矢量量化技术和隐马尔柯夫模型方法在韵母识别中的应用

本文应用矢量量化技术(Vector Quantization)和隐马尔柯夫模型方法(Hidden Markov Model)为一个全字汇量的孤立字普通话语音识别系统设计了韵母识别子系统.该系统由韵母信号析取器、滤波器阵列特征分析器、矢量量化器、预识别器、隐马尔柯夫模型匹配器和决策器组成.根据对汉语中1172个不同音节的语音信号测试结果,决策器输出的准确率(即系统的最后识别准确率)为89.5%,而前两个估计的识别准确率则达到97.2%.系统的训练包括生成矢量量化器的码字和为每一个韵母建立隐马尔柯夫模型,改进了Linde 等人提出的码字生成算法,提出了一个得到隐马尔柯夫模型参数的系统化方法.

Application of Vector Ouantization and Hidden Markov Model to recognition of Mandarin Finals

Based on the techniques of Vector Quantization(VQ)and Hidden Markov Model(HMM),a subsystem is proposed in this paper to recognize Ma- ndarin finals for a speech recognition system of isolated Mardarin words without being restricted by vocabulary.The system first isolates the final signals from the syllable waveform,and then performs filter bank analysis of the speech power spectrum.A vector quantizer is then employed to generate a symbol sequence for speech features.A pre-classifier is used before matching HMM to make the classification decision.The system was tested with 1172 diff- erent Mandarin syllables and a correct recognition rate of 89.5% was obtained. The training of the system includes the creation of a VQ codebook and the sett- ing up of HMMs for each final.Linde's algorithm for codebook generation has been modified and a systematic method to obtain HMM initial parame- ters is proposed.