Sensitive keyword spotting for crime analysis

Keyword spotting is the task of detecting keywords of interest in continuous speech. This work investigates the application of keyword spotting to detect crime and it can be used along with telephone tapping and audio monitoring devices by security organization. In this work phonetic based word spotter is developed. A simple method to word spotting is to build up the keyword models with filler models to account for nonkeyword intervals. Different methods are used to model the keywords and fillers. The best performance achieved is 76% for a spotter using 39 phones as filler models. In order to increase the performance tri-phones are used to model the keywords. This approach attains the best performance with hit rate of 96% with false alarm rate of 1.984 FA/KW/H.