huggingFaceModel

[Jefresongil]

Hi Nassif, I've already done some audio transcriptions using the template "huggingFaceModel = jonatasgrosman/wav2vec2-xls-r-1b-portuguese" and indeed it offers a better result. But I would like to know if there is any previous action that we can do in these audios with a view to improvements, since many times, wiretaps have a low audio quality. Could Microsoft or Google's models represent any gain in these cases?

[lfcnassif]

Hi @Jefresongil. We don't have any pre processing step to enhance audio quality. 1,5 year ago, wav2vec2 quality was similar to Google's or Microsoft's services, but algorithms evolved a lot since then. So yes, Google and Microsoft can provide better results today. We plan to offer Whisper in 4.2 version, to be released next year. If you can't wait, there is a draft code in #1335 you can try.