CN101165679A - Pattern matching device and method - Google Patents

Pattern matching device and method Download PDF

Info

Publication number
CN101165679A
CN101165679A CNA2006101356589A CN200610135658A CN101165679A CN 101165679 A CN101165679 A CN 101165679A CN A2006101356589 A CNA2006101356589 A CN A2006101356589A CN 200610135658 A CN200610135658 A CN 200610135658A CN 101165679 A CN101165679 A CN 101165679A
Authority
CN
China
Prior art keywords
mentioned
mode
compression
standard
pattern
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CNA2006101356589A
Other languages
Chinese (zh)
Other versions
CN100552664C (en
Inventor
柿野友成
栾剑
郝杰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Toshiba TEC Corp
Original Assignee
Toshiba Corp
Toshiba TEC Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Toshiba Corp, Toshiba TEC Corp filed Critical Toshiba Corp
Priority to CNB2006101356589A priority Critical patent/CN100552664C/en
Priority to JP2007071480A priority patent/JP2008102482A/en
Publication of CN101165679A publication Critical patent/CN101165679A/en
Application granted granted Critical
Publication of CN100552664C publication Critical patent/CN100552664C/en
Priority to JP2010248442A priority patent/JP5188563B2/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Landscapes

  • Compression, Expansion, Code Conversion, And Decoders (AREA)

Abstract

The mode matching device analyzes all feature elements composing the standard mode and combines the neighboring feature elements to form one feature element in order to compress the standard mode; after generating the compressed standard mode B'x, the sequence of compression ratio of each feature element is used as the compression information Nx; making correlation storage for the B'x and Nx; after mode A is inputted into the mode matching processor (15), according to the Nx, the recovery unit (21) elongates the B'x, and recovers it into the standard mode; the distance calculating unit (23) calculates the inter-mode distance between the standard mode and the input mode A.

