Improvement of Statistical model-based noise-robust voice activity detector

Abstract

Statistical model-based voice activity detector (SMVAD) is a robust algorithm in various noise conditions to detect speech region from input signal using noise and noisy speech statistical models such as complex Gaussian probability density function (PDF). The decision rule of SMVAD is based on likelihood ratio test (LRT). However, the LRT-based decision rule may cause detection errors because of statistic properties of noise and speech signal. In this paper, we analyze the reasons why the detection errors occur. To decrease the detection errors, we propose two modified decision rules using reliable likelihood ratios (LRs) determined by spectral power of each frequency bin. We also propose a weighting scheme considering spectral characteristics of noise and speech signal. To decrease the spectral variation of same type of noise signal, in addition, we propose a spectral smoothing method of input signal and explain the effects of this method. The performances of our proposed methods are evaluated by receiver operating characteristic (ROC) curves and compared with three conventional methods in various noise environments. In most of noise conditions, the proposed methods show better performance than conventional methods. The experimental results also show that the proposed weighting scheme, which is applied to each LR, can guarantee the most stable performance improvement of SMVAD.