A multi-resolution hidden Markov model using class-specific features

We address the problem in signal classification applications, such as automatic speech recognition (ASR) systems that employ the hidden Markov model (HMM), that it is necessary to settle for a fixed analysis window size and a fixed feature set. This is despite the fact that complex signals such as human speech typically contain a wide range of signal types and durations. We apply the probability density function (PDF) projection theorem to generalize the hidden Markov model (HMM) to utilize a different features and segment length for each state. We demonstrate the algorithm using speech analysis so that long-duration phonemes such as vowels and short-duration phonemes such as plosives can utilize feature extraction tailored to the their own time scale.