JPS5961898A - Recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Technical Field] The present invention relates to an improvement in a recognition device, and more specifically, the present invention relates to an improvement in a recognition device, and more specifically, for example, a recognition device that recognizes a section of information such as a section of speech such as a phrase, etc. This invention relates to the improvement of a recognition device that recognizes subdivided unit elements.

<Prior art> The sound of a segment of a phrase, etc., can be expressed as a phoneme, kana, syllable, etc.
! In the case of recognition in 7 subdivided units, conventionally, generally speaking, one section of input speech information to be recognized is processed, for example, by acoustic processing to obtain a cover pattern containing feature vectors for each unit such as phoneme or syllable. Perform 77 checking between the input cover turn and the standard no (turn) stored in advance, and output the input information as a candidate unit sequence in descending order of similarity, and combine this output candidate unit sequence and bunsetsu. The information of a segment 9, such as a phrase, for the input information is recognized by checking the contents of a dictionary such as .

However, according to such conventional methods, all phonemes,
The degree of similarity is calculated by matching standard patterns such as syllables with the input cover patterns, and the ones with the highest degree of similarity are output as candidate syllables, etc.

Therefore, for example, when recognizing single syllables including syllables, it is necessary to match over 100 types of monosyllable standard patterns and input kabata patterns for each syllable, which takes a lot of processing time. It had become.

In addition, it is necessary to perform dictionary matching processing on all candidate unit sequences output from those with high similarity after that.
The processing time becomes long, the accuracy of recognizing correct phrases, etc. does not improve, and as a result, the amount of processing required for overall recognition becomes enormous.

Purpose of the present invention: It is an object of the present invention to provide a recognition device that eliminates the above-mentioned drawbacks of the conventional art, and improves the accuracy of recognizing information to be recognized at intervals such as correct phrases, and as a result, improves overall recognition efficiency. The present invention provides a recognition device that can reduce the amount of processing required for recognition.

<Example> Hereinafter, an example in which the recognition device of the present invention recognizes one segment of audio input such as a phrase using more subdivided unit elements such as syllables will be described as an example.

According to an embodiment of the present invention, in a recognition device that recognizes information to be recognized, such as a segment of speech such as a phrase, using N unit elements that are further divided into phonemes, kana, syllables, etc., the recognition target A transition matrix creation means for creating a transition matrix that describes a transition relationship that is a connection relationship between (N+1) characters (unit elements) for a character (unit element) string of a bunsetsu or sentence, and this transition matrix creation means Based on the transition matrix created by
Exclude groups of syllables (unit elements) that do not transition from the recognition target, and/or refer to the transition matrix for each candidate sequence when creating candidate sequences, and exclude candidate sequences that include combinations of syllables (unit elements) that do not transition. The system is configured to include processing means for performing recognition processing such as , etc., in order to reduce the amount of processing in the next high-level dictionary matching.

First, prior to a detailed explanation of the present invention, a transition matrix showing a transition relationship, which is a connection relationship between unit absorptive elements used in the recognition device of the present invention, will be explained.

In general, Japanese sentences can be expressed as syllable strings corresponding to kana character strings when all Japanese sentences are expressed using kana characters.

For example, the phrase “Earth’s” is
It consists of four unit elements called ``1'', which are called monosyllables.9
Are standing. Connective relationship between two syllables (“chi” gara ekyu “, “kyu kara ゝu〃, 1u I gara ゝno”)
If you look at sentences in all of Japanese or in one topic in a specific field, there are syllable pairs that do not connect (transition; hereinafter, the term transition is used). For example, before the syllable of the word ``b'', nothing other than ゛〃 or ``ge'' comes.Also, ゝnya'' does not come at the beginning of the word, and ``he〃 (also pronounced as ``he'') does not come at the end of the word.

The first-order transition relationship of the syllables constituting such a bunsetsu is written as the following equation (1) K, and a transition matrix M(X, Y) as shown in FIG. 1 is created.

