Hulan-2013-11-30
来自cslt Wiki
ASR
ASR Kernel development
TTS
- 5000 utterance done.
- 500 utterance TN recording. Quality controlled failed. Resulting synthesis is unsatisfied.
- 41 WD utterance recording. Quality control fine. Adaptation done. Sound OK.
- Buzzy sound was investigated.The main source is the source model (excitation). STRAIGHT sounds better.
Dialog system
- HowNET information extracted. Coverage is limited. 50k words.
- Custom task error analysis
There are totally 2000 errors. Investigated into 600 errors.
- NULL query, 1.4%
- English upper/lower mismatch. 1.6%
- Traditional/Simple Chinese mismatch. 2.2%
- High frequency of sub-important words, like taxing. 1.3%
- Database labeling error (matched query is better than the labeled correct query). 21.8%
- Stand query or query involve many unimportant words, leading to less TF/IDF. STOP words still impact. 10.7%
- TF/IDF incorrectly weighted the matched terms. 3.9%
- Synonym can not match. 36.5%
- Category words can not match. 13.5%
- Answer label incorrect. Semantic relationship missing. 6.8%
- Word segmentation hide keywords. 4%
- Vague query. None discriminative words after stop words purging. 1.6%
Template matching
- Work on disambiguousness
- Completed the prototype
- Development source code generation
