Model-based articulatory phonetic features for improved speech recognition

We describe a neural based articulatory phonetic inversion model to improve the recognition of the acoustically varying vowels and the syllable initial plosives. The model uses a set of continuous valued articulatory phonetic features (APFs) to explore the interactions between the motor control of articulators and the acoustic phonetic events. We demonstrate that the neural model gives more accurate and robust recognition performance on the TIMIT sentences. The model offers two salient properties: it allows asynchronous feature changes at phoneme boundaries, and it accounts for the dual aspects of human speech production and perception through a heuristic learning algorithm during APFs mapping.