WORD BOUNDARY DETECTION IN STATIONARY NOISES USING CEPSTRAL MATRICES

Juraj Kačur - Gregor Rozinaj

   A new method for word boundary detection in the noisy environment is presented here. It is based on cepstral matrices constructed from blocks of CLPC vectors using 1-dimensional DCT. Our detection system measures global variations of CLPC vectors, which is a simple task having cepstral matrices. An approach utilizing normalized CLPC vectors is very effective in colour noises. For SNR values ranging from 3 to 13 dB a 50 % reduction of the detection error was reached comparing to the classical method. For white like noises the same accuracy is obtained using only plain CLPC vectors. A reliable noise classification algorithm based on CLPC vectors is proposed, too, which integrates these approaches into a single one.

Keywords: Cepstral matrices, speech detection, CLPC vectors, WSS noises

[full-paper]