Speech recognition apparatus and method utilizing an orthogonalized dictionary

Claim 2

A speech recognition apparatus comprising: acoustic analyzing means for analyzing input speech to obtain an input speed pattern; sampling pattern generating means for generating one of a learning pattern for storage and a sampling pattern for recognition by extracting a predetermined number of data samples from the input speech pattern obtained from said acoustic analyzing means; orthogonalizing time filter means for computing an average pattern from each learning pattern generated by said sampling pattern generating means, and for smoothing the average pattern along a time axis thereof to generate dictionary data of a first axis and for differentiating the average pattern along the time axis thereof to generate dictionary data of a second axis, the dictionary data of the first and second axes being representative of the learning pattern; orthogonalized dictionary means for storing dictionary data of the first and second axes generated by said orthogonalizing time filter means; and similarity computing and speech recognition means for computing similarities between each input speech pattern acoustically analyzed by said acoustic analyzing means and converted into a sampling pattern by the sampling pattern generating means, and the dictionary data of the first and the second axes generated and stored by said orthogonalizing time filter means to recognize each sampling pattern based on the computed similarities with the dictionary data.

Claim 3

An apparatus according to claim 1, wherein said orthogonalizing time filter means computes an average pattern A.sub.(j,k) of each of a plurality of learning patterns a m (j,k) of a given category i of the learning patterns a m (j,k) using a data system based on the following equation: ##EQU8## wherein j represents a number of acoustically analyzed feature vectors of the predetermined number of data samples of the learning pattern, k represents a number of subintervals the learning pattern is divided into by the predetermined number of data samples, and m represents a number of learning patterns of the given category, j and k being positive integers; obtains dictionary data b 1 (j,k) of the first axis using the average pattern A.sub.(j,k) in accordance with the following equation: b.sub.1(j,k) =A.sub.(j,k-1) +2*A.sub.(j,k) +A.sub.(j,k+1) ; thereafter, obtains dictionary data b 2 (j,k), of the second axis using the average pattern A.sub.(j,k) in accordance with the following equation: b.sub.2(j,k) =-A.sub.(j,k-1) +A.sub.(j,k+1) ; and, normalizes dictionary data b 2 (j,k) of the second axis and registers normalized dictionary data in said orthogonalized dictionary means.

Claim 4

An apparatus according to claim 2, wherein said orthogonalizing time filter means performs reorthogonalization processing of the dictionary data of the second axis based on an equation expressed as follows: B.sub.2(j,k) =b.sub.2(j,k) -(b.sub.2(j,k) ·b.sub.1(j,k))b.sub.1(j,k) normalizes the reorthogonalized dictionary data B 2 (j,k), and registers the normalized dictionary data in said orthogonalized dictionary means as the dictionary of the second axis.

Claim 5

An apparatus according to claim 1, wherein said orthogonalizing time filter means obtains dictionary data b 1 (j,k) of the first axis from a plurality of learning patterns a m (j,k) of a given category i as follows: ##EQU9## wherein j represents a number of acoustically analyzed feature vectors of each learning pattern, k represents a number of subintervals the learning pattern is divided into by the predetermined number of data samples, and m represents a number of learning patterns of the given category, j and k being positive integers; and, obtains dictionary data b 2 (j,k) of the second axis expressed by the following equation from each of the learning patterns a m (j,k), and provides dictionary data obtained therefrom to said orthogonalized dictionary means; ##EQU10##

Claim 6

An apparatus according to claim 1, wherein said orthogonalizing time filter means obtains dictionary data b 1 (j,k) of the first axis expressed by the following equation from a plurality of learning patterns a m (j,k) of a given category i: b.sub.1(j,k) =A.sub.(j,k-2) +4*A.sub.(j,k-1) +6*A.sub.(j,k) +4*A.sub.(j,k+1) +A.sub.(j,k+2) ; wherein j represents a number of acoustically analyzed feature vectors of each learning pattern, k represents a number of subintervals the learning pattern is divided into by the predetermined number of data samples, and m represents a number of learning patterns of the given category, j and k being positive integers; obtains dictionary data b 2 (j,k) of the second axis expressed as: b.sub.2(j,k) =-A.sub.(j,k-2) -2*A.sub.(j,k-1) +2*A.sub.(j,k+1) +A.sub.(j,k+2) ; and, quadratically differentiates the dictionary data of the second axis based on the following equation to generate dictionary data b 3 (j,k) of a third axis: b.sub.3(j,k) =-A.sub.(j,k-2) -2*A.sub.(j,k-1) +3*A.sub.(j,k) -2*A.sub.(j,k+1) -A.sub.(j,k+2).

