Design of Speaker-lndependent Chinese Character Speech Recognition System

Abstract

ln this thesis a speaker -independent speech recognition system of Chinese characters is presented. Using this systemˊwe do not have to train the s- ystem every time a new user wishes to use it. The speech recognition system is based on the use of one or more reference patterns for each character to be recognized rather than storing every us- erˊs training data. The multiple reference patterns are created by a clu- stering algorithm which is based on the shared nearest neighbor concept. ln the recognition procedureˊ a combination of the nearest neighbor deci- sion rule and the k-nearest-neighbor decision rule is used to cjassjfy the test utterance. ln building the recognition system many aspectsˊ such as endpoint detectionˊ time alignmentˊ feature extraction and distance mea- sureˊ are also considered in this thesis. The speech recognition experiment is done by use of Chinese spoken digits as sample characters and the results are quite satistactory.