Speaker Recognition Robustness In Noisy Environments

Due to its special merits of flexibility, economy and accuracy, speaker recognition techonology has a broad application future in biometrics security field. However, the state-of-art techniques of speaker recognition have performed well under ideal conditions, while the practical results degraded distinctly. Thus the problem of improving the system robustness has turned into the most active research field.We can improve the robustness of speaker recognition system in many different ways, because this problem relates to every parts of the system. This thesis has investigated this problem from approaches such as packets loss compensation and modeling techniques with noise and limited training data, and proposes some novel approaches:1. Aiming at the problem of packets loss, this paper proposes a new method of lost packets compensation based on lagrangian interpolation. Traditional method of lost packets compensation compensates lost packets works well when the ratio of lost packets is low but deteriorates dramaticly when the ratio of lost packets increases.2. Since GMM needs large number of data for training, when lost packets ratio is high, the performance of GMM is poor. This work presents a new classifier GMM-DM which improves the performance of GMM classifier by introducing distortion measure when training data is inadequate due to the packets loss during transportation. Experiments carried out in this work show that the compensation based on lagrangian interpolation and the GMM-DM new classifier could obtain better results than traditional methods when the ratio of lost packets is relatively high.3. We have also investigated the robustness of speaker recognition in different level of simulated noise and real-world noise obtained from NOISEX-92 database using subband processing based on Analysis of Noise Energy Distribution (ANED) and decision fusion using GMM-DM. The general principle is to split the whole frequency domain into several subbands on which statistical recognizers are independently applied and then recombined to yield a global score and a global recognition decision.