Support text prompts

[mayfer]

I would like to use a text prefix (a.k.a. prompt) to the transcription so that I can provide domain-specific context to the transcription.

This significantly improves performance on Whisper for my use cases. Would love to have it on parakeet as well

[senstella]

Hello. Thank you for the suggestion. However, the Parakeet model does not have text prompt in the sense like Whisper does. While Whisper pretrained with various prompts to support this use-case, to the best of my knowledge, Parakeet model does not utilize such training objective and directly learns to transcribe from zero-prompt. So language forcing like #31 does not work.

One possible approach is: initialize decoder's LSTM state with prompt. However, performance-wise, this isn't in training objective so further validations are required. Meanwhile, I think creating issue in the original Parakeet model repository directly could get much more helpful information from the Parakeet team directly! https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3/discussions