基于FVQ和HMM模糊训练的语音识别方法

本文提出了一种基于模糊矢量量化（ＦＶＱ）和隐马尔柯夫模型（ＨＭＭ）模糊训练的语音识别新方法．利用模糊矢量量化替代了传统方法中的矢量量化，语音特征参数序列经过模糊矢量量化后得到模糊观察符号序列．在此基础上提取出一个反映训练样本总体随机变化特性的模糊观察符号序列，然后用它对该音节的ＨＭＭ进行一次性全局训练，训练算法经传统的Ｂａｕｍ－Ｗｅｌｃｈ算法改进得到．经十个汉语数字的对比实验表明，该训练算法大大提高了系统的训练速度，模糊矢量量化与传统的矢量量化相比，不仅提高了隐马尔柯夫模型的鲁棒性，进而提高了系统的识别率，而且在语音训练数据不充足的情况下，也能得到很好的识别性能．

Fuzzy Vector Quantization and Fuzzy Training of HMM's for Speech Recognition

In this poper, a new speech recognition approach based on fuzzy vector quantization andfuzzy training of HMMs is presented. A fuzzy vector quantizer replacing of the traditional vectorquantizer iiss used to transform the continuous sets of speech feature vectors into a fuzzy observationsequence which can represent the overall random variety feature of the training sets. By use of thefuzzy observation sequence, the overall training of the HMM's is done once and for all. The trainingalgorithm developed is a modification of the Baum-Welch reanimation algorithm, and can quicken thetraining speed of HMM's. An FVQ/HMM method is investigated by comporitive experiments to greatlyimprove the performance of VQ/HMM based speech recognition system even with deficient availabletraining data.