Sinovoice-2014-04-08

Environment setting

• Sinovoice internal server deployment. The current configuration involves Gitlab + Trac.
• Some access permission problems remain, need to investigate other tools.

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.
• 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 5 iteration: 23.26
• Xent 16 iteration: 22.90

• 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

• lattice done, MPE run 4 days

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: Wang Dong provide info.
• Prepare shared DNN structure for multilingual training

Noise robust feature

• GFbank can be propagated to Sinovoice
• Liuchao will prepare fast computing code

Language modeling

• Training process was delivered.
• Some problems in encoding.
• Training text retrieval based on topic models (initial key words)

DNN Decoder

decoder optimization

• Test computation cost of each step

• beam 9/5000: netforward 65%
• beam 13/7000: netforward 28%

• Liuchao will verify the proportion number with CSLT engine.
• Zhiyong & Liuchao will deliver the frame-skipping approach.
• Investigate BigLM retrieval optimization.