In FIG. 1, the transition matrix M(X, Y) describes the transition from character X to the next character Y in a character string, which is a unit element string, and when there are N unit elements (syllables), ( N+I)
It is a matrix of X(N+I), and is stored in a ROM or the like in terms of hardware. In addition, the YO column shows whether each unit element (1 to N) comes in a clause, and the X0 row shows each unit element (1 to N).
) is written in the category.

For example, Figure 2 shows an example in which the transition of the character string "red" is written in the transition matrix.The elements of the transition matrix can be either 0 (transition not possible) or l (transition possible). revealed, 1
It is stored in bits. In FIG. 2, all matrix elements other than the notation "I" are shown as "0", and their display is omitted.

Next, the creation of the transition matrix will be explained in a little more detail.

First, when creating a transition matrix, the transition matrix memory is initially set to 0'' (M(x, y)==0).

Next, the character string A-= (at, a2.a3,..., a
l) However, when l is the number of characters in a string, the character transition relationship of character string A is expressed as a transition matrix M(X,
Write in Y). Similarly, the transition relationships are written for all character strings to be recognized, and the creation of the transition matrix (first order) is completed.

The concrete transition matrix (first order) M created in this way
An example of (X, Y) is shown in FIG. As is clear from FIG. 3, for example, (X, Y)-(Eh).

Since the bit position of
Since there is a transition to ", and the focus position of (X, Y) - (e, ke) is 0, this indicates that there is no transition from ``e'' to ゝke.

The above is a first-order transition, but there are also second-order transitions, and more generally M
The M-order transition matrix extended to the next can be similarly created according to the following equation (2).

M-order transition matrix: M(XI, X2.X3.-, XM,
Y).

(N+1) dimension M(ai-M, ai-(MD,-,ai)=I,
(i−+ ~I+1)・(2) The embodiment of the present invention uses connections between (N+1) unit elements for a predetermined unit element sequence to be recognized so that this transition matrix can be arbitrarily created. The recognition device is provided with a transition matrix creation means for creating a transition matrix that describes a transition relationship.

Next, an embodiment of the present invention will be described with reference to the drawings. FIG. 4 is a block diagram showing the configuration of an apparatus according to an embodiment of the present invention.

In FIG. 4, 1 is a floppy disk device,
A storage medium storing character strings such as phrases or sentences to be recognized is attached. 2 is the character code input terminal 1,
The character code of a character string is input from an external device. Also 3
is a keyboard device, 4 is a changeover switch means, 5 is a central processing unit (CPtJ), 6 is a character buffer, 7 is a character counter, 8 is a transition matrix memory, 9 is a recognition processing unit, and IO is the input of voice information to be recognized. input terminal 11 is the above c
This is a function key for inputting an instruction signal for creating a transition matrix to pu.

In the configuration described by Nagisa, when writing desired transition matrix information to the transition matrix memory 8, first press the function key 1.
1 to instruct the CPU to create a 1 transition matrix, and then operate the changeover switch means 4 to select the keyboard 3,
The user selects the floppy disk drive 1 or other input means and inputs the character string to be recognized as a unit.

Under the control of the CPt 15, the character string input from the input means writes transition matrix information into the transition matrix memory 8 according to the processing flow for creating a transition matrix shown in FIG.

That is, the CPU 5 first executes the transition matrix initial value setting operation (step n1), and sets all the stored contents of the transition matrix memory 8 to the initial value "0".

Next, proceeding to a specific transition matrix creation operation (n2), the character string input from the input means is encoded and temporarily stored in the character buffer 6, and the number of characters is stored in the character counter 7.
(n3).

Next, character transition relationships are written into the transition matrix memory 8 based on the encoded character string information stored in the character buffer 6 (n4). This operation is performed according to equation (1) above. Specifically, for example, the data stored in the character buffer Otsu 6 is sequentially shifted into a transition determination means consisting of one-digit character buffers 6X, 6Y and a -character delay device 6D shown in FIG. and 6Y
The (x, y) address of the transition matrix memory 8 is specified in accordance with the contents, and all ``1'' is written to that address position.Therefore, the first character code is input to the buffer 6Y by the first shift operation. , X=O in memory 8.