Description

Mode matching device and method
Technical field
The present invention relates to mode matching device and the method in whether the input pattern of judging sound, literal, figure etc. is some area of pattern recognition in the mode standard of imagining in advance, utilized.
Background technology
In existing area of pattern recognition, widespread use dynamic programming (Dynamicprogramming:DP) is as pattern matching mode.Area of pattern recognition with sound is an example, and the dynamic programming ratio juris is described.
Acoustic pattern A shows as following (1) formula.
A＝a 1、a 2、...、a i、...、a I ...(1)
In (1) formula, i{i=1,2 ..., the I} express time, a is meant the characteristic element of acoustic pattern A at time i.
Therefore, in the various words each, prepare in advance identical with acoustic pattern A, with the mode standard B of the time series performance of characteristic element.This standard picture B shows as following (2) formula.
B＝b 1、b 2、...、b j、...、b J ...(2)
In (2) formula, j{j=1,2 ..., the J} express time, b is meant the characteristic element of acoustic pattern B at time j.
Mode matching device compares the mode standard B of acoustic pattern A that is imported and various words respectively, obtains the distance between the pattern.Then, the mode standard of minor increment is exported as recognition result.
In the acoustic pattern of reality, since the change of the rate of articulation, the distortion of generation time axle.Dynamic programming is extremely effective as the method for correcting this time shaft distortion.
In dynamic programming, the recursion formula represented of following (3) formula of computing repeatedly.
g ( i , j ) = min g ( i , j - 1 ) + d ( i , j ) g ( i - 1 , j - 1 ) + 2 d ( i , j ) g ( i - 1 , j ) + d ( i , j ) . . . ( 3 )
In (3) formula, (i j) is the characteristic element a of acoustic pattern A to d iCharacteristic element b with mode standard B jKey element between distance.(i j) is Cumulative Distance between the key element of acoustic pattern A and mode standard B to g.
The above-mentioned recursion formula of mode matching device repeated calculation (3).Then, as shown in Figure 9, on A, B plane, obtain and make a iWith b jBest corresponding path (DP-pass).
(i j), has stipulated and below abutment points (i, relation j-1) the uppermost formula of above-mentioned recursion formula (3) at the arbitrfary point on A shown in Figure 10, the B plane.(i j), has stipulated and lower left abutment points (i-1, relation j-1) formula of the centre of this recursion formula (3) at point.(i j), has stipulated and left side abutment points (i-1, relation j) the nethermost formula of this recursion formula (3) at point.
In addition, also add to tilt restriction sometimes by omitting uppermost formula from this recursion formula (3).In this case, for point arbitrarily (i, j) regulation relation shown in Figure 11.
But, generally prepare a plurality of in pattern match necessary mode standard.Therefore, consider to seek the efficient activity in the zone of storage standards pattern by the compression standard pattern.Method as the compression standard pattern comprises following method.That is,, get its average with the characteristic element of adjacency respectively for each characteristic element that constitutes mode standard.Then, under the situation that has approximate key element, they are merged into a characteristic element.Come the compression standard pattern like this.
Yet, to compress with above-mentioned method under the situation of mode standard, temporal information useful in pattern-recognition disappears.That is, because the characteristic element of adjacency is merged into one, so the temporal information that each key element has respectively disappears.After temporal information disappears, generally be easy to take place matching error (matching error) from mode standard.Particularly, make owing to sneaking into of noise (noise) under the quantity situation how that is identified as the approximate feature mode of part, matching error significantly occurs.
Summary of the invention
The objective of the invention is to, even under the situation of having compressed mode standard, also reduce the incidence of matching error.
The mode matching device of embodiment of the present invention is merged into a characteristic element at each characteristic element that constitutes mode standard by the characteristic element that will be similar to the characteristic element of adjacency, comes the compression standard pattern.In addition, the sequence of the ratio of compression of each characteristic element of the compression standard pattern after the generation compression is as compressed information.Then, compression standard pattern after the compression and the compressed information that generates at this compression standard pattern are stored explicitly.In addition, according to the compressed information that is associated with storing with this compression standard pattern, the compression standard pattern elongation being stored is recovered to mode standard.Then, the mode interval of mode standard after utilizing dynamic programming calculate to restore and input pattern from.
