Hybrid Model of Wigner Distribution/HMM and Self-organizing Neural Network for Speech Recognition

Abstract

Based on the time-varying character of speech, this paper makes full use of excellent characters in Wigner distribution (WD) and combines the homomorphic processing technique, proposes a new cepstrum coefficient, WD-MFCC. And apply this coefficient in a hybrid model of hidden Markov models and a self-organizing neural network for speech recognition in noisy environment. Experiments use three cepstrum coefficients (MFCC, DPSCC and WD-MFCC) into three speech recognition models respectively, results prove that with low SNR, compared with traditional continuous density HMM model (CDHMM with pure speech) and CDHMM-N model (CDHMM adding addictive noise), the hybrid model which introduced in this paper with WD-MFCC can obviously improve the recognition rate and the performance of speech recognition system in noisy environment.