[Python] Paramaters when working on audio transcription of low-resource languages with Whisper

Am working on an audio transcription tool for a low-resource language (Ukrainian), and the results are unsatisfactory. I do understand that choosing a larger model and fine-tuning can solve my problems, but at the moment I do not have any budget for that.

As a solution (for some time), I was thinking play avec my transcribe params and additionally applying RAG (to ensure Whisper gets all the named entities correctly).

Below I add my current params, yet, unfortunately, the outcomes are not satisfactory. Could anyone help me with the parameters?



Continue reading...