System for recognizing speech continuously spoken with number of word or words preselected

• US 4,049,913 A
• Filed: 11/01/1976
• Issued: 09/20/1977
• Est. Priority Date: 10/31/1975
• Status: Expired due to Term

• Current Assignee: Nippon Electric Company Limited (NEC Corporation)
• Original Assignee: Nippon Electric Company Limited (NEC Corporation)
• Inventors: Sakoe, Hiroaki
• Primary Examiner(s): Claffy, Kathleen H.
• Assistant Examiner(s): Kemeny, E. S.
• Application Number:
  US05/737,454
• Time in Patent Office: 323 Days
• Field of Search: 179/1.5 A, 179/1.5 D
• US Class Current: 704/241
• CPC Class Codes: G06F 40/216 using statistical methods
• CPC Class Codes: G06F 40/284 Lexical analysis, e.g. tokenisation or collocates
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence
• CPC Class Codes: G10L 15/12 using dynamic programming techniques, e.g. dynamic time warping [DTW]

Abstract

A continuous speech recognition system comprises a word number specifier for specifying, as the number of continuously spoken word or words, either a single integer or a set of different integers. The single integer may be manually or automatically adjusted. In compliance with the specified word number or numbers, the system carries out pattern matching between an input pattern representative of the spoken word or words and a predetermined number of reference patterns. The matching may be carried into effect by dynamic programming. The input pattern is recognized to be one of the reference patterns or to be a concatenation of some or all of the reference patterns, equal in number either to the single integer or to one of the different integers.

Claims

1. In a continuous speech recognition system comprising an input pattern buffer memory for temporarily memorizing an input pattern A represented by a time sequence of input pattern feature vectors a_i, I in number, as:

space="preserve" listing-type="equation">A = a.sub.1, a.sub.2, . . . , a.sub.i, . . . , a.sub.I,

said input pattern being representative of a first predetermined number k of continuously spoken words, said first predetermined number being equal at least to unity;

a reference pattern memory for memorizing reference patterns Bⁿ representative of a second predetermined number of words n, each reference pattern being represented by a time sequence of referene pattern feature vectors b_jⁿ, Jⁿ in number, as;

space="preserve" listing-type="equation">B.sup.n = b.sub.1.sup.n, b.sub.2.sup.n, . . . , b.sub.j.sup.n, . . . , b.sub.J.spsb.n.sup.n ;

and means responsive to said input and reference pattern feature vectors for recognizing said input pattern to be a concatenation of the reference patterns, k in number;

the improvment wherein said means comprises a word number specifier for specifying the number K of words of said input pattern, said word number K being preselected for a predetermined duration to be equal at least to unity with reference to said first predetermined number; and

recognition means responsive to said word number K and said input and reference pattern feature vectors for deciding said concatenation.

2. A continuous speech recognition system as claimed in claim 1, said word number K being preselected to be equal to said first predetermined number k, wherein said recognition means comprises first matching means responsive to said input and reference pattern feature vectors for finding a partial similarity measure S<

p, m> and

a partial recognition result n<

p, m>

, said partial similarity measure being defined as a maximum of similarity measures S[A(p, m), Bⁿ ] between each partial pattern A(p, m) and said reference patterns, said partial pattern being defined by a part of said input pattern feature vector sequence as;

space="preserve" listing-type="equation">A(p, m) = a.sub.p+1, a.sub.p+2, . . . , a.sub.m,

said partial recognition result being one of said reference patterns that gives said maximum;

second matching means responsive to said word number K and the partial similarity measures for finding a maximum of summations, each being calculated for the partial similarity measures of those partial patterns, K in number, into a concatenation of which said input pattern is divided; and

decision means responsive to said word number K and the partial recognition results for deciding said input pattern to be the partial recognition results for those partial patterns of a concatenation for which said maximum summation is found.

3. A continuous speech recognition system as claimed in claim 2, wherein said second matching means comprises means responsive to said partial similarity measures for finding segmentation points at which said input pattern is divided into those partial patterns of the concatenation for which said maximum summation is found, and said decision means comprises means responsive to said word number K for successively specifying consecutive integers between unity and said word number K and means responsive to said partial recognition results, said consecutive integers, and said segmentation points for successively selecting those partial recognition results with refrence to said segmentation points which are for said partial pattern concatenation.

4. A continuous speech recognition system as claimed in claim 1, wherein said word number specifier comprises means for successively specifying said word number K as a finite sequence of different integers k, said integers including said first predetermined number k, and said recognition means comprises first matching means responsive to said input and reference pattern feature vectors for finding a partial similarity measure S<

p, m> and

a partial recognition result n<

p, m>

, said partial similarity measure being defined as a maximum of similarity measures S[A(p, m), Bⁿ ] between each partial pattern A(p, m) and said reference patterns, said partial pattern being defined by a part of said input pattern feature vector sequence as;

space="preserve" listing-type="equation">A(p, m) = a.sub.p+1, a.sub.p+2, . . . , a.sub.m, overall

said partial recognition result being one of said reference patterns that gives said maximum;

second matching means comprising, in turn, first means for successively providing serial integers x between unity and a greatest one of said different numbers and second means responsive to each of said serial integers and the partial similarity measures for finding an overall similarity measure T<

k>

defined as a maximum of summations, each being calculated for the similarity measures of those partial patterns, x in number, into a concatenation of which said input pattern is divided;

third matching means responsive to said word number K and the overall similarity measures for deciding said first predetermined number to be one of said different integers that gives a maximum of the overall similarity measures found by said second matching means for said different integers; and

decision means responsive to the first predetermined number found by said third matching means and the partial recognition results for deciding said reference pattern concatenation.

5. A patterns speech recognition system as claimed in claim 4, wherein said second matching means further comprises third means responsive to said serial integers and said partial similarity measures for finding a point sequence of provisional segmentation points at which said input pattern is divided into those partial paterns of the concatenation for which said maximum summation is found, and said decision means comprises means responsive to the first predetermined number found by said third means for successively specifying consecutive integers between unity and the last-mentioned predetermined number and means responsive to said partial recognition results, said consecutive integers, and the provisional segmentation points for finding those segmentation points among the point sequences at which said input pattern is divided into the partial patterns, equal in number to said first predetermined number, the partial recognition results for the last-mentioned partial patterns giving said reference pattern concatenation.

6. A continuous speech recognition system as claimed in claim 5, further comprising fourth means operatively coupled to said second matching means for finding in response to said serial integers, said partial recognition results, and said overall similarity measures a sequence of provisional recognition results for those partial patterns, x in number, of a concatenation into which said input pattern is divided to maximize the summation therefor, wherein said segmentation finding means comprises means operatively coupled to said fourth means and said consecutive integer specifying means for selecting a provisional recognition result sequence which consists of the provisional recognition results, equal in number to said first predetermined number, and which maximizes the overall similarity measure.