Speech recognition apparatus for recognizing the category of an input speech pattern

Claim 1

A speech recognition apparatus for recognizing an input speech comprising: (a) a means for evaluating a noise pattern N by time-averaging feature vectors Nl, where l is a subscript indicative of a frame number, within a prescribed noise detection interval prior to inputting the input speech, each of said feature vectors Nl being composed of a plurality of channel components evaluated by frequency-analyzing environmental noise; (b) a first memory for storing said noise pattern N; (c) a means for frequency-analyzing the input speech to evaluate a feature vector A i , where i is a subscript indicative of a speech frame number, each composed of a plurality of channel components of a plurality of channels corresponding to frequency components of the input speech at a prescribed speech frame period; (d) a means for evaluating a speech feature vector B i of the input speech by subtracting said noise pattern N from each feature vector A i of the input speech evaluated in each of said speech frame when the speech is inputted after said evaluation of the noise pattern; (e) a means for evaluating frame power in each of said speech frame from said speech feature vector B i to detect a start point and an end point of the input speech using said frame power; (f) a means for converting said speech feature vector B i of the input speech to local peaks vector R i including, a means for evaluating a least square fit line from said speech feature vector B i in each of said speech frame, a means for evaluating a spectrum normalized speech feature vector Z i composed of components Z i k , where i is a subscript indicative of a speech frame number and k is a superscript indicative of the channel number k=1 to K, by spectrum normalizing said speech feature vector B i using said least square fit line, and a means for evaluating a local peaks vector R i of the input speech composed of binary components r.sub.[1]i k by making a binary value "1" the components Z i k of said spectrum normalized speech feature vector Z i corresponding to channels being the local maximum in a frequency direction of said vector Z i while making a binary value "0" the other components Z i K of said vector Z i , said local maximum being extracted based on the following conditions: (Z i k >Z i k+1 ) and (Z i k-1 <Z i k ) and (Z i k >0), wherein k=1 to K, Z i 0 =-∞, and Z i K+1 =-∞, (g) a second memory for storing an input pattern composed of a time series of said local peaks vector R i of the input speech; (h) a third memory for storing a plurality of previously prepared reference patterns each of which is represented by a time series of local peaks vectors and corresponds to some word to be recognized respectively; (i) a means for evaluating pattern similarities between an input pattern composed of a time series of local peaks vectors R i from the start point to the end point of an input speech and a plurality of said reference patterns of the same format as that of said input pattern by making use of pattern matching; and (j) a means for judging a word name given to a reference pattern having a maximum pattern similarity among said pattern similarities evaluated for each reference pattern as a recognized result.