Nippon Electric Co., Ltd.

Nec Corporation

Hiroaki Sakoe

Seibi Chiba

Masao Watari

Shichiro Tsuruta

For use in recognizing an input pattern consisting off feature vectors positioned at a first, . . . , an i-th, . . . , and an I-th pattern time instant along a pattern time axis, a connected word recognition system comprises first through N-th neural networks B(1) to B(N) which are assigned to reference words identified by a first, . . . , an n-th, . . . , and an N-th word identifier and are arranged along a signal time axis divisible into a first . . . , a j-th, . . . , and a J-th signal time instant. The n-th word identifier n corresponds to consecutive ones of the pattern time instants by a first function n(i). The time axes are related to each other by a second function j(i). Among various loci (n(i), j(i)) in a space defined by the word identifiers and the time axes, an optimum locus (n(i), j(i)) is determined to maximize a summation of output signals of the respective neural networks when the consecutive pattern time instants are varied between the first and the I-th pattern time instants for each word identifier. The input pattern is recognized as a concatenation of word identifiers used as optimum first functions n(i) in the optimum locus. Preferably, the optimum locus is determined by a dynamic programming algorithm. If possible, use of a finite-state automaton is more preferred.

Connected word recognition system including neural networks arranged

A pattern matching system including a calculating arrangement for calculating a current value of cumulative distances between an input pattern and a reference pattern of a current pattern number at a current input pattern time instant and a current reference pattern time instant by using a dynamic programming algorithm which decreases workload because the current value is selected as a selected distance only when the current value is not greater than a threshold value which is determined at the current input pattern time instant. Work areas or memories are used as a result memory device accessed by a combination of the current input and reference pattern time instants for memorizing or storing the selected distance, the current pattern number, and the current reference pattern time instant, which are for later use in calculating another of the cumulative distances. Preferably, the threshold value should monotonously increase with time. More preferably, two of the work areas are alternatively used.

Pattern matching system

In a neural network, input neuron units of an input layer are grouped into first through J-th input layer frames, where J represents a predetermined natural number. Intermediate neuron units of an intermediate layer are grouped into first through J-th intermediate layer frames. An output layer comprises an output neuron unit. Each intermediate neuron unit of a j-th intermediate layer frame is connected to the input neuron units of j'-th input layer frames, where j is variable between 1 and j and j' represents at least two consecutive integers, one of which is equal to j and at least one other of which is less than j. Each output neuron unit is connected to the intermediate neuron units of the intermediate layer. For recognition of an input pattern represented by a time sequence of feature vectors, each consisting of K vector components, where K represents a predetermined positive integer, each input layer frame consists of K input neuron units. Each intermediate layer frame consists of M intermediate neuron units, where M represents a positive integer which is less than K. The vector components of each feature vector are supplied to the respective input neuron units of one of the input layer frames that is preferably selected from three consecutively numbered input layer frames. The neural network is readily trained to make a predetermined one of the output neuron units produce an output signal indicative of the input pattern and can be implemented by a microprocessor.

Multi-layer neural network to which dynamic programming techniques are

A pattern matching apparatus for comparing an input pattern of features with a reference pattern. Address information is stored in the reference pattern along with features, so that branching in a work memory which stores cumulative distances between the input and reference patterns may be effected. In this manner, memory requirements for storing reference patterns are reduced, and the number of required distance calculations also is reduced.

Pattern matching apparatus employing compensation for pattern deformation

A voice verification system in which multiple generic reference patterns are obtained by speaking the password in a number of different ways and in which a speaker specific reference pattern is generated by the speaker undergoing registration. A subset of the generic reference patterns are selected having the greatest similarity to the registered speaker's pattern. During verification, the speaker's identity is verified if the dissimilarity between the input pattern and the registered speakers reference pattern is both less than any dissimilarity between the input pattern and the selected generic reference patterns and also less than a threshold value.

Speaker verification system

For improved pattern recognition, the reference pattern feature sequence contains control parameters (operators) which provide branching and/or omission of those portions of words which may be non-standard due to speaker or dialect deformations. A pattern matching apparatus comprising a stack controller with two PUSH/POP stacks for addressing reference patterns to be used for correlations against the detected pattern information. The reference patterns may contain control characters indicating alternative reference pattern segments or segments which may be omitted.

Pattern matching apparatus

A continuous speech recognition system determines the similarity between input patterns and reference patterns over time such that similarities between previously spoken speech patterns and reference patterns are determined while speech continues to be spoken. Degrees of dissimilarity at arbitrary reference pattern word times are determined asymptotically and are recorded. The minimum degree of dissimilarity is determined and the corresponding word is categorized. Recognition decisions are ultimately made in reverse chronological order.

Continuous speech recognition system

There is provided a voice recognition system comprising a standard pattern memory in which a voice pattern of a predetermined word is stored as a positive reference pattern and also voice patterns of words similar to but different from the first-mentioned word are stored as negative reference patterns, a pattern comparator for calculating dissimilarities of an input voice pattern with respect to the positive reference pattern and negative reference patterns, and a discriminator for providing a coincidence confirmation output signal when the dissimilarity with respect to the positive reference pattern is less than a predetermined threshold value and less than the dissimilarities with respect to the negative reference patterns while otherwise rejecting the result of recognition.

Voice recognition system

A connected word recognition system operable according to a DP algorithm and in compliance with a regular grammar, is put into operation in synchronism with successive specification of feature vectors of an input pattern. In an m-th period in which an m-th feature vector is specified, similarity measures are calculated (58, 59) between reference patterns representative of reference words and those fragmentary patterns of the input pattern, which start at several previous periods and end at the m-th period, for start and end states of the reference words. In the m-th period, an extremum of the similarity measures is found (66, 69, 86), together with a particular word and a particular pair of start and end states thereof, and stored (61-63). Moreover, a particular start period is selected (67, 86) and stored (64). A previous extremum found and stored (61) during the (m-1)-th period for the particular start state found in the (m-1)-th period, is used in the m-th period as a boundary condition in calculating each similarity measure. After all input pattern feature vectors are processed, a result of recognition is obtained (89) by referring to the stored extrema, particular words, particular start states, and particular start periods.

