Speech Recognition, Using Analog Threshold Logic

Abstract

A speech-analysis system using analog threshold logic (ATL) for feature abstraction is being developed to recognize consonants in utterances of CVC words by a number of talkers. The ATL element, originally modeled after the biological neuron, has an output that is linearly proportional to the net sum of excitatory and inhibitory inputs, provided that this net sum is greater than some adjustable threshold. The speech-analysis system abstracts relatively steady-state as well as transient features over almost 60 dB of dynamic range from the logarithmitized outputs of 19 low-Q filters. Feature abstraction takes place in real time, does not require segmentation, utilizes parallel processing, and provides information about both the presence and magnitude of the frequency-energy-temporal characteristics of speech sounds. Many features corresponding to results of speech-recognition experiments described in the literature have been automatically abstracted by ATL networks, which measure the slope of the spectral envelope in real time. Using the experimental system, recognition scores in the range of 82%–99% were obtained for 16 consonants. Results for the fricative consonants are discussed and some of their invariant features described. [Work supported by the U. S. Air Force Research and Technology Division, Air Force Avionics Laboratory.]