Word spotting method taking account of duration change characteristics for stable and transient parts of speech

This paper describes a new word spotting method which takes into consideration the duration change characteristics of stable and transient parts of speech. An isolated word to be used as a reference pattern is divided into phoneme-like segments which have minimum and maximum permissible segment durations. This paper proposes two techniques for setting segment durations. One is based on phoneme-context rules. The other is based on supervised learning. The word spotting process is carried out using a segment-durationcontrolled dynamic-programming algorithm. The word detection rate for the 7 keywords in 13 sentences spoken by 5 speakers was 94.2% using the rulebased method and 96.9% for the supervised-learning method.