Speech recognition apparatus

Claim 2

A speech recognition apparatus comprising: partial pattern forming means for obtaining a partial pattern A.sub.(p,q) =a p ,a p+1 , . . . , a q (where 1≦p≦q≦I) whose starting point is time point i=p in an input pattern A=a 1 , a 2 , . . . , a i , . . . , a I , which is expressed in the time series of feature vectors, and whose endpoint is time point i=q; reference pattern memory means for storing the reference patterns B n for each word number "n" of a reference pattern B n =b 1 n , b 2 n , . . . , b j n , . . . , b jn n , which is previously set for each word number "n"; normalized similarity measure computing means in which a function j(i) is set up which makes the time base "i" of the partial pattern obtained by said partial pattern forming means correspond to the time base "j" of the reference pattern stored by said reference pattern memory means, and a maximum similarity measure S{A.sub.(p,q),B n }as a maximum value of the sum of the vector similarity measure s {a i , b n , j(i)} as defined by said time base "i" and said function j(i) is obtained as a normalized similarity measure as defined by a normalizing function r[q-p+1, Jn, S{A.sub.(p,q), B n }] which provides a normalized similarity measure not dependent on the time duration t=q-p+1 of said partial pattern and the time duration of said reference pattern; partial maximum similarity computing means for maximizing for each "p" the normalized similarity measure obtained by said normalized similarity measure computing means by using the dynamic programming algorithm, and for obtaining a maximum value of the result of said maximization for "n", as a partial maximum similarity measure; ##EQU22## partial decision means for successively repeating the operation that the registered word "n" providing the partial maximum similarity measure D q obtained in said partial maximum similarity computinq means is treated as a partial decision result W q , in the range of q=1 to I, and for obtaining the partial similarity measure D i and the partial decision result W i (i=1 to I) at all the time points along the time series of the input pattern; and output means in which the largest partial maximum similarity measure of those measures obtained in said partial decision means, is treated as a maximum similarity measure D max = i max {Di} (where i=1 to I), and the partial decision result W i providing that maximum similarity measure D max is obtained, and said result is output as the recognition result.

Claim 3

An apparatus according to claim 1, in which said normalizing function for providing a normalized similarity measure in said normalized similarity measure computing means is given as an average similarity measure for each unit time, and mathematically expressed by: ##EQU23##

Claim 4

An apparatus according to claim 1, in which said normalizing function for providing a normalized similarity measure in said normalized similarity measure computing means is given as an average similarity measure for each unitary path length in the pattern matching, and mathematically expressed by: ##EQU24## where p ass =matching path length.

Claim 5

An apparatus according to claim 1, in which said normalizing function for providing a normalized similarity measure in said normalized similarity measure computing means is given by: ##EQU25## where α, β, and γ are constants.

