Is Punjabi on your roadmap?
Happy to work with your team if there is guidance.
Ping, checking in for updated guidance / sample code