Sinovoice-2016-6-16

Data

• WeiYu

• 188h 8k new data for labeling

Big-Model Training

16k

• Done!

8k Project

• PingAn

• CNN+TDNN

• LiaoNingYiDong

• CNN+TDNN

Embedding

• Done!
• Fix-point computation is slow than floting-point on Android device
• Optimize the NN structure

• Normalization component optimization
• Splice-context
• Narrow and deep structure

LM

• Prepare specific domain vocabulary

• Dianxin/Baoxian/Dianli

• DT lm training

• ReFr

Problem

• Stream-mode decoding(cmvn) causes performance reduction.

SiaSun Robot

• Beam-forming + adaptation
• NN-model based beam-forming

SID

Digit

• Engine Package