Claim 6

A speech recognition apparatus comprising: input pattern forming means for obtaining a partial pattern A.sub.(p,q) =a p , a p+1 , . . . , a q (where 1≦p<q≦I) whose starting point is time point i=p in an input pattern A=a 1 , a 2 , . . . , a i , . . . , a I , which is expressed in the time series of feature vectors, and whose endpoint is time point i=q; reference pattern memory means for storing the reference patterns B n for each word number "n" of a reference pattern B n =b 1 n , b 2 n , . . . , b j n , . . . b jn n , which is previously set for each word number "n"; similarity measure computing means in which a function j(i) is set up which makes the time base "i" of the partial pattern obtained by said input pattern forming means correspond to the time base "j" of the reference pattern stored by said reference pattern memory means, and a maximum similarity measure S{A.sub.(p,q), B n } of the sum of the vector similarity measure s {a i , b j n .sub.(i) } as defined by said time base "i" and said function j(i) is obtained; first intermediate similarity measure computing means for obtaining said obtained similarity measure S{A.sub.(p,q), B n }, as a maximum similarity measure, as mathematically given below, of the maximum results for "n" as given by maximizing, by using the dynamic programming algorithm, for "p" at time point i=q, the sum of the weighting function U p =0 (where p=1) and U p =U.sub.(p) (where 1≦p≦I) for the time duration p-1 of the partial pattern A.sub.(1, p-1) =a 1 , a 2 , . . . , a i , . . . , a p-1 whose starting point is time point i=1 of the input pattern and the endpoint is time point i=p-1 ##EQU26## means for successively repeating the operation that the registered word "n" providing the intermediate similarity measure D q obtained in said first intermediate similarity computing means is treated as an intermediate decision result W q , in the range of q=1 to I, and for obtaining the intermediate similarity measure D i and the intermediate decision result W i (i=1 to I) at all the time points along the time series of the input pattern; maximum similarity computing means for obtaining said intermediate similarity measure D i as a maximum similarity measure which is given below and the largest normalized similarity measure of those measure R q =D q +U q (where q=1 to I) as defined by the weighting function U q =0 (where q=I) and U q =U.sub.(I-q) (where 1<q<I) for the time duration "I-q" of the partial pattern: A.sub.(q+1 I) =a.sub.q+1, a.sub.q+2, . . . , a.sub.i, . . . , a.sub.I of an input pattern whose starting point is said time point i=q+1 and the endpoint is time point i=I; and output means for obtaining the intermediate result W i giving said maximum similarity measure D max and outputting a registered word giving said W i as the recognition result.

Claim 7

A speech recognition apparatus comprising: input pattern forming means for obtaining a partial pattern A.sub.(p,q) =a p , a p+1 , . . . , a q (where 1≦p<q≦I) whose starting point is time point i=p in an input pattern A=a 1 , a 2 , . . . , a i , . . . , a I , which is expressed in the time series of feature vectors, and whose endpoint is time point i=q; reference pattern memory means for storing the reference patterns B n for each word number "n" of a reference pattern B n =b 1 n , b 2 n , . . . , b j n , . . . b jn n , which is previously set for each word number "n"; similarity measure computing means in which a function j(i) is set up which makes the time base "i" of the partial pattern obtained by said input pattern forming means correspond to the time base "j" of the reference pattern stored by said reference pattern memory means, and a maximum similarity measure S{A.sub.(p,q), B n } of the sum of the vector similarity measure s {a i , b j n .sub.(i) } as defined by said time base "i" and said function j(i) is obtained; first intermediate similarity measure computing means for obtaining said obtained similarity measure S{A.sub.(p,q), B n }, as a maximum similarity measure, as mathematically given below, of the maximum results for "n" as given by maximizing, by using the dynamic programming algorithm, for "p" at time point i=q, the sum of the weighting function U p =0 (where p=1) and U p =U.sub.(p) (where 1≦p≦I) for the time duration p-1 of the partial pattern A.sub.(1, p-1) =a 1 , a 2 , . . . , a i , . . . , a p-1 whose starting point is time point i=1 of the input pattern and the endpoint is time point i=p-1 ##EQU26## means for successively repeating the operation that the registered word "n" providing the intermediate similarity measure D q obtained in said first intermediate similarity computing means is treated as an intermediate decision result W q , in the range of q=1 to I, and for obtaining the intermediate similarity measure D i and the intermediate decision result W i (i=1 to I) at all the time points along the time series of the input pattern; maximum similarity computing means for obtaining said intermediate similarity measure D i as a maximum similarity measure which is given below and the largest normalized similarity measure of those measure R q =D q +U q (where q=1 to I) as defined by the weighting function U q =0 (where q=I) and U q =U.sub.(I-q) (where 1<q<I) for the time duration "I-q" of the partial pattern: A.sub.(q+1 I) =a.sub.q+1, a.sub.q+2, . . . , a.sub.i, . . . , a.sub.I of an input pattern whose starting point is said time point i=q+1 and the endpoint is time point i=I; and output means for obtaining the intermediate result W i giving said maximum similarity measure D max and outputting a registered word giving said W i as the recognition result.