Feature space derivation for isolated utterance reading with discriminant deformable models

The following paper proposes a speech-reading method that combines discriminant deformable model-based approach to individual frame analysis with the hidden Markov model technique, applied to a sequence analysis. A discriminant deformable model, which is a modification of the original deformable grid paradigm, allows for fast phone-image analysis, however it requires derivation of new feature space for the purpose of grid-matching result evaluation. This critical issue has been a focus of research presented in the paper, yielding matched-grid representation, which provides good class separation for several phones. Discrete-observation HMMs are used for sequence analysis and the experiments are restricted to speaker-dependent isolated digit reading.