Sinovoice-2014-04-01

Environment setting

• Create cluster accounts for Wufei and Xiaoxi.

Corpora

• New Beijing Mobile labeling done (109h).
• Next will label the corrupted speech from Huawei(97h).
• 300h text transcription will be ready in 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.
• LM corpus preparation done.

Acoustic modeling

Telephone model training

1000h Training

• Baseline: 8k states, 470+300 MPE4, 20.29
• Jietong phone, 200 hour seed, 10k states training:

• Error in training found.
• Xent run into 5 iterations

• CSLT phone, 8k states training

• MPE1: 20.60
• MPE2: 20.37
• MPE3: 20.37

6000 hour 16k training

Training progress

• 6000h/CSLT phone set alignment/denlattice completed

• Xent: 12.83
• MPE1: 9.21

• 6000h/jt phone set alignment/denlattice completed

• denlattice re-run

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
• Prepare shared DNN structure for multilingual training

Language modeling

• Training data ready
• Xiaoxi and Wufei will collaborate to make familiar the training & testing process
• Initial optimization from telecom.

DNN Decoder

Online decoder adaptation

• Incremental training finished (stream mode)
• Online decoder completed

• Test the proportion of DNN forward and Graph search in decoding