“2013-06-28”版本间的差异

2013年6月28日 (五) 09:56的最后版本

Data sharing

• LM count files still undelivered!

DNN progress

Experiments

• Sparse DNN.

1. With Atlas without any change, on the ARM platform, obtained RT 2.0. Will change the Atlas code to support sparse matrices.

Tencent exps

1：本周大规模数据DNN实验结束，基本在第8次迭代收敛，

Old Baseline New Baeline DNN-1000小时 DNN-6000小时（7次迭代）DNN-6000小时（8次迭代）

1900		8.4		6.8		4.3		3.9			3.7
2044		22.4		15.7		12.7		10.7			10.6
online1	35.6		32.7		25.8		24.6			24.6
online2	29.6		27.3		22.1		21.1			20.8
map		24.5		15.8		13.4		8.7			8.4
general	36		25.1		19.3		15.9			16

2：dnn+dt代码完成，下周可以开始实验。首先将在小数据上进行。

GPU & CPU merge

1. Hold

RNN LM progress

• Use 100M text, 10k lexicon in training. Validation test set is obtained from the transcription of the Tencent online1 speech data.

100 hidden layer, 1 hidden layer, 3-gram

training time: 7 hour, 8GB

prediction time: quick, 8GB

3-gram PPL: 227.323 WER: 36%

RNN PPL: 170.056129 WER: 41%

3-gram+RNN: PPL: 0.25RNN+0.75 3-gram: 180.0 WER 35%

• possibly a bug when computing PPL with the RNN toolkit.

Embedded progress

• Status:

1000 test words + 2000 noise words

        before  |  after 
 utt    952        3317
%wer    6.26%     11.04%
 RT     0.07      0.20

This means the GMM-based system highly relies on the vocabulary. It may work well with small lexica, but difficult with large ones.

Run other optimization parameters:

 option   |   RT    |   %wer

 original    0.07       6.28

  -ds        0.06       6.33%
  -topn      0.06       6.80%  
  -maxwpf     -          -
  -maxhmmpf   -          - 
  -kdmaxdepth -          -
  -kdmaxbbi   -          -
  -pl_window  -          -

• To be done

1. sparse DNN based Kaldi engine
2. sparse DNN based PS engine