Integration of phoneme pattern recognition with hidden Markov models to enhance performance of low level speech recognition

The hidden Markov model (HMM) is commonly employed in automatic speech recognition (ASR). The HMM has been shown to have a good performance in many applications, although it has some well-known limitations in modelling speech. Therefore, the standard HMM topology has been modified in a variety of ways to reduce errors, including factorization of the HMM into multiple-streams. However, the gap between the theoretical representation of speech and its acoustic representation can be further reduced. This paper describes a new method of correcting the HMM based on matching two dimensional templates of word time-frequency patterns to assist in low level speech recognition. This is shown to be a promising method to enhance speech recognition performance.