Sinovoice-2016-4-28

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
  -----------------------------------------------------------------------------

• 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
  -----------------------------------------------------------------------------

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

Big-Model Training

• 7*2048-10000h net weight-matrix factoring, to improve the decoding speed --SVD

• SVD looks OK, but fine-tuning still didn't work.

 Base WER:
 relu_2000_mpe_1000H: 17.72
 relu_1200_mpe_1000H: 18.60

 |layer / nodes retaind|  200  |  400  |  600  |  800  | 1000  | 1200  | 1400  |  1600  |
 |      hidden 2       |       |       | 22.53 | 20.30 | 19.01 |       |       |        |
 |      hidden 7       |       | 18.92 | 18.30 | 17.92 |       |       |       |        |     
 |       final         |       |       | 18.32 | 18.00 | 17.83 |       |       |        |     

• 7*1024 cross-entropy total train, then mpe, 0.2 improvment
• 7*1024 svd factoring, speed the decoding

• 8k

Embedding

• 10000h-chain 5*400+800 DONE.

• Beam affect the performance of chain model significantly, need more investigation.

• 5*576-2400 TDNN model

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.

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

• Same Channel test EER: 100%

• Speaker confirm
• phone channel

• Cross Channel

• Mic-wav PLDA adaptation EER from 9% to 7% (20-30 persons)