Sub-band Selection-based Speech Detector for Robust Keyword Recognition under Noisy Enviroment

M. Ji and H.R. Kim (Korea)

Keywords

EPDVAA (End Point Detector based on Voice ActivityAlgorithm), speech detection, keyword recognition

Abstract

In this paper, we propose a sub-band selection-based speech detector for robust keyword recognition. In order to show the excellence of the proposed algorithm, we compare it with a conventional speech detection algorithm, which is called EPD-VAA (End-Point Detector based on Voice Activity Algorithm). Their performances, advantages and disadvantages are compared and evaluated. The proposed algorithm is the combination of word boundary detection algorithm using ATF (Adaptive Time-Frequency) parameters and EPD-VAA. The proposed speech detector is trained so as to better extract keyword speech than other speech. EPD-VAA usually works well under high SNR, but it doesn’t work well any more in low SNR. Furthermore it cannot find the exact start point of the speech, especially if the speech starts with fricatives or nasal sounds. But in case of the proposed one, useful bands are selected through keyword training. Therefore it decides the boundary of the speech according to the energy levels of the sub-bands, which correspond to the pre-defined useful bands. Experimental results indicate that the proposed speech detector outperforms the EPD-VAA under noisy environment.

Important Links:



Go Back