Hulan-2013-12-13

ASR

ASR Kernel development

[ASR group weekly report]

TTS

• This week

• 2000 utterances recording finished, on labeling silence
• Design/implement the online stream service
• Solve the problem of single English letter pronunciation.
• CGI server finished

• Next week

• Deliver new male/female voice

Dialog system

Statistical approach

• Word2Vector seems use the same training approach of NN LM, however we do not know the training object. Will clarify this next Monday.
• Word2Vector can be used to expand queries based on semantic distance.
• Tested edit distance to solve unmatched similar words. A little performance gain was obtained, but not quite significant.

n-best	original	edit-distance
1	0.604		0.605
2	0.740		0.744
3	0.806		0.810
4	0.842		0.845
5	0.858		0.862
6	0.870		0.873
7	0.879		0.883
8	0.887		0.899

Template matching

• Grammar design completed

System design

• Interface design, web design completed
• Coding (with automatic web code generator)