1989 Fiscal Year Final Research Report Summary
Continuous Speech Recognition System for Noisy Environment Using Artificial Intelligence
| Project/Area Number |
62550269
|
|---|
| Research Category |
Grant-in-Aid for General Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Research Field |
計算機工学
|
|---|
| Research Institution | Keio University |
|---|
Principal Investigator |
NAKAGAWA Masao Faculty of Science and Technology, Keio University Professor, 理工学部・電気工学科, 教授 (30051882)
|
|---|
| Project Period (FY) |
1987 – 1989
|
| Keywords | HMM / Noise / Continuous Speech Recognition / Trigram |
|---|
| Research Abstract |
We have obtained following three results.
(1)We have proposed a frame-synchronous HMM connected word recognition algorithm which can obtain multiple candidates for every sentence length. The algorithm can also obtain approximate solution with maximum probability of word sequence using the trigram model. We also introduce beam search into this algorithm to reduce the amount of computation. In experiments with 100 words vocabulary, the sentence recognition rate is 71.3%, and the sentence recognition rate within the 10th best candidates is 88.8 % . The result shows the increases of 2.5% and 12.5% respectively from the results with an algorithm obtaining a single candidate using the trigram model. The increase in amount of computation is cut down by using the beam search.
(2)We have proposed a speaker-independent word-based HMM speech recognition system using Separate Vector Quantization(Band-Division Separate VQ HMM Speech Recognition). The proposed system can reduce the effects of extemal noise added to the speech and changes of utterance influenced by noise at the same time. From results of experiments we obtained 5-16 % higher recognition rate than conventional HMM speech recognition system.
(3)We have proposed a system for isolated -word recognition using two level --- syllable level and word level --- HMM. Compared to the conventional HMM isolated- word recognition system, this system can reduce the amount of memory for models. The more the vocabulary is large, the more this merit is effective. With 500 words recognition, this system reduces the amount of memory by 61 percent Using this system, the word recognition rate for 100 words by a woman speaker is 99.4 percent. This rate is equal to that of conventional system , and amount of memory for models is reduced by 44 percent.
|
