Stochastic resonance in speech recognition: differentiating between /b/ and /v

Proposes that a neural mechanism has evolved which utilizes ambient noise in the recognition of weakly uttered, barely distinguishable monosyllable segments of speech, or phonemes; this mechanism fitting the characteristics of stochastic resonance. SR was originally conceptualized in 1980 to account for the periodicity of the ice ages, and has thus far been found to account for many natural, physical, and biological phenomena. Our experiments are averages over simulations, consisting of adding artificially produced midband (with respect to the system sampling frequency) Gaussian noise to pure analog speech waveforms. Conclusions are drawn from the comparison of true-positive to false-positive recognition results. The speech recognition algorithm relies on a recently published technique that finds the minimum difference between event times over a variety of time-scales and time-shifts. Our results propose a neurological supplement to the theoretical and experimental findings of physiological stochastic resonance, expressed in the form of Brownian motion assisting cochlear channel gating of sub-threshold sound signals.