An improved robust statistical voice activity detection based on sub-band periodic intensity
Weijun He, Xiaohui Feng, Zhengyu Zhu, Weili Zhou · 2015
From an investigation of the statistical model likelihood ratio test-based voice activity detection(VAD), it was discovered that there existed false alarm problem in detecting the non verbal vocalization signal. In this paper, an improved statistical model-based VAD method is proposed for noise adverse environments, which employs reserved coefficient in the decision rule. The reserved coefficient is determined by sub-bands periodic intensity, sub-bands are divided on the basis of human auditory sensing characteristic. The final decision depends upon the geometric mean of the reserved sub-band likelihood ratios. Simulation which is carried out on the CADCC and NOISEX-92 databases, shows its promising performance in comparison with traditional robust VAD methods in both stationary and nonstationary noise conditions, in terms of improved false alarm rate and receiver operating characteristic (ROC) curve.