search

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

Question about naming of `lm_score` parameter in `decode_logits` #63

Closed patrickvonplaten closed 2 years ago

patrickvonplaten commented 2 years ago

Might not be super important, but I was a bit surprised by the naming of the lm_score variable in https://github.com/kensho-technologies/pyctcdecode/blob/9071d5091387579b4722cfcbe0c8597ad0b16227/pyctcdecode/decoder.py#L498

the lm_score corresponds to the normalized sum of both the logit_score and the lm_score no? See: https://github.com/kensho-technologies/pyctcdecode/blob/9071d5091387579b4722cfcbe0c8597ad0b16227/pyctcdecode/decoder.py#L326

Shouldn't it rather be called merged_score or something?

lopez86 commented 2 years ago

You're correct here. I added a small PR to clarify this.