Sub-band selection-based speech detector for robust keyword recognition in low SNR

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 2 conventional speech detection algorithms. Their performances, advantges, and disadvantages are compared and evaluated. One is EPD-VAA (End Point Detector - Voice Activity Algorithm). The other is word boundary detection algorithm using ATF (Adaptive Time-Frequency) parameters. The proposed speech detection algorithm is combination of these two speech detectors. The proposed speech detector is trained so as to better extract keyword speech detectors. The EPD-VAA usually works well under high SNR, but it doesn``t work well any more in low SNR environment. Therefore the frequency parameter that is relatively stronger to noise than the parameter is used here. And the EDP-VAA algorithm cannot find the exact start point of the speech, especially if the speech starts with fricatives of nasal sounds. In order words, the speech detector finds robustly only islands of reliability. Hence when the speech detector declares the start of the speech boundary, it usually puts a few extra frames before the start frame and after the end frame in order not to cut the first and last sound of the speech. The proposed algorithm is designed for a rubust keyword spotter. Before the speech detector works, it is trained with pre-defined keywords. Consequently it makes the speech detector to work well for keywords. Before the speech detector is ready to detect the speech boundary, useful bands for each keyword should be selected. Once useful bands are selected, the speech detector decides the boundary of speech according to the levels of sub-bands. Experimental results indicate that the proposed speech detector outperforms the EPD-VAA algorithm under low SNR and the speech boundaries for keywords are extracted more accurately. In addition, the start point of the speech boundary is detected more accurately than ...