Statistical Model-Based Voice Activity Detection Using Spatial Cues and Log Energy for Dual-Channel Noisy Speech Recognition

In this paper, a voice activity detection (VAD) method for dual-channel noisy speech recognition is proposed on the basis of statistical models constructed by spatial cues and log energy. In particular, spatial cues are composed of the interaural time differences and interaural level differences of dual-channel speech signals, and the statistical models for speech presence and absence are based on a Gaussian kernel density. In order to evaluate the performance of the proposed VAD method, speech recognition is performed using only speech signals segmented by the proposed VAD method. The performance of the proposed VAD method is then compared with those of conventional methods such as a signal-to-noise ratio variance based method and a phase vector based method. It is shown from the experiments that the proposed VAD method outperforms conventional methods, providing the relative word error rate reductions of 19.5% and 12.2%, respectively.