Description of drawings
Fig. 1 is the block diagram of structure of the mode matching device of expression embodiment of the present invention.
Fig. 2 is the figure that is used for illustrating the mode standard compression processing of carrying out at this mode matching device.
Fig. 3 is the mode chart that is illustrated in the data structure of storing in the storage unit of this mode matching device.
Fig. 4 is the block diagram of the major part structure of the pattern match processing unit in expression the 1st embodiment.
Fig. 5 is the block diagram of the major part structure in the pattern match processing unit of representing in the 2nd embodiment.
Fig. 6 is the figure that is used to illustrate the 1st recursion formula that the pattern match processing unit by the 2nd embodiment calculates.
Fig. 7 is the figure that is used to illustrate the 2nd recursion formula that the pattern match processing unit by the 2nd embodiment calculates.
Fig. 8 is the figure that is used to illustrate the 3rd recursion formula that the pattern match processing unit by the 2nd embodiment calculates.
Fig. 9 is the figure that is used to illustrate existing dynamic programming.
Figure 10 is the figure that is used to illustrate the recursion formula that calculates with existing dynamic programming.
Figure 11 is the figure that is used to illustrate other example of the recursion formula that calculates with existing dynamic programming.
Embodiment
Below, use description of drawings the present invention to be applicable to the embodiment in acoustic pattern identification field.At first, use Fig. 1～Fig. 4 that the 1st embodiment is described.
Fig. 1 is the block diagram of major part structure of the mode matching device 10 of expression present embodiment.Mode matching device 10 possesses phonetic analysis unit 11, mode compression unit 12, compressed information generation unit 13, storage unit 14 and pattern match processing unit 15.
To phonetic analysis unit 11 input audio signal M.Voice signal M carries out the electric signal that obtains after the conversion from the sound by microphone (microphone) input.The voice signal M that is imported is analyzed in phonetic analysis unit 11.Then, extract the characteristic element a of sound import, generate the acoustic pattern A that represents with above-mentioned (1) formula according to time series.Acoustic pattern A outputs to pattern match processing unit 15.
In mode compression unit 12, import a plurality of mode standards (pattern) B1, B2, B3 ....At in the various words each prepare in advance each mode standard B1, B2, B3 ....These mode standards B1, B2, B3 ... be not compressed.Mode compression unit 12 along time orientation compress respectively each mode standard B1, B2 of being imported, B3 ....
About its compression method, use Fig. 2 to describe.In the figure, b 1, b 2, b 3..., b 9Represent each characteristic element of mode standard Bx arbitrarily.Mode standard Bx is each characteristic element b 1, b 2, b 3..., b 9Become by its series arrangement and to obtain after the time series.
Mode compression unit 12 is at each the characteristic element b that constitutes mode standard Bx 1, b 2, b 3..., b 9, obtain distance respectively with the characteristic element of adjacency.Then, under the situation of distance, a plurality of characteristic elements of these adjacency are replaced as an average characteristics key element b ' smaller or equal to pre-set threshold k(k=1,2,3 ...).By this processing, partly a plurality of characteristic element boil down to average characteristics key elements.Should handle by whole implementation, generate compression standard Mode B ' x mode standard Bx.
Under the situation of the example of Fig. 2, the characteristic element b of adjacency 2With b 3Distance smaller or equal to threshold value, these characteristic elements b 2, b 3Be replaced as average characteristics key element b ' 2B ' 2=(b 2+ b 3)/2}.In addition, the characteristic element b of adjacency 4, b 5And b 6Distance smaller or equal to threshold value, these characteristic elements b 4, b 5And b 6Be replaced as average characteristics key element b ' 3B ' 3=(b 4+ b 5+ b 6)/3}.In addition, the characteristic element b of adjacency 7With b 8Distance smaller or equal to threshold value, these characteristic elements b 7, b 8Be replaced as average characteristics key element b ' 4B ' 4=(b 7+ b 8)/2}.Like this, mode standard Bx (=b 1, b 2, b 3..., b 9) be compressed into compression standard Mode B ' x (b ' 1, b ' 2, b ' 3, b ' 4, b ' 5).
Compressed information generation unit 13 is each the characteristic element b ' by the compression standard Mode B ' x after 12 compressions of above-mentioned mode compression unit 1, b ' 2, b ' 3, b ' 4, b ' 5Ratio of compression n1, n2, n3, n4, the sequence of n5 be generated as compressed information Nx.In the present embodiment, as shown in Figure 2, each characteristic element b ' of expression compression standard Mode B ' x 1, b ' 2, b ' 3, b ' 4, b ' 5The value defined of representing the several characteristic key element of original mode standard Bx respectively is ratio of compression n 1, n 2, n 3, n 4, n 5
