patrickvonplaten / Wav2Vec2_PyCTCDecode

Small repo describing how to use Hugging Face's Wav2Vec2 with PyCTCDecode
110 stars 17 forks source link

Not able to reproduce greedy search results with beam_width=1 #6

Closed ashu5644 closed 2 years ago

ashu5644 commented 2 years ago

Hi @patrickvonplaten , I am following this tutorial https://huggingface.co/blog/wav2vec2-with-ngram . I have tried with some beam width variations, but getting large WER's compared to greedy search. So, I tried it with passing beam_width=1 in decode function of pyctcdecode but getting similar WER to beam_width = {4,10, default=100}. Is it normal behaviour? I suppose , beam_width=1 should give similar results as of greedy decoding.

patrickvonplaten commented 2 years ago

Interesting! To be honest I don't really know what the best beam width is...maybe ask on https://github.com/kensho-technologies/pyctcdecode

ashu5644 commented 2 years ago

Ok, Thanks. I will check.