Keyword spotting based on the analysis of template matching distances
This paper presents a system for speakerindependent keyword spotting (KWS) in continuous speech usinga spoken example template. The approach, based on DynamicTime Warping (DTW) for matching the template to a testutterance, does not require any modelling or training as requiredin alternative techniques such as the Hidden Markov Model(HMM). This is of particular relevance to applications such asdetection of words that have not been adequately represented ina training database (e.g. searching for topical words that areemerging in society). Introduced is the use of the DTW distancehistogram for automatic estimation of similarity thresholds forevery keyword-utterance pair. Experiments conducted on a widerange of speech sentences and keywords show that when only afew examples of the keyword are available, the proposed systemhas higher recall ratio than a HMM-based approach.