Claim 7

A speech recognition apparatus comprising: speech converting means for converting input speech into an electrical signal; acoustic analyzing means for acoustically analyzing the electrical signal; speech interval detecting means for detecting a speech interval of the acoustically analyzed electrical signal; sampling pattern generating means for extracting a predetermined number of sampling data points from a speech pattern of the speech interval detected by said speech interval detecting means so as to obtain one of a sampling pattern for recognition and a learning pattern for storage; sampling pattern storage means for storing a predetermined number of learning patterns generated by said sampling pattern generating means for a given category i of learning patterns wherein i is a positive integer equal to or larger than 1; orthogonalizing time filter means for computing an average pattern from each of the learning patterns of the given category i, and for smoothing the average pattern along a time reference axis thereof to generate dictionary data of a first axis and differentiating the average pattern along the time axis thereof to generate dictionary data of a second axis; orthogonalized dictionary means for storing the dictionary data of the first and second axes generated by said orthogonalizing time filter means from the learning patterns; orthogonalizing means for generating additional dictionary data orthogonal to dictionary data of the first and second axes stored by orthogonalized dictionary means from learning patterns stored in the sampling pattern storage means and for adding the additional dictionary data to said orthogonalized dictionary means; similarity computing means for computing similarities between each of the sampling patterns generated by the sampling pattern generating means and the dictionary data of the first and second axes stored in said orthogonalized dictionary means; and sampling pattern recognizing means for recognizing each of the sampling patterns based on similarities with the dictionary data of said orthogonalized dictionary means computed by said similarity computing means.

Claim 8

A speech recognition apparatus comprising: speech converting means for converting input speech into an electrical signal; acoustic analyzing means for acoustically analyzing the electrical signal; speech interval detecting means for detecting a speech interval of the acoustically analyzed electrical signal; sampling pattern generating means for extracting a predetermined number of sampling data points from a speech pattern of the speech interval detected by said speech interval detecting means so as to obtain one of a sampling pattern for recognition and a learning pattern for storage; sampling pattern storage means for storing a predetermined number of learning patterns generated by said sampling pattern generating means for a given category i of learning patterns wherein i is a positive integer equal to or larger than 1; orthogonalizing time filter means for computing an average pattern from each of the learning patterns of the given category i, and for smoothing the average pattern along a time reference axis thereof to generate dictionary data of a first axis and differentiating the average pattern along the time axis thereof to generate dictionary data of a second axis; orthogonalized dictionary means for storing the dictionary data of the first and second axes generated by said orthogonalizing time filter means from the learning patterns; orthogonalizing means for generating additional dictionary data orthogonal to dictionary data of the first and second axes stored by orthogonalized dictionary means from learning patterns stored in the sampling pattern storage means and for adding the additional dictionary data to said orthogonalized dictionary means; similarity computing means for computing similarities between each of the sampling patterns generated by the sampling pattern generating means and the dictionary data of the first and second axes stored in said orthogonalized dictionary means; and sampling pattern recognizing means for recognizing each of the sampling patterns based on similarities with the dictionary data of said orthogonalized dictionary means computed by said similarity computing means.

Claim 9

An apparatus according to claim 6, wherein said orthogonalizing means generates the dictionary orthogonal to said orthogonalized dictionary means by Gram Schmidt orthogonalization.

Claim 10

An apparatus according to claim 6, wherein said orthogonalizing time filter means computes an average pattern A.sub.(j,k) of each of a plurality of learning patterns a m (j,k) of the give category i using a data system based on the following equation: ##EQU11## wherein j represents a number of acoustically analyzed feature vectors of the sampling data points of the learning pattern, k represents a number of subintervals the learning pattern is divided into by the predetermined number of sampling data points, and m represents a number of learning patterns of the given category i, j and k being positive integers; obtains dictionary data b 1 (j,k) of the first axis using the average pattern A.sub.(j,k) in accordance with the following equation: b.sub.1(j,k) =A.sub.(j,k-1) +2*A.sub.(j,k) +A.sub.(j,k+1) ; thereafter, obtains dictionary data b 2 (j,k) using the average pattern A.sub.(j,k) in accordance with the following equation b.sub.2(j,k) =-A.sub.(j,k-1) +A.sub.(j,k+1) ; and, normalizes the dictionary data b 2 (j,k) of the second axis and registers the normalized dictionary data in said orthogonalized dictionary means.

