A Robust Feature for Speech Detection
Atanas Ouzounov
Institute of Information Technologies, 1113 Sofia
E-mail: atanas@iinf.bas.bg
Abstract: A new robust feature intended for speech detection is proposed in this paper. This feature is the mean of the absolute values of the delta spectral autocorrelation function of the power spectrum and it is named the Mean-Delta (MD) feature. For noisy speech, the trajectory variations of the MD feature are compared with those ones of the other robust feature - the Energy-Entropy (EE) feature, developed in [6]. The selected noisy speech samples from two databases (the SpEAR database [8] and the BG-SRDat corpus [9]) are used in the experimental study. In all the tests the MD feature has demonstrated the same or better results than those of the EE.
Keywords: speech detection, voice activity detection, spectral autocorrelation, speaker recognition.