Comparison of keyword spotting methods for searching in speech

This paper presents and discusses keyword spotting methods for searching in speech. In contrast with searching in text, the searching in speech or generally in multimedia data still represents a challenge. The aim of the paper is to present a keyword spotting (KWS) method based on a large vocabulary continuous speech recognition (LVCSR) system, based on phonetics decoder, and keyword spotting using a filler model. All the methods are evaluated and compared from various points of view – speed, quality, requirements on training data and so on. All experiments are done using a telephone-quality speech corpus. Furthermore, this paper presents a new block decision in filler model-based keyword spotting which brings the speedup of decision together with better detection.