Multi-SNR GMMs-Based Noise-Robust Speaker Verification Using 1/fa Noises

摘要

Environmental noises greatly degrade the performance of speaker verification system. This paper designs a novel noise-robust speaker verification system based on multi-SNR techniques using 1/f^alpha noises. During training phase, the 1/f^alpha noises are employed to simulate environmental noises. To simplify the computation load, an optimal value of alpha is selected. To evaluate the performance of the system, speaker verification experiments are conducted on 5 kinds of environmental noises: "airport", "bus", "lobby", "restaurant", and "office" at 0 dB SNR. The experimental results demonstrate that a notable degradation of verification error rates is obtained while using 1/f^alpha noises