Sinovoice-2014-01-20

DNN training

Environment setting

• Cluster accounts rearrangement
• Withdraw root/sudo previelege
• Changed NFS server to 40 processes, hope to increase the disk reading speed
• Create a RAID-0 with 3 or 4 3T disks

Corpora

• Change the data labeling strategy: do not label gender and the length of noise in the rest of the corpora.
• Automatic labeling

• Xiaoming will work with Zhiyong to discover how to generate transcriptions with confidence score embedded.
• The first step is to investigate the raw accuracy on the domain-dependent test, and then decide the quality of automatic labeling

• Xiao Na prepare 300h telephone data (Sinovoice recording) to improve the 8k model.

470 hour 8k training

• MPE training done

Model	CE	MPE1	MPE2	MPE3	MPE4
4k states	23.27/22.85	21.35/18.87	21.18/18.76	21.07/18.54	20.93/18.32
8k states	22.16/22.22	20.55/18.03	20.36/17.94	20.32/17.78	20.29/17.80

6000 hour 16k training

• Feature extraction done: solved three problems in the data: (1) short wave (2) unmatched file length (3) unmatched sample rate
• Training goes to tri4b, quick increase of states/pdfs
• DNN training could be started from Tuesday

DNN Decoder

• Sinovoice decoder: some errors in FST building. Many triphones are lost after graph building. Problems in cdgen?
• Kaldi decoder:

• A minor difference between CLG/HCLG results was find. Debugging into the problem.
• CLG RT is comparable to the HCLG RT, 0.3-0.4 in CSLT grid-2.
• Additional optimization on pdf-pre-computing will be investigated.
• Code deliver today.