Claim 11

An apparatus according to claim 8, wherein said orthogonalizing time filter means performs reorthogonalization processing of the dictionary data of the second axis based on an equation expressed as follows: B.sub.2(j,k) =b.sub.2(j,k) -(b.sub.2(j,k) ·b.sub.1(j,k))b.sub.1(j,k) normalizes the reorthogonalized dictionary data B 2 (j,k), and registers the normalized dictionary data in said orthogonalized dictionary means as the dictionary of the second axis.

Claim 12

An apparatus according to claim 6, wherein said orthogonalizing time filter means obtains dictionary data b 1 (j,k) of the first axis from learning patterns a m (j,k) of the given category i as follows: ##EQU12## wherein j represents a number of acoustically analyzed feature vectors of the sampling data points of the learning pattern, k represents a number of subintervals the learning pattern is divided into by the predetermined number of sampling data points, and m represents a number of learning patterns of the given category i, j and k being positive integers; and, obtains dictionary, data b 2 (j,k) of the second axis, expressed by the following equation, from each learning pattern a m (j,k), and provides the dictionary data obtained therefrom in said orthogonalized dictionary means: ##EQU13##

Claim 13

An apparatus according to claim 6, wherein said orthogonalizing time filter means obtains dictionary data b 1 (j,k) of the first axis expressed by the following equation from learning patterns a m (j,k) of the given category i: b.sub.1(j,k) =A.sub.(j,k-2) +4*A.sub.(j,k-1) +6*A.sub.(j,k) +4*A.sub.(j,k+1) +A.sub.(j,k+2) ; wherein j represents a number of acoustically analyzed feature vectors of the sampling data points of the learning pattern, k represents a number of subintervals the learning pattern is divided into by the predetermined number of sampling data points, and m represents a number of learning patterns of the given category i, j and k being positive integers; obtains dictionary data b 2 (j,k) of the second axis expressed as: b.sub.2(j,k) =-A.sub.(j,k-2) -2*A.sub.(j,k-1) +2*A.sub.(j,k+1) +A.sub.(j,k+2) ; and, quadratically differentiates the dictionary data of the second axis based on the following equation to generate dictionary data b 3 (j,k) of a third axis: b.sub.3(j,k) =-A.sub.(j,k-2) -2*A.sub.(j,k-1) +3*A.sub.(j,k) -2*A.sub.(j,k+1) -A.sub.(j,k+2).

Claim 14

A speech recognition apparatus comprising: speech converting means for converting input speech into an electrical signal; acoustic analyzing means for receiving the electric signal from said speech converting means and acoustically analyzing the electrical signal; speech interval detecting means for detecting a speech interval of the acoustically analyzed electrical signal; sampling pattern generating means for extracting a predetermined number of sampling data points from a speech pattern of the speech interval detected by said speech interval detecting means so as to obtain one of a sampling pattern for recognition and a learning pattern for storage; sampling pattern storage means for storing a predetermined number of learning patterns generated by said sampling pattern generating means for a given category i of learning patterns wherein i is a positive integer equal to or larger than 1; orthogonalizing time·frequency filter means, comprising at least three types of filters, for obtaining an average pattern from each of the learning patterns of the given category i, smoothing the average pattern along a time reference axis thereof to generate dictionary data of a first axis, differentiating the average pattern along the time axis to generate dictionary data of a second axis, and differentiating the average pattern along a frequency axis thereof to generate dictionary data of a third axis; orthogonalized dictionary means for storing the dictionary data of the first, second, and third axes generated by said orthogonalizing time·frequency filter means; similarity computing means for computing similarities between each of the sampling patterns generated by said sampling pattern generating means and the dictionary data of the first, second, and third axes stored in said orthogonalized dictionary means; and sampling pattern recognizing means for recognizing each of the sampling patterns based on similarities with the dictionary data of the first, second and third axes as computed by said similarity computing means.