That is the characteristic element b ' of compression standard Mode B ' x, 1Only represent the characteristic element b of mode standard 1, so ratio of compression n1 is " 1 ".Equally, characteristic element b ' 2Represent characteristic element b 2And b 3, so ratio of compression n2 is " 2 ".Characteristic element b ' 3Represent characteristic element b 4, b 5And b 6, so ratio of compression n3 is " 3 ".Characteristic element b ' 4Represent characteristic element b 7And b 8, so ratio of compression n4 is " 2 ".Characteristic element b ' 5Only represent characteristic element b 9, so ratio of compression n5 is " 1 ".
Like this, the compressed information Nx at compression standard Mode B ' x is " 1,2,3,1,2 ".
Storage unit 14 makes by compression standard Mode B ' x and compressed information generation unit 13 after 12 compressions of mode compression unit and stores explicitly at the compressed information Nx that this compression standard pattern generates as shown in Figure 3.
Pattern match processing unit 15 possesses restoration unit 21, smoothing processing unit 22 and metrics calculation unit 23 as shown in Figure 4.
Restoration unit 21 bases and the compressed information Nx that this compression standard Mode B ' x stores explicitly the compression standard Mode B of storage in the storage unit 14 ' x elongation, are recovered to mode standard B with it X1For example, under the situation of the example of Fig. 2, with compression standard Mode B ' x (b ' 1, b ' 2, b ' 3, b ' 4, b ' 5) the compressed information Nx (n1, n2, n3, n4, n5) that is associated is " 1,2,3,1,2 ".
Thereby, characteristic element b ' 1Remain unchanged.Characteristic element b ' 2Become 2 characteristic element b ' of adjacency 2, b ' 2Characteristic element b ' 3Become 3 characteristic element b ' of adjacency 3, b ' 3, b ' 3Characteristic element b ' 4Become 2 characteristic element b ' of adjacency 4, b ' 4Characteristic element b ' 5Remain unchanged.Like this, compression standard Mode B ' x is recovered to mode standard B X1(b ' 1, b ' 2, b ' 2, b ' 3, b ' 3, b ' 3, b ' 4, b ' 4, b ' 5).
Smoothing processing unit 22 is at the mode standard B after being restored by restoration unit 21 X1, carry out smoothing by low-pass filter (low pass filter) and handle.Mode standard B after the recovery X1The middle noise that causes by compression that produces.Handle by implementing smoothing, can remove this noise.
Metrics calculation unit 23 utilizes well-known dynamic programming to calculate via the acoustic pattern A of phonetic analysis unit 11 inputs and the mode standard B that has been undertaken by smoothing processing unit 22 after smoothing is handled X1Mode interval from Gx.For example, the computing repeatedly by the recursion formula shown in above-mentioned (3) formula comes between computation schema apart from Gx.
Pattern match processing unit 15 is carried out the processing of above-mentioned restoration unit 21, smoothing processing unit 22 and metrics calculation unit 23 repeatedly at all compression standard Mode B ' x.Then, obtain the mode interval that calculates and be minimum mode standard B from Gx X1, as the recognition result G output of acoustic pattern A.
Like this, in the mode matching device 10 of present embodiment, passing through the compression standard Mode B xWhen generating compression standard Mode B ' x, the ratio of compression sequence of each characteristic element that generates this compression standard Mode B ' x is as compressed information Nx.Then, compression standard Mode B ' x and the compressed information Nx that generates at this compression standard Mode B ' x are stored in the storage unit 14 explicitly.
Behind input audio signal M, extract the characteristic element a of sound import by phonetic analysis unit 11 according to time series, generate acoustic pattern A.Acoustic pattern A outputs to pattern match processing unit 15.Pattern match processing unit 15 is carried out following pattern match and is handled at all compression standard Mode B ' x of storage in the storage unit 14.
At first, the compressed information Nx that reads compression standard Mode B ' x arbitrarily and be associated from storage unit 14 with it.Then, compression standard Mode B ' x elongation, be recovered to mode standard B according to compressed information Nx X1Then, to the mode standard B after restoring X1Carrying out smoothing handles.Then, acoustic pattern A that utilizes dynamic programming to calculate to be imported and the mode standard B that has implemented after smoothing is handled X1Mode interval from Gx.
Like this, pattern match processing unit 15 is at each compression standard Mode B ' x, and the mode interval of calculating and acoustic pattern A is from Gx.Then, be minimum mode standard B with mode interval from Gx X1Recognition result G output as acoustic pattern A.
