2013年9月27日 (五) 01:36的最后版本

ASR

ASR Kernel development

[ASR group weekly report]

TTS

• Lab format learned
• All the details of the label format are clear
• Construct label files from word/pingyin/phone transcription. Use csep word-segmentation tool to obtain these transcriptions from the original text.
• Monophone, Triphone Chinese prototype system is ready. 500 sentences from 863 data are used for training. The trivial questions were used for clustering. 16k signals with 256 FFT transform. GV model used.
• The voice is funny.

Next week:

• Keep on collecting context-dependent labels

Dialog system

• MH, RY help compose 60 questions, which are being used for testing.
• Conducting the initial experiment:

1. Using 9k dim TF/IDF, compose feature vectors for each query, each answer. Mach the TF/IDF of query+answer to match the TF/IDF of new queries. Add the scores of the Cosine score of the match with queries and answers directly.

• CER: 8.3%
• Query time: very slow

1. Remove 506 stop words. No significant change on CER & Query time.

1. Fast match by listing the words only in the query & answer as the feature, and matching by order to speed up the score calculation.

1. Hierarchical matching. First split all the answers to 11 top-level categories, and then split them into 1030 second-level categories. query+answer TF/IDF score.

• CER: top-category: 6.7%, second-level category: 18.3%
• Query time: 6/sec

1. Keep two top-level categories, try to reduce top-level errors:

• CER: two top-category, no errors. second-level category: still on going

Next week:

• Reverse index-based fast match, on going.