Claim 15

A speech recognition apparatus comprising: speech converting means for converting input speech into an electrical signal; acoustic analyzing means for receiving the electric signal from said speech converting means and acoustically analyzing the electrical signal; speech interval detecting means for detecting a speech interval of the acoustically analyzed electrical signal; sampling pattern generating means for extracting a predetermined number of sampling data points from a speech pattern of the speech interval detected by said speech interval detecting means so as to obtain one of a sampling pattern for recognition and a learning pattern for storage; sampling pattern storage means for storing a predetermined number of learning patterns generated by said sampling pattern generating means for a given category i of learning patterns wherein i is a positive integer equal to or larger than 1; orthogonalizing time·frequency filter means, comprising at least three types of filters, for obtaining an average pattern from each of the learning patterns of the given category i, smoothing the average pattern along a time reference axis thereof to generate dictionary data of a first axis, differentiating the average pattern along the time axis to generate dictionary data of a second axis, and differentiating the average pattern along a frequency axis thereof to generate dictionary data of a third axis; orthogonalized dictionary means for storing the dictionary data of the first, second, and third axes generated by said orthogonalizing time·frequency filter means; similarity computing means for computing similarities between each of the sampling patterns generated by said sampling pattern generating means and the dictionary data of the first, second, and third axes stored in said orthogonalized dictionary means; and sampling pattern recognizing means for recognizing each of the sampling patterns based on similarities with the dictionary data of the first, second and third axes as computed by said similarity computing means.

Claim 16

An apparatus according to claim 12, wherein said orthogonalizing time·frequency filter means obtains average pattern A.sub.(j,k) from a plurality of learning patterns a m (j,k) of the given category i using the following equation: ##EQU14## wherein j represents a number of acoustically analyzed feature vectors of the sampling data points of the learning pattern, k represents a number of subintervals the learning pattern is divided into by the predetermined number of sampling data points, and m represents a number of learning patterns of the given category i, j and k being positive integers; obtains dictionary data b 1 (j,k) of the first axis using the following equation: b.sub.1(j,k) =A.sub.(j-1,k-1) +A.sub.(j-1,k) +A.sub.(j-1,k+1) +A.sub.(j,k-1) +2*A.sub.(j,k) +A.sub.(j,k+1) +A.sub.(j+1,k-1) +A.sub.(j+1,k) +A.sub.(j+1,k+1) ; obtains dictionary data b 2 (j,k) of the second axis using the following equation: b.sub.2(j,k) =-A.sub.(j-1,k-1) +A.sub.(j-1,k+1) +{(-A.sub.(j,k-1) +A.sub.(j,k+1))}+{(-A.sub.(j+1,k-1) +A.sub.(j+1,k-1) }; obtains dictionary data b 3 (j,k) of the third axis using the following equation: b.sub.3(j,k) =-A.sub.(j-1,k-1) -A.sub.(j-1,k) -A.sub.(j-1,k+1) +A.sub.(j+1,k-1)} +A.sub.(j+1,k) +A.sub.(j+1,k+1) ; and normalizes the dictionary data of the first, second, and third axes, and registers the normalized dictionary data in said orthogonalized dictionary means.

Claim 17

An apparatus according to claim 13, wherein said orthogonalizing time·frequency filter means performs reorthogonalization processing of the dictionary data of the second axis based on the following equation: B.sub.2(j,k) =b.sub.2(j,k) -(b.sub.2(j,k) ·b.sub.1(j,k))b.sub.1(j,k) and registers the reorthogonalized dictionary data B 2 (j,k) in said orthogonalized dictionary means as new dictionary data b 2 (j,k) of the second axis.

Claim 18

A speech recognition method comprising the steps of: (a) converting unit speech into an electrical signal and acoustically analyzing the electrical signal; (b) detecting a speech interval of the acoustically analyzed electrical signal; (c) extracting a predetermined number of sampling points from a speech pattern of the speech interval detected so as to obtain one of a sampling pattern for recognition and a learning pattern for storage; (d) storing a predetermined number of learning patterns for a given category i of learning patterns wherein i is a positive integer equal to or larger than 1; (e) obtaining an average pattern from each of the learning patterns of the given category i, smoothing the average pattern along a time reference axis thereof to generate dictionary data of a first axis, differentiating the average pattern along the time axis thereof to generate dictionary data of a second axis, and differentiating the average pattern along a frequency axis thereof to generate dictionary data of a third axis; (f) storing the dictionary data of the first, second, and third axes; (g) computing similarities between each of the sampling patterns and the stored dictionary data of the first, second, and third axes; and (h) recognizing the sampling patterns based on the computed similarities with the stored dictionary data.

