k2-fsa / snowfall

Moved to https://github.com/k2-fsa/icefall
Apache License 2.0
143 stars 42 forks source link

Update the aishell recipe to adapte the new version snowfall #246

Closed luomingshuang closed 2 years ago

luomingshuang commented 2 years ago

We have a new version snowfall. But the aishell recipe codes don't update. The aishell recipe codes can't run based on the current version snowfall. I notice some people still use this repository for Chinese ASR. So I take some time to update the aishell recipe codes to adapt the current version snowfall. Also, I have run experiments based on the new aishell recipe and got some good results. The results are as follows:

AiShell New Training Results

2021-08-25

(Mingshuang Luo): All the following results are based on a 10-epoch training process.

Aishell CTC_Train

2021-08-25 00:46:57,911 INFO [ctc_decode.py:182] %WER 24.24% [15616 / 64428, 1289 ins, 2041 del, 12286 sub ]
2021-08-25 00:46:57,911 INFO [ctc_decode.py:186] %CER 15.02% [15737 / 104765, 415 ins, 484 del, 14838 sub ]

Aishell MMI_Bigram_Train

2021-08-25 00:48:33,789 INFO [mmi_bigram_decode.py:198] %WER 17.11% [11026 / 64428, 1014 ins, 1536 del, 8476 sub ]
2021-08-25 00:48:33,789 INFO [mmi_bigram_decode.py:202] %CER 8.79% [9206 / 104765, 354 ins, 346 del, 8506 sub ]

AiShell MMI_Att_Transformer_Train

2021-08-25 00:51:22,460 INFO [mmi_att_transformer_decode.py:551] %WER 14.00% [9019 / 64428, 797 ins, 1228 del, 6994 sub ]
2021-08-25 00:51:22,460 INFO [mmi_att_transformer_decode.py:555] %CER 6.28% [6580 / 104765, 237 ins, 156 del, 6187 sub ]
luomingshuang commented 2 years ago

About the directory "local2", I copy the files from librispeech/asr/simple_v1/local2 and just adjust the add_silence_to_transcript.py.

danpovey commented 2 years ago

Thanks! Merging.