Spoken term detection based on acoustic speech segmentation

Abstract

This paper presents a system for spoken term detection in a continuous speech stream. Spoken terms are predefined through a set of acoustic examples provided by the users. Spoken term detection proceeds in two steps: speech segmentation and term verification. We suggest the use of an acoustic-based algorithm for the segmentation which exploits acoustic particularities of the speech stream to detect word frontiers. From the segmentation stage a collection of utterances are delimited, and they are aligned with the spoken term acoustic representation. Tests were conducted on an Arabic corpus using spectral features of the signal. A correct detection rate of about 100% was reached with a false alarm of 20%.