摘要
本文提出一种新的矢量量化方法压缩语音特征用于孤立字语音识别.新方法借鉴了隐马尔柯夫模型(HMM)中状态的概念来规划模板;用动态规划(Dyna-mic Progamming)的技术优化矢量量化产生的初始码字.新方法使得识别所需的模板库的尺寸明显减小,识别响应时问缩短,而且由于模板的优化,使得系统的正确识别率显著提高.
The paper presents a new VQ method to compress speech features for isolated speech recognition.The method cites from the concept of“State”in Hidden Markov Model(HMM),and uses Dynamic programming technique to optimize the VQ codewords.It not only decreases the size of pattern database, but also makes recognition more faster.Since the templates are optimized by using DP method,the correct recognition rate is high.
关键词
特征抽取
元音识别
矢量量化
feature extraction
dynamic programming/speech recognition
vector quantization
