Poor Result on Using LM boosted Wav2Vec 2.0

kensho-technologies / pyctcdecode

A fast and lightweight python-based CTC beam search decoder for speech recognition.

Apache License 2.0

416 stars 89 forks source link

Poor Result on Using LM boosted Wav2Vec 2.0 #53

Closed rohitashwa1907 closed 2 years ago

rohitashwa1907 commented 2 years ago

Hey guys, While using pyctcdecode package with Wav2Vec2 from huggingface, I came across a strange issue. It combines many words generated by a normal acoustic model. Eg: Acoustic Model: " I will definitely like to visit that place". LM Model: " I will definitely liketo visitthat place".

gkucsko commented 2 years ago

Hey, my guess would be a vocabulary mismatch. Do you have an example to reproduce? happy to look

rohitashwa1907 commented 2 years ago

Silly Me ! I made a small mistake which led to vocab mismatch. Thanks a lot for your time. Closing the issue.

gkucsko commented 2 years ago

Great :)