Modelling durations in hidden Markov models with application to word spotting

The conventional hidden Markov model models speech fairly well with the exception of parameters that are segment-based rather than frame-based. Duration of different acoustic segments, e.g. phonemes, falls into this category. Methods are described which solve this problem. For computational purposes sub-optimal methods are often used. This paper also deals with an improvement to such an algorithm that is in current use. Some results utilizing this algorithm are given