Noise robust speaker verification using mel-frequency discrete wavelet coefficients and parallel model compensation

Interfering noise severely degrades the performance of a speaker verification system. The Parallel Model Combination (PMC) technique is one of the most efficient techniques for dealing with such noise. Another method is to use features local in the frequency domain. Recently, Mel-Frequency Discrete Wavelet Coefficients (MFDWCs) [1, 2] were proposed as speech features local in frequency domain. In this paper, we discuss using PMC along with MFDWCs features to take advantage of both noise compensation and local features (MFDWCs) to decrease the effect of noise on speaker verification performance. We evaluate the performance of MFDWCs using the NIST 1998 speaker recognition and NOISEX-92 databases for various noise types and noise levels. We also compare the performance of these versus MFCCs and both using PMC for dealing with additive noise. The experimental results show significant performance improvements for MFDWCs versus MFCCs after compensating the Gaussian Mixture Models (GMMs) using the PMC technique. The MFDWCs gave 5.24 and 3.23 points performance improvement on average over MFCCs for -6 dB and 0 dB SNR values, respectively. These correspond to 26.44% and 23.73% relative reductions in equal error rate (EER), respectively