Comparative evaluation of voice activity detectors in single microphone noise reduction algorithms

Abstract

Purpose: Voice activity detection (VAD) is significantly associated with the performance of single microphone noise reduction algorithms. In this study, the cause and extent of the VAD effect on noise reduction performance were analyzed by comparing conventional VAD algorithms with ideal VAD. Methods: A Wiener filter and logMMSE estimator employing various VAD algorithms were implemented using Simulink models and simulated with various 30-second speeches mixed with different noises at -5, 0 and 5 dB SNR. In addition, a noise estimation method not using VAD was also compared to facilitate a more extensive evaluation. The VAD and noise estimation accuracy were analyzed and objective speech quality measures such as fwSNR, PESQ and LLR were calculated. Results: Statistical analysis showed that the noise reduction performance using ideal VAD was significantly better than using conventional methods, the performance of which is dependent on noise types and SNR conditions. Conclusions: Consequently, there is still much room to improve VAD algorithms to have a high non-speech hit rate and low non-speech false alarm rate for better noise reduction performance in all noisy circumstances.