Further Studies in Acoustic Modeling for Continuous Mandarin Speech Recognition

• Main Authors: Pei-Jie Hong, 洪培傑
• Other Authors: Lin-shan Lee
• Format: Others
• Language: zh-TW
• Published: 1999
• Online Access: http://ndltd.ncl.edu.tw/handle/85958209795056036079

碩士 === 國立臺灣大學 === 電信工程學研究所 === 87 === In continuous speech recognition, acoustic modeling is always a very important part. No matter in large vocabulary speech recognition or other application of speech recognition like key-word spotting, dialogue system etc..., there is always the need of good acoustic model. So the improvement of acoustic modeling technique is indeed very important. There are two parts of the thesis. The first part concerned with methods to improve the performance of decision tree. In this part, we consider from two respects. The first one is the constructing process of the decision tree. The other one is the way to use the question set. In the second part, we focus on the recognition problems that may cause by spontaneous speech and the solutions to solve the problems. We mention two problems, unclear pronunciation and phone missing. To solve the first problem, we try to use different transition topology to solve it. The performance is dependent to the testing data, but it is still better than the original models. To solve the phone missing problem, we use the hypothesized syllable to extend the range of syllable space. Hope that even syllables with phone missing can still be recognized.