Abstract: The purpose of speech recognition is able to extract the content of the speech in various conditions and transform from speech to text automatically. Based on human hearing perception mechanism, we propose a decomposition method using independent parallel subbands for speech recognition. In this method, wavelet processing is introduced into Hidden Markov Model (HMM) and Fuzzy Neural Network (FNN) is used to improve the convergence speed and to avoid local optimal in speech detection. Experiment results show that the proposed hybrid speech recognition model is more robust when noise presents. We combine the dynamic modeling of CDHMM and the classification capability of fuzzy neutral network, this has been a hot research area in recent years and can be applied to speech-text transform devices.