Sinovoice-2014-02-17

DNN training

Environment setting

• 2nd GPU machine is ready. 3T * 4 RAID-0 is fast enough.
• The new machine has been added into the SGE env.

Corpora

• Beijing mobile 120h speech data are ready.
• PICC data are under labeling (200h), ready in two weeks.
• Now totally 1100h telephone speech will be ready soon.

470 hour 8k training

• 470 + 300h + Beijing mobile 120h training

• Re-train the whole models including gmm+dnn, with noise model involved.
• Train noise model by treating noise as a special phone
• The noise should be treated specifically in L construction
• 7.2h per iteration, the xEnt training might be finished in 1 week
• Run incremental DT training on the CSLT cluster, by mapping noise to the silence phone.

6000 hour 16k training

• Ran CE DNN to iteration 8 (8400 states, 80000 pdf)
• Testing results go down to 12.69% WER (Sinovoice results: 10.70).

Model	WER	RT
small LM, it 4, -5/-9	15.80	1.18
large LM, it 4, -5/-9	15.30	1.50
large LM, it 4, -6/-9	15.36	1.30
large LM, it 4, -7/-9	15.25	1.30
large LM, it 5, -5/-9	14.17	1.10
large LM, it 5, -5/-10	13.77	1.29
large LM, it 6, -5/-9	13.64	-
large LM, it 6, -5/-10	13.25	-
large LM, it 7, -5/-9	13.29	-
large LM, it 7, -5/-10	12.87	-
large LM, it 8, -5/-9	13.09	-
large LM, it 8, -5/-10	12.69	-

• A new round of training with shared trees for tone variations has been kicked off and run into dnn training again.
• Need to test the new gmm model, need to compare to Xiaoming's original settings

Adaptation

• Adaptation with 10, 20, 30 sentences are conducted
• 30 sentences can reach reasonable performance (from 14.6 to 11.2).
• Hidden layer adaptation is better than input and output adaptation
• Cross entropy regularization with P=0.3 is reasonable
• Results are here

DNN Decoder

• Faster decoder

• The new RT is reported here
• The RT of the latest decoder on train203 is 0.144 (HCLG) 0.148 (CLG).

• Online decoder

• Interface design completed
• CMN strategy is clear: (1) global CMN model be trained first (2) Apply the model in decoding directly (3) may need to adapt the DNN model slightly with the feature.