不定人语音识别系统

Speaker Independent Speech Recognition System

采用了说话人自适应和大数量人的训练数据两种方法解决不定人语音识别问题.在离散隐马尔可夫模型DHMM孤立字语音识别系统中,采用了3种说话人自适应方法,方法1,方法2为码本自适应,方法3为新人数据自适应,并将新建的1000人的语音库用于不定人语音识别.实验结果表明,说话人自适应方法均有一定的自适应效果,特别是多码本自适应后,识别率可提高16%,达到93%以上;大数据库的采用,使得不定人的数字识别率达到了94%.

To solve Speaker-Independent Speech Recognition problem,we use two kinds of method:speaker adaptation and large amount of people's training data. In Discrete Hidden Markov Model (DHMM)isolated word recognition system,we use three adaptation methods. Method 1 and method 2 are codebook adaptation, method 3 is new speaker's data adapta-tion, We also use. new speech database containing 1 000 people for speaker-independent speech recognition. Results of experiments indicate that speaker adaptation has an effect on improving recognition accuracy to a certain extent, specially after multi-codebook adaptation, recognition rate can be increased about 16% and is above 93%. Using large speech database to train Speaker-independent speech recognition system, its recognition rate can reach 94%.