A Novel Approach For Speaker Verification Under Various Noisy Environments

Since voice always has large amount of information, it is widely used as one of biometric features. However, the most challenging problem of its real-world applications is to reduce the damaging effect of noise. Many speaker verification approaches have succeeded in achieving high accuracy in clean conditions but failed in practical applications. This problem occurs partly due to the mismatch between enrollment(labs) and test(real-world) acoustic conditions.In order to compensate the mismatch between enrollment and test acoustic conditions, this thesis presents a novel approach based on Gaussian Mixture Model-Universal Background Model(GMM-UBM) algorithm. Noisy background adaption is proposed to make speaker models more close to the one in real-world scenarios. Considering the superiority of Maximum A-Posteriori(MAP) adaptation in GMM-UBM, noisy speeches are used as training sequences of background model so that adapted speaker models can become closer to the one in test environment. In testing step, at first noise power spectral density is estimated from noisy speech, and then features extracted from the estimated noise signal are used for noise recognition. In this way the appropriate GMM-UBM can be correctly chosen according to certain noise type. Moreover, in order to improve the overall performance of the speaker verification system, this thesis has done many experiments to achieve the best combination. In noise reduction, a 2-step Mel-domain de-noise method is used as a suitable front-end. In voice activity detection, based on pitch estimation method, it extracts feature vectors only from the voiced waveform, for there is little speaker information in unvoiced waveform, in that way to improve the correct recognition rate furthermore. In noise estimation, by means of minimum mean-square error(MMSE) optimal estimation we can estimate the noise power spectral. In noise identification, we couple many features and models and provide a comparative performance analysis of them. At last an excellent couple is obtained through the comparison.Testing environments include various different types of noise. The clean speeches, from both English and Chinese speech corpus, mixed with noise at the SNR of 5dB, are used as testing sequences. Finally, experiments on corrupted speeches showed that the proposal could reduce up to 26.38% and an average of 16.44% equal error rate(EER) compared to the baseline, indicating its advantages on speaker verification under various noisy conditions.