“Sinovoice-2016-5-26”版本间的差异

2016年5月26日 (四) 03:53的版本

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

• 8k more data
• 50h for training
• 120h labeled ready

• PingAn

• 100h User data done

Model training

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

Big-Model Training

16k

8k=

Model

• Add noise phone

• 1300.mdl done

• CNN + TDNN

• 280/900 mdl
• Need about 12 days Xent

Project

• PingAn

• Add noise phone to phone-list

 =========================================================================================
 |     AM / config                     |      all      |    KeHu wer   ||  KeHu no-ins  |
 -----------------------------------------------------------------------------------------
 | tdnn 7-2048 xEnt                    |     16.45     |     36.49     ||     25.18     |
 | tdnn 7-2048 MPE                     |     15.22     |     32.77     ||     23.48     |
 | tdnn 7-2048 MPE adapt-PABX          |     14.67     |     31.33     ||     22.76     |
 -----------------------------------------------------------------------------------------
 | tdnn 7-1024 xEnt                    |     16.60     |     35.91     ||     25.58     |
 | tdnn 7-1024 MPE 2e-6                |     15.67     |     32.77     ||     26.09     |
 | tdnn 7-1024 MPE 2e-5 1.mdl          |     15.54     |     32.77     ||     26.29     |
 | tdnn 7-1024 MPE 1e-5 4.mdl          |     15.76     |     33.55     ||     27.20     |
 | tdnn 7-1024 MPE adapt-PABX          |     14.80     |     30.48     ||     22.56     |
 -----------------------------------------------------------------------------------------
 | spn 7-1024 xEnt                     |     16.49     |     36.23     ||     24.59     |
 | spn 7-1024 xEnt xEnt-PA_user 101.mdl|     16.19     |     33.22     ||     22.69     |
 | spn 7-1024 xEnt xEnt-PA_user mpe    |     15.24     |     32.77     ||     21.65     |
 | spn 7-1024 MPE-1000H 23.mdl         |     15.29     |     33.09     ||     21.65     |
 | spn 7-1024 MPE adapt-PA_all 29.mdl  |     15.11     |     33.42     ||     21.84     |
 | spn 7-1024 MPE adapt-PA_user 2e-5   |     15.31     |     31.79     ||     20.14     |
 | spn 7-1024 MPE adapt-PA_user Hs 2e-5|     15.32     |     32.24     ||     20.93     |
 =========================================================================================

• LiaoNingYiDong:

 =========================================================================
 |     AM / config                     |     LNYD      |  LNYD re-tag  |
 -------------------------------------------------------------------------
 | tdnn 7-2048 xEnt                    |     21.51     |               |
 | tdnn 7-2048 MPE                     |     20.09     |               |
 | tdnn 7-2048 MPE adapt-LNYD          |     17.92     |     16.29     |
 -------------------------------------------------------------------------
 | tdnn 7-1024 xEnt                    |     21.72     |               |
 | tdnn 7-1024 MPE                     |     20.99     |               |
 | cnn 7-1024 xEnt 600.mdl             |     21.03     |               |
 | cnn 7-1024 MPE 12.mdl               |     19.80     |               |
 | cnn 7-1024 MPE adapt-LNYD 41.mdl    |     17.96     |     15.93     |
 -------------------------------------------------------------------------
 | spn 7-1024 xEnt                     |     21.70     |               |
 | spn 7-1024 MPE-1000H 23.mdl         |     19.97     |               |
 | spn 7-1024 MPE adapt-LNYD           |     18.67     |               |
 | spn cnn 7-1024 xEnt 300.mdl         |     22.26     |               |
 ========================================================================

Embedding

• The size of nnet1 AM is 6.4M (3M after decomposition). So we need to control AM size within 10M.
• 5*500-2400 TDNN no-svd/svd100 model, MPE training done

LM=1e-5, beam=9, max-active=5000
 =============================================================================================================
 |         AM / testset              |  test_1000ju  |  test_2000ju  |  test_8000ju  |  test_10000ju  |
 -------------------------------------------------------------------------------------------------------------
 | nnet1 4*600+800 xEnt (6.4M)       |     25.30     |     40.48     |               |                |
 | nnet1 4*600+800 mpe  (6.4M)       |     20.75     |     35.33     |               |                |
 -------------------------------------------------------------------------------------------------------------
 | nnet3 5*500 mpe (13M)             |     16.18     |     29.53     |               |                |
 | nnet3 5*500 svd-100 mpe (9.5M)    |     17.69     |     30.11     |               |                |
 =============================================================================================================

Character LM

• Except Sogou-2T, 9-gram has been done.
• Add word boundary tag to Character-LM trainig done

• 9-gram
• Except Weibo & Sogou-2T
• 1e-7(13M) wer17.91 compared with 1e-7(no-boundary,71M) 13.4
• 1e-8(54M) wer17.54

• Prepare specific domain vocabulary

• Dianxin/Baoxian/Dianli

• DT lm training

• ReFr

• Merge Character-LM & word-LM

• Union
• Compose, success.

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

SiaSong Robot

• Beam-forming algorithm test
• NN-model based beam-forming

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

• Engine Package