Speech recognition apparatus

Claim 1

A speech recognition apparatus comprising an input terminal for receiving an analog speech signal, and a sequence pattern matching circuit responsive to said analog speech signal for producing a sensible manifestation of a word in said speech signal, said pattern matching circuit including: an A/D converter coupled to said terminal for converting said analog speech signal into a digital signal at a predetermined sampling frequency; feature parameter calculating means responsive to said digital signal for calculating a sequence of feature parameters A=a 1 a 2 . . . a n of said speech signal; an input signal pattern memory for storing the calculated sequence of the feature parameters received from said feature parameter calculating means; a reference pattern memory for storing sequences of reference feature parameters B=b 1 b 2 . . . b m of reference patterns to be compared with the calculated feature parameters; comparison means comprising a distance matrix computing unit for computing a distance matrix {d ij } between the calculated feature parameters and the reference feature parameters of one of the reference patterns for i+r≧j≧i-r where r is a window width; a distance matrix buffer for storing elements d i ,j on lattice points (i,j) of the calculated distance matrix; a cost function computing unit for computing values of cost function g(i,j) on the basis of the elements d i ,j for the lattice points satisfying the equation i+j=M+lN where M is a constant and N≧3; a cost function buffer memory for storing the computed values of cost function, said cost function buffer memory comprising memory regions for storing the values G(r), . . . , G(0), . . . , G(-r) of the cost function; control means for effecting a successive iteration computation of the cost function for l=0,1,2 . . . , the paths to reach each lattice point on i+j=M+lN where a corresponding cost function is to be calculated by said cost function computing unit being selected by said control means to be one of at least two oblique paths having the same distance from the lattice points on i+j=M+(l-1)N; and means responsive to completion of the comparison of the calculated feature parameters with said reference feature parameters for producing a sensible manifestation of a word in said analog speech signal.

Claim 25

An apparatus according to any one of claims 1 to 23 wherein the two sequences of feature parameters are A=a n , . . . a 0 , a 1 , . . . a n and B=b m , . . . b 0 , b 1 , . . . b m , where a 1 and b 1 are regarded as forming start points of the sequences A and B and render the cost function value g(k) (where k=j-i) corresponding to the point satisfying i+j=M into zero.

Claim 27

A method of effecting speech recognition by use of a sequence pattern matching procedure to determine the similarity between two sequences of feature parameters A=a 1 . . . a n of an input speech signal and feature parameters B=b 1 . . . b m of a reference pattern, comprising the steps of sampling an input speech signal at a predetermined frequency to produce a sequence of sample data, computing a feature parameter for each predetermined number of sample data forming one frame period, detecting the start of the speech signal, storing into an input speech pattern memory feature parameters of successive frames, said storing step successively storing all the computed feature parameters regardless of whether they precede or follow the detected start of the speech signal, reading out of the input speech pattern memory, upon detection of the start of the speech signal, feature parameters n' frames before the detected start of the speech signal and succeeding feature parameters to obtain a sequence of feature parameters A=a -n' a -n'+1 . . . a 0 a 1 . . . a n where n' is an integer satisfying n'≧(r-1)/2 and r is a window width, computing distance matrix elements d i ,j at lattice points (i,j) defined between the two sequences of the feature parameters A and B satisfying i+r≧j≧i-r, executing a successive iteration computation of dynamic programming by computing a cost function g(i,j) for all the lattice points satisfying i+j=2+l, i≧1, j≧1 and l=0, 1, 2, . . . starting with all the lattice points satisfying i+j=2, i≧-n', J≧-n' and i+r≧j≧i-r to determine a minimum score value for each sequence of the reference feature parameters, and producing a sensible manifestation of a word in said input speech signal upon completion of said sequence pattern matching procedure.

Claim 28

A method of effecting speech recognition by use of a sequence pattern matching procedure to determine the similarity between two sequences of feature parameters A=a 1 . . . a n of an input speech signal and feature parameters B=b 1 . . . b m of a reference pattern, comprising the steps of sampling an input speech signal at a predetermined frequency to produce a sequence of sample data, computing a feature parameter for each predetermined number of sample data forming one frame period, detecting the start of the speech signal, storing into an input speech pattern memory feature parameters of successive frames, said storing step successively storing all the computed feature parameters regardless of whether they precede or follow the detected start of the speech signal, reading out of the input speech pattern memory, upon detection of the start of the speech signal, feature parameters n' frames before the detected start of the speech signal and succeeding feature parameters to obtain a sequence of feature parameters A=a -n' a -n'+1 . . . a 0 a 1 . . . a n where n' is an integer satisfying n'≧(r-1)/2 and r is a window width, computing distance matrix elements d i ,j at lattice points (i,j) defined between the two sequences of the feature parameters A and B satisfying i+r≧j≧i-r, executing a successive iteration computation of dynamic programming by computing a cost function g(i,j) for all the lattice points satisfying i+j=2+l, i≧1, j≧1 and l=0, 1, 2, . . . starting with all the lattice points satisfying i+j=2, i≧-n', J≧-n' and i+r≧j≧i-r to determine a minimum score value for each sequence of the reference feature parameters, and producing a sensible manifestation of a word in said input speech signal upon completion of said sequence pattern matching procedure.