The address position of Y=al is specified, and "1" is written to that address position.The next shift operation inputs the first character code to buffer 6X, and the second character code to buffer 6Y. is input into memory 8, X=
The address position of a 1 , Y=a2 is specified, l1'' is written to that address position, and the transition relationship from al to a2 is written.
- Completes the writing of the transition relationship for the character string.

Below, similar operations are performed for all character strings to be recognized to complete the creation of the transition matrix (n5, n6
).

In addition, when adding a new recognition target word, except for the initial setting operation of step nl in FIG.
3 to n5 are executed and the transition relationship is written in the transition matrix.

The above is a first-order transition, but for a second-order transition and even an M-order transition matrix, the transitions shown in FIGS. 6(b) and (C) can be made similarly according to equation (2) above. It can be created by a determining means.

Next, a recognition operation using the transition matrix created as described above will be explained.

FIG. 7 is a detailed block diagram of the recognition processing section 9 shown in FIG. 4 above.

In FIG. 7, the speech information input to the phrase speech input section 2I is input to the next stage acoustic processing/comparison section 22. This acoustic processing/comparison section 22 includes a transition matrix memory 26 (corresponding to the memory 8 in FIG. 5). The sound processing section 22 performs feature extraction processing on the phrase audio signal for each single syllable, and the feature pattern for each single syllable is temporarily stored in a buffer within the processing section 22. On the other hand, the storage device 23 stores standard patterns Pi (i=1 to
N) is stored, and this standard pattern Pi is sequentially read out and the processing/comparison section 22 performs a matching calculation with the input feature pattern of the input voice stored in the buffer within the processing section.

According to the prior art, the matching calculation process between the standard pattern and the input feature pattern was all performed on the standard pattern of rebellion, but according to the embodiment of the present invention ((as described later), the matching calculation process between the standard pattern and the input feature pattern is Based on the stored information, the syllable previously recognized as a candidate is matched with a standard pattern of connectable syllables (in the first case, the syllable that may come at the beginning), and the closest match is calculated. The first candidate and the sequentially approximated ones are selected as the next candidates, and the result is stored in the candidate syllable memory 24.In other words, when generating a syllable lattice, the syllable group that does not transition from any of the candidate syllables before -syllable 1 is selected as the next candidate. Processed to exclude it from recognition targets.

The time series of a plurality of I-H sounds % stored in the candidate syllable lattice memory 24 is input to a candidate string output section 27 consisting of a candidate string creation section 25 and a transition matrix memory 26. , by referring to the contents of the transition matrix memory 26 and excluding candidate sequences containing syllable transitions that cannot be transitioned.
Only transitionable candidate sequences are created in the order of combinations with high reliability, and the dictionary collation unit 29 collates these candidate sequences with phrases stored in the dictionary 28. If they match, the result is output to the phrase output unit 30. It is configured to

Next, syllable recognition processing using the transition matrix M(X, Y) will be explained with reference to a block diagram of candidate syllable creation processing using the transition matrix shown in FIG.

In this embodiment, the resulting candidate syllables are temporarily stored in the candidate syllable lattice buffer 24 in chronological order. Further, the above transition matrix information is stored in the memory 26, and the syllable standard pattern is stored in the memory 23.

The recognition results are stored in the complimentary syllable lattice 24 as shown in the following table, and when recognizing the syllable i, the following process is executed.

However, J(i): number of i-th syllable candidates Sij: j-th syllable i candidate syllable number order, previous syllable candidate as i-I) (St when z=o, j=0), the immediately preceding plurality (J(i-1
) candidate syllables, calculate the sum of each row of the transition matrix,
Syllables whose resulting row m (Y) is 0 are designated as non-transitionable.

m(Y)=VM(,5i-1,j,Y)・=・4
3)””M(,fii,-1,1,Y)1M(51-1
,2,Y)+...1M (Si-1,JG-1),Y
) In this formula (3), m(Y) = 0, and the syllable group designated as impossible to transition is excluded and the next similarity comparison process is performed to output the candidate syllable of the i-th syllable. Write to lattice 7. However, when 1=I (syllable of clause class), the syllable group designated as non-transitionable by the 0th row M (0, Y) is excluded and the similarity comparison process is performed.

