Robust Features and Neural Network for Noisy Speech Detection

Atanas Ouzounov
Institute of Information Technologies-BAS
Acad. G. Bonchev Str. bl. 29A,
Sofia 1113, Bulgaria,
E-mail: atanas@iinf.bas.bg

Abstract

In this paper, the effectiveness of three features in speech detection tasks is experimentally studied. The first feature is obtained by processing of the spectral autocorrelation function (Ouzounov, 2004) while the second one is based on the multi-band spectral entropy (Misra et al, 2005). The well-known mel-cepstrum is utilized as a third feature. A multi-layer perceptron based speech detector is developed and speech detection tasks with noisy data are carried out for each feature. The performance analysis of the speech detection results is done using the ROC curves and measures. The experimental results revealed that the feature obtained by processing of the spectral autocorrelation function is more suitable for noisy speech detection than the other two features.