Robust Speech Recognition Using Perceptual Wavelet Denoising and Mel-frequency Product Spectrum Cepstral Coefficient Features

To improve the performance of Automatic Speech Recognition (ASR) Systems, a new method is proposed to extract features capable of operating at a very low signal-to-noise ratio (SNR). The basic idea introduced in this article is to enhance speech quality as the first stage for Mel-cepstra based recognition systems, since it is well-known that cepstral coefficients provided better performance in clean environment. In this speech enhancement stage, the noise robustness is improved by the perceptual wavelet packet (PWP) based denoising algorithm with both type of thresholding procedure, soft and modified soft thresholding procedure. A penalized threshold was selected. The next stage of the proposed method is extract feature, it is performed by the use of Mel-frequency product spectrum cepstral coefficients (MFPSCCs) introduced by D. Zhu and K.K and Paliwal in [2]. The Hidden Markov Model Toolkit (HTK) was used throughout our experiments, which were conducted for various noise types provided by noisex-92 database at different SNRs. Comparison of the proposed approach with the MFCC-based conventional (baseline) feature extraction method shows that the proposed method improves recognition accuracy rate by 44.71 %, with an average value of 14.80 % computed on 7 SNR level for white Gaussian noise conditions.