SPEECH SIGNAL DETECTION IN A NOISY ENVIRONMENT USING NEURAL NETWORKS AND CEPSTRAL MATRICES

Juraj Kačur - Gregor Rozinaj - Sergio Herrera-Garcia

   In this article a new flexible speech detection method comprising two relatively modern approaches like artificial neural networks (ANN) and cepstral matrices is presented. Cepstral matrices obtained via linear prediction coefficients were chosen as the eligible speech features. This technique is known to provide reliable log spectrum estimation at a low cost. Furthermore, both spectral and time characteristics can be efficiently, which is an essential aim here. Several WSS noises and different SNR settings were tested. In the range of 3 to 13 dB the ANN approach remarkably outperformed the energy and zero crossing method and improved the accuracy of the other algorithm based on cepstral matrices as well.

Keywords: cepstral matrices, neural networks, MLP, speech detection, CLPC vectors, WSS noises

[full-paper]