Zhiyong Zhang 2016-04-18

• Weekly report

1. Character-LM and Word-LM merge

:# Union the two can achieve 9.53-wer compared with only word-LM 9.67-wer, but the HCLG is twice bigger than only word-lm;
:# Compose get an wer of 64-wer, need deep investigation

1. Re-compile nnet3-mpe under the 20160410 kaldi version
2. Assist the restaurant recording of Siason
3. Assist the implementation of NG-SGD in nnet1 version with Xiangyu

• Next week

1. To solve the deletion error in discriminative training
2. To check the SVD re-train failure problem