I’m just curious about cuda_ctc_decoder and ctc_decoder in src/torchaudio/model/decoder/*.py (git hash ccd78ffba3208b4a9f5face4bd76e942ddc2d4a3)
actually What I want to know is that why I couldn’t get
timestepslike a output of ctc_decoder when I use cuda_ctc_decoder?
It seem a differences of beamsearch score method(Lexicon vs Prefix), but Isn’t it possible getting a timesteps with cuda_ctc_decoder?
Another question is that How can I use ctc_decoder without n-gram LM…?
like a cuda_ctc_decoder, only with token_list, Can I use ctc_decoder.py…?