Add lm decode for the Python API. #353
Conversation
Shouldn't double lm_log_prob when merge same prefix path
| @@ -17,11 +17,6 @@ void Hypotheses::Add(Hypothesis hyp) { | |||
| hyps_dict_[key] = std::move(hyp); | |||
| } else { | |||
| it->second.log_prob = LogAdd<double>()(it->second.log_prob, hyp.log_prob); | |||
|
|
|||
| if (it->second.lm_log_prob != 0 && hyp.lm_log_prob != 0) { | |||
There was a problem hiding this comment.
Is there a reason to remove it?
There was a problem hiding this comment.
here two decoding paths that have the same prefix and we need to merge them into one path,
we can refer the Python implementation at here :https://github.com/k2-fsa/icefall/blob/2b3c5d799f3a585dc22071a9148424ff77aefd47/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py#L843
python implementation perform (path1_am_prob * path1_lm_prob ) + (path2_am_prob * path2_lm_prob )
Since paths with the same prefix must have the same language model probability, we can simplify the calculation to 【path1_lm_prob * ( path1_am_prob + path2_am_prob )】
If we don't remove it , we actually perform (path1_am_prob + path2_am_prob ) * ( path1_lm_prob + path2_lm_prob)
simplified to 【 2 * path1_lm_prob * ( path1_am_prob + path2_am_prob )】
I have tested two opensource test set and compared result :
| 双人对话 | 中英混句 | |
|---|---|---|
| lm-shallow_fusion_0.1 | 15.18% | 22.26% |
| lm-shallow_fusion_0.1_bug_fix | 15.03% | 21.99% |
test set reference ：
https://www.openslr.org/123/
https://magichub.com/cn/competition/code-switching-asr-challenge/
There was a problem hiding this comment.
Thanks for the detailed explanation!
|
Thanks for your contribution! |
No description provided.