Noise-robust speech detection using spectral variation information

Abstract

This paper deals with a new parameter for voice detection which is used in various speech engineering areas such as speech synthesis, speech recognition and speech coding. Coefficient of variation (CV) of speech spectrum as well as other feature parameters is used for the detection of speech. CV is calculated only in the specific range of speech spectrum and gives information of existence of pitch in speech. Average magnitude and spectral magnitude are also employed to improve the performance of detector. Those parameters helps to detect unvoiced sound on start point. We evaluate the performance of detector by frame difference between proposed algorithm and hand labeled one. Proposed algorithm is compared with energy -based speech detection algorithm which using LCR instead of ZCR. Energy-based algorithm with noise reduction by Kalman filter is also compared with proposed algorithm. From the flag value which is results of three feature parameters, we are able to detect speech boundaries more accurately in low SNR noisy environment. Proposed algorithm outperform conventional energy-based algorithm in most kinds of noise and coefficient of variation parameter can be used to voice activity detection or real-time end point detection algorithm.