/Chinese-ASR

Chinese-ASR built on kaldi

Primary LanguageShell

Chinese-ASR

Chinese-ASR built on kaldi

Dependencies:

Opencc: convert simplified Chinese to traditional Chinese

https://github.com/yichen0831/opencc-python

jeiba zh version : Traditional Chinese word segmentation tool:

https://l.facebook.com/l.php?u=https%3A%2F%2Fgithub.com%2Fldkrsi%2Fjieba-zh_TW&h=ATPhhi1b7UYw84pPzgAz4MDbn3MRo7oFLAuhBLW8geUqHF0O1YZDnXsNh5qe7tQVWGQ5uaocYvuV-UsuvALNeN3LRaq68ACLMfbWE2RivhiCHoyjcFtNTVy6XG0sh5MJTp5tYEZm0xA

Usage

1.modify kaldi path in path.sh

2.modify corpus path in local/data/corpus_path.sh

3.Install sequitar(G2P), sox, kaldi_lm in kaldi/tools/

4.bash run.sh

some useful scripts

  1. LM training and interpolation : local/lm

  2. Customed WFST for multiple choice problem : local/lm/wfst

-> Force the outputs to the format of " XXX XXX XXX XXX"

  1. scripts of training DFSMN : local/nnet

Experiment

Ｍodel TOCFL(CER%) Cyberon_Chinese_test(CER%)
mono0a 97.76 100.71
tri1 50.55 63.64
tri2 56.62 46.65
tri3 34.78 46.78
tri4 37.02 34.02
tri5 65.60 49.96
tdnn_lstm1 18.30 24.82
tdnn_lstm(realign) 15.88 22.24
DFSMN(Alibaba) 11.22 12.14