A stochastic articulatory-to-acoustic mapping as a basis for speech recognition

Hidden Markov models of speech acoustics are the current state-of-the-art in speech recognition, but these models bear little resemblance to the processes underlying speech production. We describe two stochastic models of speech production: maximum likelihood continuity mapping (MALCOM) and conditional observable maximum likelihood continuity mapping (CO-MALCOM). The main component of these models is a stochastic mapping between acoustics and speech articulation. Perhaps surprisingly, the parameters of the stochastic mapping can be found using only acoustic data. We discuss theoretical and experimental reasons to believe that MALCOM and CO-MALCOM learn a stochastic mapping between articulator positions and speech acoustics. CO-MALCOM can be combined with standard recognition algorithms to do speech recognition based on a production model