A robust feature extraction method based on CZCPA model for speech recognition system

To solve the problem that recognition rates of most speech recognition systems decrease severely in the noise environments, this paper presents a new feature extraction method based on zero-crossing with peak amplitude model combining difference information. The frequency information of speech signal is obtained by Computing the zero-crossing intervals of speech signal and difference speech signal. And the intensity information of speech signal is obtained by detecting peak amplitudes between the intervals and making nonlinear amplitude compression for speech and its difference signal. Thus, speech features having relationship with auditory model are obtained. The recognition network employs BP neural network or HMM. And a 50 isolated words speech recognition system based on above features is simulated in different signal-noise ratio. It is showed that new features have better robustness than traditional features.