Claim 19

A speech recognition method comprising the steps of: (a) converting unit speech into an electrical signal and acoustically analyzing the electrical signal; (b) detecting a speech interval of the acoustically analyzed electrical signal; (c) extracting a predetermined number of sampling points from a speech pattern of the speech interval detected so as to obtain one of a sampling pattern for recognition and a learning pattern for storage; (d) storing a predetermined number of learning patterns for a given category i of learning patterns wherein i is a positive integer equal to or larger than 1; (e) obtaining an average pattern from each of the learning patterns of the given category i, smoothing the average pattern along a time reference axis thereof to generate dictionary data of a first axis, differentiating the average pattern along the time axis thereof to generate dictionary data of a second axis, and differentiating the average pattern along a frequency axis thereof to generate dictionary data of a third axis; (f) storing the dictionary data of the first, second, and third axes; (g) computing similarities between each of the sampling patterns and the stored dictionary data of the first, second, and third axes; and (h) recognizing the sampling patterns based on the computed similarities with the stored dictionary data.

Claim 20

A method according to claim 15, wherein the step (e) comprises obtaining average pattern A.sub.(j,k) for each of the learning patterns a m (j,k) of the given category i using the following equation: ##EQU15## wherein j represents a number of acoustically analyzed feature vectors of the sampling data points of the learning pattern, k represents a number of subintervals the learning pattern is divided into by the predetermined number of sampling data points, and m represents a number of learning patterns of the given category i, j and k being positive integers; obtaining dictionary data b 1 (j,k) of the first axis using the following equation: b.sub.1(j,k) =A.sub.(j-1,k-1) +A.sub.(j-1,k) +A.sub.(j-1,k+1) +A.sub.(j,k-1) +2* A.sub.(j,k) +A.sub.(j,k+1) +A.sub.(j+1,k-1) +A.sub.(j+1,k) +A.sub.(j+1,k+1) ; obtaining dictionary data b 2 (j,k) of the second axis using the following equation: b.sub.2(j,k) =-A.sub.(j-1,k-1) +A.sub.(j-1,k+1) +{-A.sub.(j,k-1) +A.sub.(j,k+1) }+{-A.sub.(j+1,k-1) +A.sub.(j+1,k-1) }; obtaining dictionary data b.sub.(j-1,k) of the third axis using the following equation: b.sub.3(j,k) =-A.sub.(j-1,k-1) +A.sub.(j-1,k) -A.sub.(j-1,k+1) +A.sub.(j+1,k-1) +A.sub.(j+1,k) +A.sub.(j+1,k-1) }; and, normalizing the dictionary data of the first, second, and third axes.

Claim 21

A method according to claim 16, wherein the step (e) further comprises performing reorthogonalization processing of the dictionary data of the second axis based on the following equation: B.sub.2(j,k) =b.sub.2(j,k) -(b.sub.2(j,k) ·b.sub.1(j,k))b.sub.1(j,k) ; and generating reorthogonalized dictionary data B 2 (j,k) as new dictionary data b 2 (j,k) of the second axis.

Claim 22

A speech recognition apparatus comprising: speech converting means for converting input speech into an electrical signal; acoustic analyzing means for receiving the electrical signal from said speech converting means and acoustically analyzing the electrical signal; speech interval detecting means for detecting a speech interval of the acoustically analyzed electrical signal; sampling pattern generating means for extracting a predetermined number of sampling points from a speech pattern of the speech interval detected by said speech interval detecting means so as to obtain one of a sampling pattern for recognition and a learning pattern for storage; sampling pattern storage means for storing a predetermined number of learning patterns generated by said sampling pattern generating means for a given category i of the learning patterns wherein i is a positive integer equal to or larger than 1; orthogonalizing time·frequency filter means, comprising at least three types of filters, for obtaining an average pattern from each of the learning patterns of the given category i, smoothing the average pattern along a time reference axis thereof to generate dictionary data of a first axis, differentiating the average pattern along the time axis to generate dictionary data of a second axis, and differentiating the average pattern along a frequency axis thereof to generate dictionary data of a third axis; orthogonalized dictionary means for storing the dictionary data of the first, second, and third axes generated by said orthogonalizing time·frequency filter means; orthogonalizing means for generating additional dictionary data orthogonal to dictionary data stored by the orthogonalized dictionary means from learning patterns stored in the sampling pattern storage means and for adding said additional dictionary data to said orthogonalized dictionary means; similarity computing means for computing similarities between each of the sampling patterns generated by said sampling pattern generating means and the dictionary data of the first, second, and third axes stored in said orthogonalized dictionary means; and sampling pattern recognizing means for recognizing each of the sampling patterns based on similarities with the dictionary data stored by the orthogonalized dictionary means computed by said similarity computing means.

