A Statistical Approach for Voiced Speech Detection

Detection of Voice in speech signal is a challenging problem in developing high-performance systems used in noisy environments. In this paper, we present an efficient algorithm for robust voiced speech detection and for the application to variable-rate speech coding. The key idea of the algorithm is considering speech energy and zero crossings rate (ZCR) information simultaneously when processing speech signals and finding the end point of the signal. Next to it a decision rule and a background noise statistics estimator, by applying a statistical model. A robust decision rule is derived from the generalized likelihood ratio test (LRT) by assuming that the noise statistics are known a priori. The algorithm is most efficient for the time-varying noise. According to our simulation results, the proposed algorithm shows significantly better performance in low signal-to-noise ratio and in noisy environments.