IPED icon indicating copy to clipboard operation
IPED copied to clipboard

Published 20 hours ago • verified publisher icon sepinf-inc

Option to use a language model with Wav2Vec2 transcription

Open lfcnassif opened this issue 2 years ago • 1 comments

This was left as a future improvement of #1214. This should be investigated: https://github.com/jonatasgrosman/huggingsound/issues/62

lfcnassif avatar Sep 11 '22 19:09 lfcnassif

I still wasn't able to use the (suggested) KenshoLMDecoder implementation for a language model from huggingsound library properly to evaluate this. But I managed to use ParlanceLMDecoder implementation together with Jonatas Grosman's fine tuned wav2vec2 large portuguese model, results were not good: image

lfcnassif avatar Jul 28 '23 20:07 lfcnassif