-
I'm relatively new to ASR and to speechbrain, so pardon me if the questions are stupid. My search for CTC decoders led me to https://speechbrain.readthedocs.io/en/latest/API/speechbrain.decoders.ctc.html
Thanks in advance. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
Dear @tshmak, Thanks for your question. First, I'd like to say that You should use Also, I would like to invite you to have a look at https://github.com/speechbrain/speechbrain/blob/develop/speechbrain/decoders/ctc.py , this is the new pure CTC interfaces that implements CTC beam searchers (frame-synchronous CTC). One CTC Beam searcher (https://github.com/speechbrain/speechbrain/blob/develop/speechbrain/decoders/ctc.py#L1912) is a wrapper around torchaudio CTC decoder (https://pytorch.org/audio/main/tutorials/asr_inference_with_ctc_decoder_tutorial.html) and leave you the possibility to have constraint search. Additionally, if you want, you can also train a n-gram model on your constrain vocab, and then use it with any of our CTC decoders in (https://github.com/speechbrain/speechbrain/blob/develop/speechbrain/decoders/ctc.py#L1912) as it is very similar to having a constraint dict (the ngram will penalise unknown tokens/words). Hope I was helpful. Please let me know if you need aditionnal informations. These news decoders are still in beta but we got very good results on a variety of datasets. Best, Finally, we are working on k2 integration inside of speechbrain (#2345 should be merged very very soon) and it leaves also the possibility to the user to perform constraint search. |
Beta Was this translation helpful? Give feedback.
Thanks Adel! That's very clear.