Like this, be not with the mode standard B ' x after the compression be directly used in calculate with the mode interval of acoustic pattern A from, but utilizing compressed information Nx to be recovered to mode standard B X1Use later on.Thereby,, therefore reduced the incidence of matching error owing to added the temporal information of losing owing to the compression of mode standard Bx.
Then, the 2nd embodiment is described.In the 2nd embodiment, only pattern match processing unit 15 is different with the 1st embodiment.The structural drawing of Fig. 5 illustrates the major part structure of the pattern match processing unit 15 in the 2nd embodiment.
Pattern match processing unit 15 possesses recursion formula setup unit 31 and metrics calculation unit 32.Metrics calculation unit 32 is by the computing repeatedly of the recursion formula of setting in recursion formula setup unit 31, and the mode interval that calculates acoustic pattern A and compression standard Mode B ' x is from Gx.At this moment, use the compressed information Nx that stores explicitly with compression standard Mode B ' x, carry out the calculating of recursion formula.
In recursion formula setup unit 31, set the recursion formula shown in following (4) formula.
g ( i , j ) = min g ( i - 1 , j ) + d ( i , j ) g ( i - n j - 1 , j - 1 ) + Σ k = 1 n j - 1 - 1 d ( i - k , j - 1 ) + 2 d ( i , j ) ( · · · n j - 1 > 1 ) g ( i - 1 , j - 1 ) + 2 d ( i , j ) ( · · · n j - 1 = 1 ) . . . ( 4 )
In (4) formula, (i j) is the characteristic element a of acoustic pattern A to d iCharacteristic element b ' with compression standard Mode B ' x jKey element between distance.(i j) is Cumulative Distance between acoustic pattern A and compression standard Mode B ' x key element to g.n J-1It is the key element (ratio of compression) of compressed information Nx.
(i j), has stipulated and left side abutment points (i-1, relation j) the uppermost formula of above-mentioned recursion formula (4) at the arbitrfary point on A shown in Figure 6, the B plane.
The nethermost formula of this recursion formula (4) is compressed information key element n J-1Situation for " 1 ".In this case, (i j), has stipulated and oblique below, a left side abutment points (i-1, relation j-1) at point.
The formula of the centre of this recursion formula (4) is compressed information key element n J-1Than " 1 " big situation.In this case, at point (i, j), stipulated with from a left side tiltedly the below abutment points (i-1, j-1) beginning, according to characteristic element b ' J-1Corresponding compressed information key element n J-1Extended the point (i-n after the restriction path J-1, relation j-1).
By carrying out the computing of this recursion formula (4) repeatedly, the characteristic element b ' of compression standard Mode B ' x J-1Quilt has been assigned the n with input pattern A J-1Individual key element a iCorrespondence.Thereby, owing to added because mode standard B X1Compression and therefore the temporal information of losing has reduced the incidence of matching error.
The recursion formula that can obtain effect same is not limited to above-mentioned (4) formula.For example, also can in recursion formula setup unit 31, set as shown in Figure 7 at characteristic element b ' jLimit the recursion formula of following (5) formula of the elongation of path.
g ( i , j ) = min g ( i - n j , j ) + Σ k = 1 n j - 1 d ( i - k , j ) + d ( i , j ) ( · · · n j > 1 ) g ( i - 1 , j ) + d ( i , j ) ( · · · n j = 1 ) g ( i - 1 , j - 1 ) + 2 d ( i , j ) . . . ( 5 )
Perhaps, also can in recursion formula setup unit 31, set as shown in Figure 8 at characteristic element b ' jAnd b ' J-1Two sides limit the recursion formula of following (6) formula of the elongation of path.
g ( i , j ) = min g ( i - n j , j ) + Σ k = 1 n j - 1 d ( i - k , j ) + d ( i , j ) ( · · · n j > 1 ) g ( i - 1 , j ) + d ( i , j ) ( · · · n j = 1 ) g ( i - n j - 1 , j - 1 ) + Σ k = 1 n j - 1 - 1 d ( i - k , j - 1 ) + 2 d ( i , j ) ( · · · n j - 1 > 1 ) g ( i - 1 , j - 1 ) + 2 d ( i , j ) ( · · · n j - 1 = 1 ) . . . ( 6 )
In the respective embodiments described above, at the compression standard Mode B xThe time, get the average of approximate a plurality of characteristic elements, but a conduct representative in the characteristic element that also can select to be similar to.In addition, can also use cluster (clustering) method based on code book (codebook).
In addition, in the above-described embodiment, directly use the key element (ratio of compression) of compressed information Nx, but also can use predetermined multiplying power, for example 0.8 times value is come distance between computation schema.
In addition, show the situation of the area of pattern recognition that is applicable to sound at above-mentioned embodiment, but the present invention can be applicable in the area of pattern recognition of literal, figure etc. too.In addition, certainly carrying out various distortion without departing from the spirit and scope of the present invention implements.

