Noise robust speaker recognition under unknown noise environment

Many approaches designed to increase the performance and stability of the speaker verification system under adverse conditions were studied, performance will be at its peak when no mismatch occurs between training and testing conditions. Therefore, among all these methods, the Parallel Model Combination (PMC) appears to be the most adequate and capable techniques to handle such issue, where it compensates by minimizing the mismatch occurring between the test and the training conditions. In this study the main goal is to increase the performance of the speaker verification system. In previous studies, the (PMC) method was used to estimate the noisy speech parameters by using clean speech and noise model, assuming noise statistics are known. In this study, it is assumed that noise is not known. Noise is estimated using common VAD techniques from the noisy speech. Accordingly non-speech that is characterized by a certain VAD technique can be considered to estimate the noise model. In this study two common VAD techniques are used to to estimate the noise model, and PMC is used to estimate the noisy speech for all methods. The method that estimates the noise model directly from the noise signal is referred to the baseline method.Thereafter, the performance of the baseline is compared with that of the VAD techniques. NIST 1988 speaker recognition databases and NOISEX-92 databases were used to evaluate the performance of the speaker verification system. Expiremental results shows that the performance of the method that used the VAD techniques to estimate the noise model is comparable with the baseline method in the case of high signal-to-noise-ratio (SNR) levels, however in the case of low SNR levels, baseline method yielded better results in terms of equall error rate (EER).