By repeating the above steps, the creation of the candidate syllable lattice of the -phrasal speech is completed.

Now, when "Kokuminwa" is input as the -bunsetsu speech, the acoustic processing unit 22 extracts features for each syllable, and the feature pattern yi for each syllable is stored in the input pattern time series buffer 31. Next, the process moves on to candidate syllable creation processing using the transition matrix. First, the feature pattern y1 of the first syllable is read into the input cover turn buffer 32, and then step n3
Then, the rows of the transition matrix are specified using the previous candidate syllable group according to Equation 9 (3). In the first case, 2 in step n4.
Then, M(0, Y) in the 0th line is specified, its contents are temporarily stored in buffer 3, and the occurring syllable is specified in step n5.

Next, proceeding to step n6, the input cover turn buffer 32
The feature pattern of the first syllable y1 stored in is loaded, and this feature pattern yl and the syllable standard pattern memory 23
Among the standard patterns stored in the buffer 33, a similarity comparison is made between standard patterns that are designated as occurring syllables and sequentially read out to the standard pattern buffer 34 (step n7), and the results are used as candidate syllables. The syllables are output (step n8), and the result is the candidate syllable lattice 24.
Write it! Sign. In this example, the first syllable candidates are ゝl K OL+, 11 Go'' 11 B □
“is memorized.

Next, the process returns to step n2, and the second syllable feature pattern y2 is input to the pan 7732, and the process proceeds to step n3, where it is added to the first candidate syllable of the candidate syllable lattice 24.
K O” 11 G OLL, 11 B □ LL
M('Sl, l~3.Y) of each row corresponding to is specified, and in step n4, the sum (OR) of the transition matrices is created and the result is temporarily stored in the buffer 33. The designation of the n50 Nashokan was made.

Next, the process moves to step n6, and similar steps n6 to
n9i is executed and the second candidate syllable 11 K U rJ,
11 cu” is stored in the memory 24.

By repeating the above operations, the creation of a candidate syllable lattice for one phrase is completed.

Candidate strings are stored in the candidate syllable lattice 24 as described above.The following table shows examples of the conventional method and the present method when no transition matrix is used for the input speech "Kokumin wa". show.

As is clear from the above example, it can be seen that the correct character strings rise to the top of the candidate strings using this method.

Although the above transition matrix is a first-order transition, it can be extended to a second-order transition and even a general M-order transition using the same method.

Note that the M-order transition matrix can be created according to the above equation (2), and the syllable designation from the previous candidate syllable (up to M syllables) can be performed according to the following equation (4)K.

That is, M-order transition matrix M(Xi, X2, ., XM, Y)
In the case of expansion to
l si−(M−t), j2°°Si−+, jM1jl
"I-J (i-M) j2=1~J (i-(M-1)) jM=l~J(i-I) Number of combinations: J(i-M)・J (i-(M -1))・
・If zu(i-1) (sz, j=o when t≦0), the syllable specification is j+ =I-J(i-M) j2=I~J (i-(M-1)) j M=I~J(i-1) 0 Note that if the order of M is increased, the syllables to be generated will be more limited, and the effect will be greater.

Next, the operation of creating a candidate syllable string using a transition matrix, which is executed by the candidate string output unit 27, will be explained with reference to the processing block diagram of creating a candidate string using a transition matrix shown in FIG. Based on the contents of the candidate syllable lattice memory 24 that stores the time series of a plurality of candidate syllables output from the acoustic processing/comparison section 22 shown in FIG. Candidate sequences are created in order, and the results are temporarily stored in the candidate syllable sequence buffer 42. The candidate syllable string stored in the candidate syllable string ノ(・nofa 42) is determined in the transition matrix reference unit 43 whether transition is possible or not by referring to the transition matrix: M(X, Y) stored in the memory 26. The determination unit 44 makes a determination using the following equation (5), and only possible candidate sequences are stored in the candidate syllable sequence cover sofa 46 via the candidate syllable sequence writing unit 45.

