Sinovoice-2014-02-25

DNN training

Environment setting

• Two queues: 100.q dedicated to decoding, all.q dedicated to GMM training/MPE lattice generation
• disk203-disk210: distributed disks, for parallel jobs
• /nfs/disk1: train212 GPU task; /nfs/disk2: train215 GPU task

Corpora

• PICC data are under labeling (200h), ready in one week.
• 105h data from BJ mobile
• Now totally 1121h (470 + 346 + 105 + 200) telephone speech will be ready soon.
• 16k 6000h data: 978h online data from DataTang + 656h online mobile data + 4300h recording data

Telephone model training

470 + 300h + BJ mobile 105h training

(1) 105 BJ mobile re-training without NOISE: 33.97% WER

(2) 105 BJ mobile re-training with NOISE phone in training, but decoding without NOISE: 34.27%

(3) (2) + noise-decoding (with noise phone in lexicon/LM), still under investigation

BJ mobile incremental training

(1) Original 470 + 300 model: 30.24% WER

(2) Incremental DT training with 105h BJ data, 27.01% WER

6000 hour 16k training

Training progress

• Ran CE DNN to iteration 11 (8400 states, 80000 pdf)
• Testing results go down to 12.46% WER (Sinovoice results: 10.46).

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	1.12
large LM, it 6, -5/-10	13.25	1.33
large LM, it 7, -5/-9	13.29	1.12
large LM, it 7, -5/-10	12.87	1.17
large LM, it 8, -5/-9	13.09	-
large LM, it 8, -5/-10	12.69	-
large LM, it 9, -5/-9	12.87	-
large LM, it 9, -5/-10	12.55	-
large LM, it 10, -5/-9	12.83	1.51
large LM, it 10, -5/-10	12.48	1.65
large LM, it 11, -5/-9	12.87	1.61
large LM, it 11, -5/-10	12.46	1.28

• Additional xEnt training with DNN alignment, should be completed in 2 days
• DT training is still on the queue, waiting for lattice generation
• First version of DT model would be trained with online data (1700h)

Training Analysis

• Shared tree GMM model training completed, WER% is similar to non-shared model .
• Selected 100h online data, trained two systems: (1) di-syllable system (2) jt-phone system

        di-syl      jt-ph
Xent    15.42%      14.78%       
MPE1    14.46%      14.23%
MPE2    14.22%      14.09%
MPE3    14.26%      13.80%
MPE4    14.24%      13.68%

Auto Transcription

• PICC development set decoding obtained 45% WER.
• PICC auto-trans incremental DT training completed

Threshold  WER
org:     45.03%
0.9:     41.89%
0.8:     41.64%

• Current training data with 0.8 involve 80k sentences, amounting to about 60h data.
• Sampling 60h labelled data to enrich the training

DNN Decoder

• Online decoder

• Integration almost completed
• Initial CMN implementation finished
• The first step is to tune the prior prob of the global CMN, and then consider re-training with DT.