speechlib is a library that can do speaker diarization, transcription and speaker recognition on an audio file to create transcripts with actual speaker names
MIT License
114
stars
10
forks
source link
Feature Request: An option to NOT use faster-whisper (for intel, amd, mali and other GPUs) #37
faster-whisper relies on ctransformers2, which only work on NVIDIA cards at the moment.
With normal whisper (whisper library from OpenAI or whisper imported from transformers library) you can use GPU accelerated transcription on any card that reports cuda. For example I'm using pytorch-rocm and can use cude accelerated pytorch with my AMD 6900XT.
It would be great to have an option when calling Transcriptor class to pass a variable indicating if it should use faster-whisper or regular whisper. (It would be amazing if it could autodetect, but as this things change week to week I believe a variable would be enough for this)
faster-whisper relies on ctransformers2, which only work on NVIDIA cards at the moment.
With normal whisper (whisper library from OpenAI or whisper imported from transformers library) you can use GPU accelerated transcription on any card that reports cuda. For example I'm using pytorch-rocm and can use cude accelerated pytorch with my AMD 6900XT.
It would be great to have an option when calling Transcriptor class to pass a variable indicating if it should use faster-whisper or regular whisper. (It would be amazing if it could autodetect, but as this things change week to week I believe a variable would be enough for this)
Thank you!