DP Matching system for recognizing a string of words connected according

A system for recognizing a continuously spoken word sequence with reference to preselected reference words with the problem of coarticulation removed, comprises a pattern memory for memorizing demi-word pair reference patterns consisting of a former and a latter reference pattern segment for each reference word and a word pair reference pattern segment for each permutation with repetition of two words selected from the preselected reference words. A recognition unit is operable as a finite-state automaton on concatenating the demi-word pair reference patterns so that no contradiction occurs at each interface of the reference patterns in every concatenation. It is possible to use the automaton in restricting the number of reference patterns in each concatenation either to an odd or an even positive integer.

System operable as an automaton for recognizing continuously spoken

A pattern matching device generally comprises a first circuit (36) for calculating an elementary similarity measure between two feature vectors, one and the other selected from two feature vector sequences representative of two patterns, respectively, and a second circuit (37) for iteratively calculating a recurrence formula which defines a recurrence value by a sum of such an elementary similarity measure and an extremum of a prescribed number of previously calculated recurrence values. The recurrence formula eventually gives an overall similarity measure between the two patterns. The elementary similarity measure is now calculated by calculating a primitive similarity measure by a conventional circuit (15) and subtracting a predetermined value therefrom by a compensation circuit (31). Preferably, the second circuit (37) comprises circuitry (41, 42) for preventing the sum from overflowing outwardly of a preselected range.

Pattern matching device operable with signals of a compressed dynamic

This pattern matching system features the calculation of a weighting factor based on the variable interval between feature vector samples. On carrying out matching of two information compressed patterns, a weighted similarity measure calculator (64) calculates a weighted similarity measure by multiplying an intervector similarity measure between one each feature vector of the respective patterns by a weighting factor calculated by the use of a variable interval between each feature vector and a next previous one. A recurrence formula is calculated by the use of such weighted similarity measures instead of the intervector similarity measures. A predetermined value .delta. may be used in reducing the number of signal bits used for the recurrence formula. Preferably, a sum for the recurrence formula is restricted by two preselected values. Most preferably, an additional similarity measure is used for the recurrence formula.

Pattern matching device with a DP technique applied to feature vectors

Speaker recognition is decided by a similarity measure (D) calculated from comparing selected feature vectors among an input speech signal sequence of feature vectors (A) and a selected sequence (B) of reference vectors selected from a plurality of pre-stored reference sequences. Prior to comparison of the input and reference vector sequences, the two sequences are time normalized to align corresponding feature vectors. A significant sound specifying signal (V) including a time sequence of elementary signals is generated in synchronism with one of the input and reference sequences and indicates which feature vectors in that one of the input and reference sequences are considered to represent significant sound. The similarity measure (D) is then calculated in accordance with the comparison of those feature vectors in the one sequence which are indicated by the significant sound specifying signal as representing significant sound and the corresponding feature vectors of the other sequence.

Speaker recognizer in which a significant part of a preselected one of

Operation of a continuous speech recognition system operable according to the dynamic programming technique, is controlled by a state transition diagram in compliance with which word sequences to be recognized by the system with reference to a predetermined number of reference words B.sup.n 's are pronounced. The system comprises a state transition table accessed by the reference words B.sup.n 's to successively produce particular states y's in the diagram and previous states z's for each particular state y. In cooperation with a recurrence value and an optimum parameter table, a matching unit determines a recurrence value T.sub.y (m) and an optimum parameter set ZUN.sub.y (m) according to: ##EQU1## where u and m represent an end and a start point of a fragmentary pattern A(u, m) of an input pattern A representative of a word sequence and D(u, m, n), a similarity measure between the fragmentary pattern A(u, m) and a reference word B.sup.n assigned to a permutation of the previous and the particular states z and y. By referring to the optimum parameter table and, as the case may be, to the recurrence value table, a decision unit recognizes the word sequence as a concatenation of optimum ones of the reference words B.sup.n 's.

System for recognizing a word sequence by dynamic programming and by the

A similarity calculator for calculating a set of similarity measures S(A(u, m), B.sup.c)'s according to the technique of dynamic programming comprises an input pattern buffer for successively producing input pattern feature vectors of an input pattern A to be pattern matched with reference patterns B.sup.c, an m-th input pattern feature vector a.sub.m at a time. The similarity measure set is for a set of fragmentary patterns A(u, m)'s defined by a common end point m and start points u's predetermined relative to the end point m. Scalar products (a.sub.m .multidot.b.sub.j.sup.n) are calculated between the m-th input pattern feature vector and reference pattern feature vectors b.sub.j.sup.n of an n-th reference pattern B.sup.n and stored in a scalar product buffer. Recurrence values are calculated according to a recurrence formula for each end point m, rather than for each fragmentary pattern set, and for each reference pattern B.sup.n to provide a similarity measure subset S(A(u, m), B.sup.n)'s, with a recurrence value for each reference pattern feature vector b.sub.v calculated by the use of the scalar product (a.sub.m .multidot.b.sub.v) and recurrence values calculated for a previous end point (m-1) and for at least three consecutive reference pattern feature vectors preselected relative to that reference pattern feature vector b.sub.v. Instead of the scalar product, it is possible to use any one of other measures representative of a similarity or a dissimilarity between an input pattern feature vector and a reference pattern feature vector.

Growing community of inventors

Tokyo, Japan

Hiroaki Sakoe