Improving Speaker Verification in Noisy Environments using Adaptive Filtering and Hybrid Classification Technique

摘要

This study describes two approaches of improving speaker verification in noisy environments. The first approach is implementation of a speaker verification classification technique base on hybrid Vector Quantization (VQ) and Hidden Markov Models (HMMs) in clean and noisy environments. The second approach is implementation of Adaptive Noise Cancellation (ANC) as pre-processing for noise removal. The motivation to implement hybrid classification technique is to improve the HMMs performance. It is shown that, by using the hybrid technique, an Equal Error Rate (EER) of 11.72% is achieved compared to HMM alone, which achieved 16.66% in clean environments. However, both techniques show degradation in noisy environments. In order to address these problems, an Adaptive Noise Cancellation (ANC) technique using adaptive filtering is implemented in the pre-processing stage due to its ability to separate overlapping speech frequency bands. Investigations using Least-Mean-Square (LMS), Normalized Least-Mean-Square (NLMS) and Recursive Least-Squares (RLS) adaptive filtering are conducted to find the best solution for the speaker verification system.