HMM Based Acoustic-Phonetic Decoding with Constrained Transitions and Speaker Topology

Abstract

Used for acoustic-phonetic decoding, hidden Markov models (HMM) perform well, but they induce phoneme boundaries problems; furthermore, the inter-speaker variability makes probability distributions difficult to learn. Two efficient signal processing techniques are used in a HMM-based phonetic decoding system: the Forward-Backward Divergence method detecting discontinuities of the speech signal, which are used to constrain the phonetic transitions between the models; the auto-regressive vector model allowing a definition of a speaker topology, which improves the quality of the training set. A new decoding system using these two processings is compared with a standard HMM-based system on the TIMIT database. -A significant 4% improvement of the accuracy rate is observed on the 7000 test phonemes.

Thanks to R. André-Obrecht for her software performing the FBD method.