About fine-tuning this checkpoint

#2
by OREOOOO - opened

Hi @patrickvonplaten .

I tried to fine-tune this checkpoint with librispeech train-clean-100h datasets by loading it with Wav2Vec2ForCTC.from_pretrained() function.
However, WER is not as low as I expected. (about 46%, fine-tuned with same datasets with 'facebook/wav2vec2-base' : 6%)
Have you fine-tuned the wav2vec2 model with this checkpoint? If so, did you use a different configuration file?

Thank you.

Sign up or log in to comment