Claims (9)

1. mode matching device, the mode interval of obtaining respectively the mode standard represented with the time series of characteristic element and input pattern from, and with this mode interval from as recognition result output, it is characterized in that possessing:
Mode compression unit (12) at each characteristic element that constitutes above-mentioned mode standard, is merged into a characteristic element by the characteristic element that will be similar to the characteristic element of adjacency, compresses above-mentioned mode standard;
Compressed information generation unit (13) generates sequence by the ratio of compression of each characteristic element of the compression standard pattern after the above-mentioned mode compression cell compression as compressed information;
Storage unit (14) will be stored at the above-mentioned compressed information that this compression standard pattern generates explicitly by the above-mentioned compression standard pattern after the above-mentioned mode compression cell compression and by above-mentioned compressed information generation unit;
Pattern restoration unit (21), according to the above-mentioned compressed information of storing explicitly with above-mentioned compression standard pattern, this compression standard pattern elongation by the said memory cells storage is recovered to above-mentioned mode standard; And
Metrics calculation unit (23), the above-mentioned mode standard after utilizing dynamic programming to calculate to restore by above-mentioned pattern restoration unit and with the mode interval of above-mentioned input pattern from.
2. mode matching device according to claim 1 is characterized in that also possessing:
Smoothing processing unit (22) makes by the mode standard smoothing after the above-mentioned pattern restoration unit recovery,
Above-mentioned metrics calculation unit utilize mode interval that dynamic programming calculates by above-mentioned smoothing processing unit mode standard after level and smooth and above-mentioned input pattern from.
3. mode matching device, the mode interval of obtaining respectively the mode standard represented with the time series of characteristic element and input pattern from, and with this mode interval from as recognition result output, it is characterized in that possessing:
Mode compression unit (12) at each characteristic element that constitutes above-mentioned mode standard, is merged into a characteristic element by the characteristic element that will be similar to the characteristic element of adjacency, compresses above-mentioned mode standard;
Compressed information generation unit (13) generates sequence by the ratio of compression of each characteristic element of the compression standard pattern after the above-mentioned mode compression cell compression as compressed information;
Storage unit (14) will be stored at the above-mentioned compressed information that this compression standard pattern generates explicitly by the above-mentioned compression standard pattern after the above-mentioned mode compression cell compression and by above-mentioned compressed information generation unit;
Metrics calculation unit (32), utilize the recursion formula of dynamic programming, calculating by the mode interval of the above-mentioned compression standard pattern of said memory cells storage and above-mentioned input pattern from, wherein, above-mentioned recursion formula has the above-mentioned compressed information that generated at this compression standard pattern by above-mentioned compressed information generation unit as variable.
4. mode matching device according to claim 3 is characterized in that,
If the characteristic element a of above-mentioned input pattern iCharacteristic element b with above-mentioned compression standard pattern jKey element between distance be d (i, j), between the key element of above-mentioned input pattern and above-mentioned compression standard pattern Cumulative Distance be g (i, j), be n as the compressed information of the sequence of above-mentioned ratio of compression n i, then the recursion formula that uses in above-mentioned metrics calculation unit is expressed as follows:
g ( i , j ) = min g ( i - 1 , j ) + d ( i , j ) g ( i - n j - 1 , j - 1 ) + Σ k = 1 n j - 1 - 1 d ( i - k , j - 1 ) + 2 d ( i , j ) ( · · · n j - 1 > 1 ) g ( i - 1 , j - 1 ) + 2 d ( i , j ) ( · · · n j - 1 = 1 ) .
5. mode matching device according to claim 3 is characterized in that,
If the characteristic element a of above-mentioned input pattern iCharacteristic element b with above-mentioned compression standard pattern jKey element between distance be d (i, j), between the key element of above-mentioned input pattern and above-mentioned compression standard pattern Cumulative Distance be g (i, j), be n as the compressed information of the sequence of above-mentioned ratio of compression n j, then the recursion formula that uses in above-mentioned metrics calculation unit is expressed as follows:
g ( i , j ) = min g ( i - n j , j ) + Σ k = 1 n j - 1 d ( i - k , j ) + d ( i , j ) ( · · · n j > 1 ) g ( i - 1 , j ) + d ( i , j ) ( · · · n j = 1 ) g ( i - 1 , j - 1 ) + 2 d ( i , j ) .
6. mode matching device according to claim 3 is characterized in that,
If the characteristic element a of above-mentioned input pattern iCharacteristic element b with above-mentioned compression standard pattern jKey element between distance be d (i, j), between the key element of above-mentioned input pattern and above-mentioned compression standard pattern Cumulative Distance be g (i, j), be n as the compressed information of the sequence of above-mentioned ratio of compression n j, then the recursion formula that uses in above-mentioned metrics calculation unit is expressed as follows:
g ( i , j ) = min g ( i - n j , j ) + Σ k = 1 n j - 1 d ( i - k , j ) + d ( i , j ) ( . . . n j > 1 ) g ( i - 1 , j ) + d ( i , j ) ( . . . n j = 1 ) g ( i - n j - 1 , j - 1 ) + Σ k = 1 n j - i - 1 d ( i - k , j - 1 ) + 2 d ( i , j ) ( . . . n j - 1 > 1 ) g ( i - j , j - 1 ) + 2 d ( i , j ) ( . . . n j - 1 = 1 ) .
7. the method for mode matching in the mode matching device, the mode interval that this mode matching device is obtained respectively the mode standard represented with the time series of characteristic element and input pattern from, and with this mode interval from exporting as recognition result, it is characterized in that above-mentioned method for mode matching comprises:
Compression step at each inscape that constitutes above-mentioned mode standard, is merged into a characteristic element by the characteristic element that will be similar to the characteristic element of adjacency, compresses above-mentioned mode standard;
Generate step, the sequence that is created on the ratio of compression of each characteristic element of the compression standard pattern after the compression in the above-mentioned compression step is as compressed information;
Reconstitution steps, according to the above-mentioned compressed information that is generated at above-mentioned compression standard pattern by above-mentioned generation step, this compression standard pattern elongation after the compression in above-mentioned compression step is recovered to above-mentioned mode standard; And
Calculation procedure, utilize dynamic programming calculate by above-mentioned mode standard after the above-mentioned reconstitution steps recovery and above-mentioned input pattern mode interval from.
8. method for mode matching according to claim 7 is characterized in that, also comprises:
The smoothing step makes by the mode standard smoothing after the above-mentioned reconstitution steps recovery,
The aforementioned calculation step utilize mode interval that dynamic programming calculates by above-mentioned smoothing step mode standard after level and smooth and above-mentioned input pattern from.
9. the method for mode matching in the mode matching device, the mode interval that this mode matching device is obtained respectively the mode standard represented with the time series of characteristic element and input pattern from, and with this mode interval from exporting as recognition result, it is characterized in that above-mentioned method for mode matching comprises:
Compression step at each inscape that constitutes above-mentioned mode standard, is merged into a characteristic element by the characteristic element that will be similar to the characteristic element of adjacency, compresses above-mentioned mode standard;
Generate step, the sequence that is created on the ratio of compression of each characteristic element of the compression standard pattern after the compression in the above-mentioned compression step is as compressed information; And
Calculation procedure, utilize the recursion formula of dynamic programming, the above-mentioned compression standard pattern after calculating is compressed in above-mentioned compression step and the mode interval of above-mentioned input pattern from, wherein, above-mentioned recursion formula has the above-mentioned compressed information that generated at above-mentioned compression standard pattern by above-mentioned generation step as variable.
CNB2006101356589A 2006-10-20 2006-10-20 mode matching device and method Expired - Fee Related CN100552664C (en)

