added end detection #46
Conversation
src/nets/e2e_asr_attctc_th.py
Outdated
| @@ -697,6 +697,27 @@ def recognize(self, h, recog_args): | |||
|
|
|||
| return y_seq | |||
|
|
|||
| # end detection desribed in Eq. (50) of | |||
| # S. Watanabe et al "Hybrid CTC/Attention Architecture for End-to-End Speech Recognition" | |||
| def end_detect(self, ended_hyps, i, M=3, D_end=np.log(1 * np.exp(-10))): | |||
There was a problem hiding this comment.
I think this function should be placed outside e2e_asr_attctc[_th].py and be shared from them like from e2e_common import end_detection. Because this function seems to be static (not use self) and be free from numpy/torch operations.
| @@ -570,8 +596,11 @@ def recognize_beam(self, h, recog_args, char_list): | |||
|
|
|||
| # preprate sos | |||
| y = self.xp.full(1, self.sos, 'i') | |||
| # maxlen >= 1 | |||
| maxlen = max(1, int(recog_args.maxlenratio * h.shape[0])) | |||
| if recog_args.maxlenratio == 0: | |||
There was a problem hiding this comment.
please add this specialized behavior in argparse help of asr_recog[_th].py . If you think it is worth being default, set it to 0.0 (now default is 0.5)
There was a problem hiding this comment.
and I also think argparse.ArgumentParser should be shared like from e2e_args import asr_recog_parser as I commented on def end_detect because chainer/pytorch tools should be consistent.
There was a problem hiding this comment.
I'll do it for end_detect, but let me make "argparse part" common later.
Actually, asr_recog.py and asr_recog_th.py are almost same, and I'm thinking of using the common asr_recog.py for both, which is more simple for me, but I'm not sure it would be applicable to asr_train.py and asr_train_th.py, and it needs some more consideration.
There was a problem hiding this comment.
OK. It may be too much to this PR.
src/nets/e2e_asr_attctc.py
Outdated
| @@ -546,6 +551,27 @@ def recognize(self, h, recog_args): | |||
|
|
|||
| return y_seq | |||
|
|
|||
| # end detection desribed in Eq. (50) of | |||
| # S. Watanabe et al "Hybrid CTC/Attention Architecture for End-to-End Speech Recognition" | |||
| def end_detect(self, ended_hyps, i, M=3, D_end=np.log(1 * np.exp(-10))): | |||
There was a problem hiding this comment.
ditto (see e2e_asr_attctc_th.py)
…enratio behaviour in the end detect case in argparse
|
@ShigekiKarita Done! |
Added end detection described in Eq. (50) of S. Watanabe et al "Hybrid CTC/Attention Architecture for End-to-End Speech Recognition,"
If we set maxlenratio 0.0, this automatic detection is triggered.
With WSJ and Voxforge experiments, it is confirmed that the detection works well.