Wav2Vec ASR with SentencePiece vocabulary

164 views

Skip to first unread message

Igor Shalyminov

unread,

Feb 28, 2022, 3:56:59 PM2/28/22

to fairseq Users

Hi all,

How do I run Wav2Vec2-ASR models (AudioFinetuning task) with SPM vocabulary instead of character-based?

Are there any examples in the repo?

Alexei Baevski

unread,

Feb 28, 2022, 3:59:32 PM2/28/22

to Igor Shalyminov, fairseq Users

Create a dict.spm.txt and train.spm (or whatever you choose) and then set task.labels=spm. It works fine using standard recipes, but it will not work for very low resource splits like 10m or 1h

--
You received this message because you are subscribed to the Google Groups "fairseq Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to fairseq-user...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/fairseq-users/79e7fa9e-b979-4e95-b806-0463a17cd8d8n%40googlegroups.com.

Reply all

Reply to author

Forward

0 new messages