hirofumi0810 / neural_sp

End-to-end ASR/LM implementation with PyTorch
Apache License 2.0
594 stars 140 forks source link

How can I reproduce WER in Librispeech? #177

Closed FzuGsr closed 3 years ago

FzuGsr commented 4 years ago

Hi,I want to reproduce Librispeech WER in this project. I tried to run the script blstm_las with 3 gpus and batch_size 128, but I couldn't get the result in readme. My training results are as follows

2020-10-25 17:06:31,478 __main__ line:467 INFO: WER (dev_other_960_wpbpe10000, ep:1): 33.09 %
2020-10-25 18:45:39,793 __main__ line:467 INFO: WER (dev_other_960_wpbpe10000, ep:2): 24.93 %
2020-10-25 20:24:06,648 __main__ line:467 INFO: WER (dev_other_960_wpbpe10000, ep:3): 22.18 %
...
2020-10-27 12:21:00,003 __main__ line:467 INFO: WER (dev_other_960_wpbpe10000, ep:28): 13.30 %
2020-10-27 13:56:46,384 __main__ line:467 INFO: WER (dev_other_960_wpbpe10000, ep:29): 13.51 %
2020-10-27 15:30:51,522 __main__ line:467 INFO: WER (dev_other_960_wpbpe10000, ep:30): 13.34 %

Is there a problem? How to achieve the accuracy of readme (dev-other: 91.7).

image image Look forward to your reply~ Thank you

hirofumi0810 commented 4 years ago

@FzuGsr Did you use SpecAugment?

FzuGsr commented 4 years ago

SpecAugment

no,I run run.sh directly

hirofumi0810 commented 4 years ago

That is the reason. Please use --conf2 data/spec_augment.yaml and initialize parameters with the pre-trained model by setting --asr_init path_to_your_model.

FzuGsr commented 4 years ago

--asr_init path_to_your_model

Thank you . I will try it later.