Priority Applications (3)

Application Number Priority Date Filing Date Title
CNB2006101356589A CN100552664C (en) 2006-10-20 2006-10-20 mode matching device and method
JP2007071480A JP2008102482A (en) 2006-10-20 2007-03-19 Pattern matching device and method
JP2010248442A JP5188563B2 (en) 2006-10-20 2010-11-05 Pattern matching apparatus and method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CNB2006101356589A CN100552664C (en) 2006-10-20 2006-10-20 mode matching device and method

Publications (2)

Publication Number Publication Date
CN101165679A true CN101165679A (en) 2008-04-23
CN100552664C CN100552664C (en) 2009-10-21

Family

ID=39334378

Family Applications (1)

Application Number Title Priority Date Filing Date
CNB2006101356589A Expired - Fee Related CN100552664C (en) 2006-10-20 2006-10-20 mode matching device and method

Country Status (2)

Country Link
JP (2) JP2008102482A (en)
CN (1) CN100552664C (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110880315A (en) * 2019-10-17 2020-03-13 深圳市声希科技有限公司 Personalized voice and video generation system based on phoneme posterior probability
CN111918204A (en) * 2014-10-03 2020-11-10 交互数字专利控股公司 Method for limiting direct discovery

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR102240455B1 (en) * 2019-06-11 2021-04-14 네이버 주식회사 Electronic apparatus for dinamic note matching and operating method of the same

