An Experimental Comparative Study of Three Robust Features for Speech Detection
Atanas Ouzounov
Institute of Information Technologies, 1113 Sofia
E-mail: atanas@iinf.bas.bg
Abstract:
The results from an experimental comparative study of three robust features intended for trajectory-based speech detection are presented in the paper. These features are the Mean-Delta (MD) feature [6], the Spectral Entropy (SE) [3] and the Spectral Entropy with Normalized frame Spectrum (SENS) [7]. Two experiments with noisy speech samples from two databases (the SpEAR database [2] and the BG-SRDat corpus [5]) are carried out. In the first experiment, the trajectory’s variations of the features are compared by visual evaluation on their graphical representations. In the second one, the noise influence on the features trajectories is estimated by computing of the Euclidean distances between z-normalized trajectories of clean speech examples and their noisy versions. Based on experimental results two main conclusions are made: in comparison with other two features the MD feature trajectories are significantly less influenced by different type of noises; the SENS and especially the MD feature are more suitable for trajectory-based speech detection than the SE.
Keywords: speech detection, voice activity detection, spectral entropy.