Evaluation of speech enhancement techniques for speaker recognition in noisy environments.

摘要

In automatic speaker recognition (ASR) applications, the presence of background noise severely degrades recognition performance. There is a strong demand for speech enhancement algorithms capable of removing background noise. In this thesis, a Gaussian mixture model based automatic speaker recognition system is used for evaluating the performance of five different speech enhancement techniques. Previously, it was shown that these techniques improved the SNR of the speech signals corrupted by noise but their effect on the speaker recognition performance was not fully investigated. In this work, we implement these enhancement techniques and evaluate their performance as preprocessing blocks to the ASR engine. We evaluate the performance based on speaker recognition accuracy, average segmental signal-to-noise ratio and perceptual evaluation of speech quality (PESQ) scores. We combine clean speech from the TIMIT database with eight different types of noise from the NOISEX-92 database representing synthetic and natural background noise samples and analyze the overall system performance. Simulation results show that the system is capable of reducing noise with little speech degradation and the overall recognition performance can be improved at a range of different signal-to-noise ratios (SNR) with different noise types. Furthermore, results show that different enhancement techniques have different strengths and weaknesses, depending on their application and the background noise type.