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPS5614381A (en) * 1979-07-13 1981-02-12 Nec Corp Pattern recognizing system
JPS59195699A (en) * 1983-04-21 1984-11-06 三菱電機株式会社 Word voice recognition equipment
JPS60146300A (en) * 1984-01-10 1985-08-01 株式会社リコー Pattern generation system
JPS60158498A (en) * 1984-01-27 1985-08-19 株式会社リコー Pattern collation system
JPS60201398A (en) * 1984-03-26 1985-10-11 日本電信電話株式会社 Continuous word recognition
JPS6170595A (en) * 1984-09-14 1986-04-11 株式会社リコー Voice recognition system
JPS61129700A (en) * 1984-11-29 1986-06-17 株式会社リコー Dictionary preparation for voice recognition equipment
JPH0635494A (en) * 1992-07-16 1994-02-10 Nec Corp Speech recognizing device
KR100324988B1 (en) * 1994-06-13 2002-08-27 마츠시타 덴끼 산교 가부시키가이샤 Signal analyzer
US5960395A (en) * 1996-02-09 1999-09-28 Canon Kabushiki Kaisha Pattern matching method, apparatus and computer readable memory medium for speech recognition using dynamic programming
US7437398B2 (en) * 2004-07-30 2008-10-14 Intel Corporation Pattern matching architecture

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111918204A (en) * 2014-10-03 2020-11-10 交互数字专利控股公司 Method for limiting direct discovery
US11350263B2 (en) 2014-10-03 2022-05-31 Interdigital Patent Holdings, Inc. Methods for restricted direct discovery
US11812499B2 (en) 2014-10-03 2023-11-07 Interdigital Patent Holdings, Inc. Methods for restricted direct discovery
CN110880315A (en) * 2019-10-17 2020-03-13 深圳市声希科技有限公司 Personalized voice and video generation system based on phoneme posterior probability

Also Published As

Publication number Publication date
JP2008102482A (en) 2008-05-01
CN100552664C (en) 2009-10-21
JP2011022621A (en) 2011-02-03
JP5188563B2 (en) 2013-04-24

Similar Documents

Publication Publication Date Title
AU2017404565B2 (en) Electronic device, method and system of identity verification and computer readable storage medium
Taieb et al. Boosting multi-step autoregressive forecasts
CN109658943B (en) Audio noise detection method and device, storage medium and mobile terminal
JP2001356791A (en) Noise adaptation method for automatic speech recognition using transformation matrix
CN103258533B (en) Novel model domain compensation method in remote voice recognition
CN100552664C (en) mode matching device and method
CN104995673B (en) Hiding frames error
CN115661911A (en) Face feature extraction method, device and storage medium
Yoon et al. Multiple points input for convolutional neural networks in replay attack detection
US20050185834A1 (en) Method and apparatus for scene learning and three-dimensional tracking using stereo video cameras
JP2017059193A (en) Time series image compensation device, time series image generation method, and program for time series image compensation device
CN112837670A (en) Voice synthesis method and device and electronic equipment
Liu et al. Depth-first neural architecture with attentive feature fusion for efficient speaker verification
CN106531156A (en) Speech signal enhancement technology method based on indoor multi-mobile source real-time processing
CN115169366A (en) Session recommendation method based on sampling convolution and interaction strategy
CN115602158A (en) Voice recognition acoustic model construction method and system based on telephone channel
CN109614587B (en) Intelligent human relationship analysis modeling method, terminal device and storage medium
CN112613614A (en) Method, system, equipment and medium for compressing and accelerating bert model
Li et al. Speech enhancement based on robust NMF solved by alternating direction method of multipliers
Ting et al. Performance analysis of single and combined bit-planes feature extraction for recognition in face expression database
CN115170916B (en) Image reconstruction method and system based on multi-scale feature fusion
Dong Influence modeling of complex stochastic processes
CN110222834B (en) Divergent artificial intelligence memory model system based on noise shielding
CN116451006B (en) PMU data recovery method and system based on enhanced time sequence mode attention
CN116469392A (en) Method for confirming speaker capturing local and global features based on parallel branch structure

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C14 Grant of patent or utility model
GR01 Patent grant
C41 Transfer of patent application or patent right or utility model
TR01 Transfer of patent right

Effective date of registration: 20090918

Address after: Tokyo, Japan, Japan

Patentee after: Toshiba-tec K. K.

Address before: Tokyo, Japan

Co-patentee before: Toshiba Corp

Patentee before: Toshiba tiger Co., Ltd.

C17 Cessation of patent right
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20091021

Termination date: 20121020