基于混合模型CHMM和MLP的数码语音识别系统

Digital Speech Recognition System Based on Hybrid Model of CHMM and MLP

针对传统的CHMM应用于语音识别系统存在的缺点,提出了一种由CHMM和MLP网构成的混合模型。该混合模型将MLP网引入到CHMM中来计算每个状态的输出概率,通过MLP网的非线性预测能力代替CHMM中的似然估计值对输出概率进行分析、分类,从而加强和提高CHMM的语音识别能力。实验结果表明,将该混合模型应用到语音识别系统中,其识别效果明显优于基于传统的CHMM的识别系统。

In order to overcome shortcomings of traditional CHMM applied in speech recognition system, the paper proposed a hybrid model composed by CHMM and MLP network. The hybrid model introduced MLP in CHMM to calculate output probability of each state. In order to strengthen and improve speech recognition ability of CHMM, it analyzed and classified output probability by nonlinear prediction capability of MLP network instead of likelihood estimated value in CHMM. The experiment result showed that the hybrid model applied in speech recognition system has higher recognition ratio than that of traditional CHMM.