Sinovoice-2014-03-04

DNN training

Environment setting

• Schedule cluster poweroff on 3/06, construct RAID-0 on train212
• The new RAID-0 used three new disks on train212
• Change nfs names: disk1 -> /nfs/disk212, the raid disks: /nfs/raid212, disk2->/nfs/raid215

Corpora

• PICC data are under labeling (200h), ready in one week.
• 105h data from BJ mobile
• 127h Hubei telecom
• 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

Training condition                    NO NOISE        NOISE in LM       opt noise   NOISE LM + opt noise        

No noise:                                30.61%           -                    -             -
noise phone added:                       31.88%          30.76%              31.27%         31.07

BJ mobile incremental training

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

MPE2      MPE3         MPE3+iLM       MPE4+iLM
27.01%     26.72%       25.09%         24.53%

6000 hour 16k training

Training progress

• Ran CE DNN to iteration 11 (8400 states, 80000 pdf)
• Testing results go down to 12.49% WER (Sinovoice results: 10.49).

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
large LM, it 12, -5/-9	12.91	1.61
large LM, it 12, -5/-10	12.49	1.28

• xEnt training is done

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%

• Xiaoming is working on HTK training on the same database

Hybrid training

• Receipe

• 100h MPE training
• 1700h MPE alignment/lattice
• 1700h MPE training

• 1 week to complete 3 MPE iterations
• MPE2 result: 1e-9: 10.67% (8.61%), 1e-10: 10.34% (8.27%)

Auto Transcription

PICC

• 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
• Prepare to compare the unsupervised incremental training and supervised training

Hubei telecom

• Hubei telecom data (127 h), retrieve 60k sentence by conf thred=0.9, amounting to 50%

xEnt org:  -             wer_15  29.05
MPE iter1：wer_14 29.23；wer_15 29.38
MPE iter2：wer_14 29.05；wer_15 29.11
MPE iter3：wer_14 29.32；wer_15 29.28
MPE iter4：wer_14 29.29；wer_15 29.28

• retrieve 30k sentences by conf thred=0.95, amounting to 25%, plus the original 770h data

xEnt org:     -             wer_15  29.05
MPE iter1:    -             wer_15: 29.36

DNN Decoder

Online decoder

• Various CMN implementation test

• 200ms/500ms frame block adaptation
• 10ms frame block adaptation: totally wrong

prior weight	-1	1	5	10	20	50	100
200ms	28.29	37.53	35.50	34.08	32.90	32.30	32.77
500ms	28.29	31.28	30.83	30.22	29.50	29.32	29.36

• CMN code delivery
• Online model adaptation