Phonemic classification in speech recognition system having accelerated response time

Claim 1

An acoustic processing system for a speech recognition system comprising: input means for inputting speech waveforms; extraction means, responsive to said input means, for extracting a feature speech parameter from a selected one of said speech waveforms, said feature speech parameter being indicative of the varying frequency of the speech waveform during a selected period of time; phonemic classifier means, responsive to said input means, for making a phonemic classification representative of said speech waveforms from a selected one of said speech waveforms, said phonemic classifier means including: correlation means, responsive to said input means, for calculating correlation functions representing the power of a said selected speech waveform; zero crossing frequency means, responsive to said input means, for calculating a zero crossing frequency for said selected speech waveform; differential zero crossing frequency means, responsive to said input means, for calculating the difference between the zero crossing frequency of said selected speech waveform sand a preceding selected speech waveform; power calculation means, responsive to said correlation means, for calculating from said correlation functions the power differential between a selected said speech waveform and a preceding speech waveform; threshold table means for storing a plurality of sets of threshold values, each said set of threshold values including values corresponding to those values produced by said correlation means, said zero crossing frequency means, said differential zero crossing frequency means and said power calculation means; total power calculation means, responsive to said correlation means and a logic detection means, for calculating the total power of said selected speech waveform, said total power being representative of the volume of the selected speech waveform; detection means, responsive to said total power calculation means, for selecting a set of threshold values stored in said threshold table means corresponding to the total power of said selected speech waveform; and said logic detection means includes means for making a phonemic classification in each of a plurality of analysis frames for said selected speech waveform responsive to said correlation means, zero crossing frequency means, differential zero crossing means, and power calculation means within said selected set of threshold values; normalization means, responsive to said extraction means and said phonemic classifier means, for normalizing said feature speech parameter to produce an acoustic pattern for said selected speech waveform; pattern memory means for storing a plurality of reference acoustic patterns; and comparison means, responsive to said normalization means and said pattern memory means, for comparing the acoustic pattern of said selected speech waveform with said plurality of reference acoustic patterns to determine a part of speech represented by said selected speech waveform.