Continuous speech phoneme recognition using two stage probabilistic modelling

In this paper we address the problem of phoneme recognition in continuous speech using a two stage probabilistic modelling method. The sub-phonemic properties of a phoneme are represented in the first stage and the broad phonemic features are captured by the second stage. Homogeneous or inhomogeneous hidden Markov models (HMMs) can be used in each of the two stages for modelling. This method allows for modelling of context dependent durations of phonemes and subphonemes by using inhomogeneous HMMs (IHMMs). The performance of the new scheme with different combinations of HMM and IHMM in each stage is compared. The results of two experiments, one on a speaker independent TIMIT database and another on a speaker dependent database, are reporte