Sinovoice-2016-5-5

Data

• 16K LingYun

• 2000h data ready
• 4300h real-env data to label

• YueYu

• Total 250h(190h-YueYu + 60h-English)
• Add 60h YueYu
• CER: 75%->76%

• WeiYu

• 50h for training
• 120h labeled ready

Deletion Error Promblem

• Add one noise phone to alleviate the silence over-training
• Omit sil accuracy in discriminative training
• H smoothing of XEnt and MPE

• Testdata: test_1000ju from 8000ju

  -----------------------------------------------------------------------------
                 model                    | ins  |  del  | sub | wer/tot-err  
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix         |  24  |  56   | 408 | 8.26/488
  -----------------------------------------------------------------------------
svd600_lr2e-5_1000H_mpe_uv-fix_omitsilacc |  32  |  48   | 409 | 8.28/489
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix_xent0.1 |  24  |  57   | 406 | 8.24/487
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix_xent0.2 |  25  |  60   | 409 | 8.36/494
  -----------------------------------------------------------------------------

• Testdata: test_8000ju

  -----------------------------------------------------------------------------
                 model                    | ins  |  del  | sub  | wer/tot-err  
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix         |  140 |  562  | 3686 | 9.19/4388     | 47753-total-word
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix_xent0.1 |  146 |  510  | 3705 | 9.13/4361
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix_xent0.2 |  139 |  492  | 3739 | 9.15/4370
  -----------------------------------------------------------------------------

• Testdata: test_2000ju from 10000ju

  -----------------------------------------------------------------------------
                 model                    | ins  |  del  |  sub | wer/tot-err  
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix         |  86  |  790  | 1471 | 18.55/2347
  -----------------------------------------------------------------------------
svd600_lr2e-5_1000H_mpe_uv-fix_omitsilacc |  256 |  473  | 1669 | 18.95/2398
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix_xent0.1 |  95  |  704  | 1548 | 18.55/2347
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix_xent0.2 |  100 |  697  | 1557 | 18.60/2354
  -----------------------------------------------------------------------------

• Testdata: test_10000ju

  -----------------------------------------------------------------------------
                 model                    | ins  |  del  | sub  | wer/tot-err  
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix         |  478 | 3905  | 7698 | 18.31/12081  | 65989-total-word
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix_xent0.1 |  481 | 3741  | 7773 | 18.18/11995
  -----------------------------------------------------------------------------
   svd600_lr2e-5_1000H_mpe_uv-fix_xent0.2 |  502 | 3657  | 7826 | 18.16/11985
  -----------------------------------------------------------------------------

• Add one silence arc from start-state to end-state

Big-Model Training

• 16k

• 8k

PingAnAll:
 ==================================================================================
 |         AM / error         | tot_err |   ins   |   del   |   sub   |   wer   |
 ----------------------------------------------------------------------------------
 | tdnn 7-1024 xEnt 2500.mdl  |  3626   |   619   |   773   |   2234  |  16.60  |
 ----------------------------------------------------------------------------------
 | spn 7-1024 xEnt 300.mdl    |  3746   |   702   |   763   |   2281  |  17.15  |
 ==================================================================================

PingAnUser:
 ==================================================================================
 |         AM / error         | tot_err |   ins   |   del   |   sub   |   wer   |
 ----------------------------------------------------------------------------------
 | tdnn 7-1024 xEnt 2500.mdl  |   549   |   158   |    75   |   316   |  35.91  |
 ----------------------------------------------------------------------------------
 | spn 7-1024 xEnt 300.mdl    |   571   |   151   |    97   |   323   |  37.34  |
 ==================================================================================

LiaoNingYiDong:
 ==================================================================================
 |         AM / error         | tot_err |   ins   |   del   |   sub   |   wer   |
 ----------------------------------------------------------------------------------
 | tdnn 7-1024 xEnt 2500.mdl  |   5873  |   879   |   1364  |   3630  |  21.72  |
 ----------------------------------------------------------------------------------
 | spn 7-1024 xEnt 300.mdl    |   6257  |   977   |   1348  |   3923  |  23.14  |
 ==================================================================================

Embedding

• The size of nnet1 AM is 6.4M (3M after decomposition). So we need to control AM size within 10M.
• 5*576-2400 TDNN model training done. AM size is about 17M
• 5*500-2400 TDNN model on training.
• making lattice for MPE training.

SinSong Robot

• Test based on 10000h(7*2048-xent) model

 ------------------------------------------------
   condition | clean  | replay(0.5m) | real-env
 ------------------------------------------------
     wer     |   3    |  18(mpe-14)  | too-bad
 ------------------------------------------------

• Plan to record in restaurant on April 10.

Character LM

• Except Sogou-2T, 9-gram has been done.
• Worse than word-lm(9%->6%)
• Add word boundary tag to Character-LM trainig
• Merge Character-LM & word-LM

• Union
• Compose, success.

• 2-step decoding: first, character-based LM. Then, word-based LM.
• Word boundary character training

Project

• Pingan & Yueyu Deletion error too more

• TDNN deletion error rate > DNN deletion error rate
• TDNN Silence scale is too sensitive for different test cases.

SID

Digit

• DNN-PLDA gets better performance than i-Vector;

DNN cosine 10.4167%, at threshold 89.3973 9.72222%, at threshold 87.8146 8.68056%, at threshold 84.2021 3.47222%, at threshold 11.5852 lda 3.125%, at threshold 54.1172 2.77778%, at threshold 50.1447 2.43056%, at threshold 48.6887 1.73611%, at threshold 14.5075 plda 2.43056%, at threshold -23.954 2.08333%, at threshold -24.6051 2.08333%, at threshold -21.0524 1.73611%, at threshold 4.83949

ivector plda 3.15789%, at threshold 0.563044 3.85965%, at threshold 0.525273 3.85965%, at threshold 0.502531 2.80702%, at threshold 0.429186