speech31

/

wavlm-large-english-phoneme

like 0

Hi

Really glad to see your work. I am trying to finetune wav2vec2-large-960h on librispeech (But on phoneme recognition task, those phoneme labels are generated from scripts.) But I always meet a problem that the loss will not go down after the first 200 steps and the output will just be UNK. If possible, May I take a look at your code of finetuning that?

Warm regards,
Xiaoliang