Claim 23

A speech recognition apparatus comprising: speech converting means for converting input speech into an electrical signal; acoustic analyzing means for receiving the electrical signal from said speech converting means and acoustically analyzing the electrical signal; speech interval detecting means for detecting a speech interval of the acoustically analyzed electrical signal; sampling pattern generating means for extracting a predetermined number of sampling points from a speech pattern of the speech interval detected by said speech interval detecting means so as to obtain one of a sampling pattern for recognition and a learning pattern for storage; sampling pattern storage means for storing a predetermined number of learning patterns generated by said sampling pattern generating means for a given category i of the learning patterns wherein i is a positive integer equal to or larger than 1; orthogonalizing time·frequency filter means, comprising at least three types of filters, for obtaining an average pattern from each of the learning patterns of the given category i, smoothing the average pattern along a time reference axis thereof to generate dictionary data of a first axis, differentiating the average pattern along the time axis to generate dictionary data of a second axis, and differentiating the average pattern along a frequency axis thereof to generate dictionary data of a third axis; orthogonalized dictionary means for storing the dictionary data of the first, second, and third axes generated by said orthogonalizing time·frequency filter means; orthogonalizing means for generating additional dictionary data orthogonal to dictionary data stored by the orthogonalized dictionary means from learning patterns stored in the sampling pattern storage means and for adding said additional dictionary data to said orthogonalized dictionary means; similarity computing means for computing similarities between each of the sampling patterns generated by said sampling pattern generating means and the dictionary data of the first, second, and third axes stored in said orthogonalized dictionary means; and sampling pattern recognizing means for recognizing each of the sampling patterns based on similarities with the dictionary data stored by the orthogonalized dictionary means computed by said similarity computing means.

Claim 24

An apparatus according to claim 18, wherein said orthogonalizing time·frequency filter means obtains average pattern A.sub.(j,k) from learning patterns a m (j,k) of the given category i using the following equation: ##EQU16## wherein j represents a number of acoustically analyzed feature vectors of the sampling data points of the learning pattern, k represents a number of subintervals the learning pattern is divided into by the predetermined number of sampling data points, and m represents a number of learning patterns of the given category i, j and k being positive integers; obtains dictionary data b 1 (j,k) of the first axis using the following equation: b.sub.1(j,k) =A.sub.(j-1,k-1) +A.sub.(j-1,k) +A.sub.(j-1,k+1) +A.sub.(j,k-1) +2*A.sub.(j,k) +A.sub.(j,k+1) +A.sub.(j+1,k-1) +A.sub.(j+1,k) +A.sub.(j+1,k+1) ; obtains dictionary data b 2 (j,k) of the second axis using the following equation: b.sub.2(j,k) =-A.sub.(j-1,k-1) +A.sub.(j-1,k+1) +{-A.sub.(j,k-1) +A.sub.(j,k+1) }+{-A.sub.(j+1,k-1) +A.sub.(j+1,k+1) }; obtains the dictionary b 3 (j,k) of the third axis using the following equation: b.sub.3(j,k) =-A.sub.(j-1,k-1) -A.sub.(j-1,k) +A.sub.(j-1,k+1) +A.sub.(j+1,k-1) +A.sub.(j+1,k) +A.sub.(j+1,k+1) ; and normalizes the dictionary data of the first, second, and third axes, and registers normalized dictionary data in said orthogonalized dictionary means.

Claim 25

An apparatus according to claim 19, wherein said orthogonalizing time·frequency filter means performs reorthogonalization processing of the dictionary data of the second axis based on the following equation: B.sub.2(j,k) =b.sub.2(j,k) -(b.sub.2(j,k) ·b.sub.1(j,k))b.sub.1(j,k) and registers reorthogonalized dictionary data B 2 (j,k) in said orthogonalized dictionary means as new dictionary data b 2 (j,k) of the second axis.