Conditioning speech recognition on sensorimotor principles – Curt Boylls (Department of Defense)
To establish better paradigms for automated speech recognition may take new science. On the other hand, if we assume that speech operates within the same neural framework that governs other sensorimotor processes, then there is plenty of “old science” that we might yet profitably exploit.In this talk, we will look at how the neuroscience community developed new theory and methodology for discovering and describing the control of complex sensorimotor tasks. We’ll also introduce some outrageous analogies with the present situation in speech recognition. Our story starts about 25 years ago, when sensorimotor neuroscience was suffering from its own version of the “independence assumption”: the notion that the brain programs the actions of the body by playing a keyboard somehow connected to thousands of independently controllable state variables (e.g., muscle tensions and related kinematics) inherent in the skeletomuscular system. We will see how that concept was displaced by the idea that, through the imposition of neurally-mediated constraints among states, the effective dimensionality of the skeletomotor apparatus is reduced to an exceptionally low order, one manageable within the resources available to attention and short-term memory. Part of this constraining operation includes the segregation (conceptually and anatomically) of processes that detail the events defining an action versus processes that modulate those sequences (in, say, rate and intensity). Each of these elements typically operates within its own characteristic timescale; and what we observe during skeletomotor activity is the superposition of all
their influences.If facts from speech eventually confirm the tortured conjectures drawn here from studies of locomotion and postural control, then corresponding recognizer architectures are fairly obvious. But the confirmatory step may not be required if one merely wishes to benefit today’s technology; and we will conclude with some speculations toward that end.
Curt Boylls, a neurophysiologist, has been with the Department of Defense for 11 years.