Sinovoice-2016-3-17

来自cslt Wiki
2016年3月17日 (四) 07:32Zhangzy讨论 | 贡献的版本

(差异) ←上一版本 | 最后版本 (差异) | 下一版本→ (差异)
跳转至: 导航搜索

Data

  • 16K LingYun
  • 2000h data ready
  • 4300h real-env data to label
  • YueYu
  • Total 250h(190h-YueYu + 60h-English)
  • CER: 75%
  • WeiYu
  • 50h for training
  • 120h labeled ready

Big-Model Training

  • 16k-10000h DNN training 7*1024-10000
  • MPE training to be done
  • Performance about 1 point worse than 7*2048-1000
  • The decoding speed is similar with old-nnet1 version
  • 7*2048-1000 net weight-matrix factoring, to improve the decoding speed --SVD
  • TDNN+LSTM chain training
  • Beam-13 performance, chain is better, but beam-9 worse.
  • 8K
  • HuaWei
  • 5000h 7*1024-XENT be training, performance is similar to 7*2048
  • To train MPE base on current XENT-model

Embedding

  • 10000h-chain 5*400+800 DONE.
  • To confirm the beam affection on the performance

SinSong Robot

  • Test based on 10000h(7*2048-xent) model
 ------------------------------------------------
   condition | clean  | replay(0.5m) | real-env
 ------------------------------------------------
     wer     |   3    |  18(mpe-14)  | too-bad
 ------------------------------------------------
  • Recording

Character LM

  • Except Sogou-2T, 9-gram has been done.
  • Worse than word-lm(9%->6%)
  • Merge Character-LM & word-LM
  • Union
  • Compose

Path Emphasize

  • Script Done.

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)