Sinovoice-2014-04-15

来自cslt Wiki
2014年4月15日 (二) 08:25Cslt讨论 | 贡献的版本

跳转至: 导航搜索

Environment setting

  • Sinovoice internal server deployment. Now a better approach by using Gitlab + Redmine.
  • Delivered part of the Kaldi code. Some fix yet waiting for check in.
  • Email notification is problematic. Need obtain a smtp server

Corpora

  • 300h Guangxi telecom text transcription prepared. 150h before 18th, April.
  • Now totally 1338h (470 + 346 + 105BJ mobile + 200 PICC + 108h HBTc + 109h New BJ mobile) telephone speech is ready.
  • 16k 6000h data: 978h online data from DataTang + 656h online mobile data + 4300h recording data.
  • Standard established for LM-speech-text labeling (speech data transcription for LM enhancement)
  • Xiaona will prepare noise database. Start from telephone speech.

Acoustic modeling

Telephone model training

1000h Training

  • Baseline: 8k states, 470+300 MPE4, 20.29
  • Jietong phone, 200 hour seed, 10k states training:
  • Xent 16 iteration: 22.90
  • MPE1 : 20.89
  • CSLT phone, 8k states training
  • MPE1: 20.60
  • MPE2: 20.37
  • MPE3: 20.37
  • MPE4: 20.37

6000 hour 16k training

Training progress

  • 6000h/CSLT phone set training
  • Xent: 12.83
  • MPE1: 9.21
  • MPE2: 9.13


  • 6000h/jt phone set phone set training
  • ran into MPE1.


Train Analysis

  • The Qihang model used a subset of the 6k data
  • 2500+950H+tang500h*+20131220, approximately 1700+2400 hours
  • GMM training using this subset achieved 22.47%. Xiaoming's result is 16.1%.
  • Seems the database is still not very consistent
  • Xiaoming kicked off the job to reproduce the Qihang training using this subset

Multilanguage Training

  • Prepare Chinglish data: contacted with a vendor for 1000 hour mobile recording. Will check how much we need
  • AMIDA database downloading
  • Build a baseline system
  • Prepare shared DNN structure for multilingual training

Noise robust feature

  • GFbank can be propagated to Sinovoice
  • Let Mengyuan prepare the experiments
  • Liuchao will prepare fast computing code

Language modeling

Training recipe transfer

  • Training process was delivered.
  • Problems in encoding were solved.
  • Initial CSLT LM buildup completed.

Domain specific atom-LM construction

Some potential problems

  • Unclear domain definition
  • Using the same development set (8k transcription) is not very appropriate

Text data filtering

  • Prepare word list
  • VSM-based topic segmentation was delivered to Sinovoice, but the tool is highly inefficient.
  • An enhanced toolkit was delivered.
  • A telecom specific word list is ready, several stop words are ready

DNN Decoder

decoder optimization

  • Test computation cost of each step
  • beam 9/5000: netforward 65%
  • beam 13/7000: netforward 28%
  • Sinovoice change in Kaldi delivered and ready to check-in
  • Need to verify the speed of the CSLT engine

Frame-skipping

  • Zhiyong & Liuchao will deliver the frame-skipping approach.

BigLM optimization

  • Investigate BigLM retrieval optimization.