The candidate syllable string of the j-th younger brother is Aj = (at, a2, "', al). However, when the i-th syllable number I of the a-bow is the number of syllables in the string, the transition matrix M (X , Y) is performed when any one of M(al, O)=0 (i=I+1) holds true.

In this equation (5), candidate syllable strings that include non-transitionable syllable strings in which any one of them is true are excluded, and the same determination is made for the next candidate syllable string VC, and only candidate syllable strings that are transitionable are excluded. It is stored in the output buffer 46.

Now, when "Kokuminwa" is used as a -bunsetsu sound, candidate syllables as shown in the following table are stored in chronological order in the candidate syllable lattice memory 24 through processing by the acoustic processing/comparison section 22.

This memo! Based on the 2417 stored syllable lattices, candidate sequences are created in descending order of reliability, and a transition matrix: M(
Only transitionable candidate sequences created by referring to X, Y)'r are output, and in this example, candidate sound 1 sequence is output as follows.

According to the conventional method that does not refer to the transition matrix, r GOKUP I NWA J would be output as the candidate string with the highest reliability, but according to this method, the syllable transition of this candidate string, for example, from 'KtJ'' to It is determined that @PI” is not transitionable using the transition matrix: M(X, Y),
Excluded from subsequent dictionary matching processing.

Although the above transition matrix is a first-order transition, it can be extended to a second-order transition and even a general M-order transition using the same method.

Note that the M-order transition matrix can be created using the above equation (2), and the candidate syllable string can be negated using the following equation (6).

That is, M-order transition matrix: M(XI,X2+”'+X
In the case of expansion of M, Y)-, the j-th candidate column f:Aj-(a
t, a2. ....

ai), then M(ai-M, ai(M-1)l''', ai)=0
(i=I~I+I) −(6) (however, t≦O, t〉
If any one of az=o) is true when I, negation is made.

Note that if the degree of M is increased, the candidate syllable string becomes more limited, and the effect becomes greater.

As described above, when creating a candidate string, the matrix M is referred to for each candidate string, and candidate strings that include combinations of syllables that do not transition are excluded. However, the subdivided units are not limited to syllables, but may also be phonemes or words.

Also, character strings such as alphabets or FORTRAN
It may be a character string of a programming language such as a language.

In general, the present invention can be applied to any character string in which there is a transition relationship between subdivided units constituting a recognition target word.

<Effects> As described above, according to the present invention, since it is possible to extract correct candidate sequences with high accuracy, the accuracy of recognizing correct phrases, etc. is increased, and as a result, the amount of processing such as high-level dictionary matching is reduced. In addition to reducing the number of types of information that should be recognized,
Since the recognition device can create transition matrices by topic, field, etc. as needed each time, depending on the content, topic, field, etc., it is possible to further increase the effectiveness of recognition processing using transition matrices. It is possible.

In addition, in the present invention, transition matrices of different types with the same degree, such as those created for sentences and clauses for each topic: Mi, M
A new transition matrix: M (M=MiUMj) can be easily created from j by summing and composing them.

[Brief explanation of the drawing]

Figure 1 shows a linear transition matrix, Figure 2 shows an example of a transition matrix in which character string transitions are written, Figure 3 shows an example of a transition matrix for bunsetsu character strings, and Figure 4 shows an example of a transition matrix in which character string transitions are written. Figure 5 is a process flow diagram for creating a transition matrix according to the present invention, and Figure 6 is a block diagram showing a specific example of a transition determination means. Figure 7 is a detailed block diagram of the recognition processing unit using a transition matrix, Figure 8 is a process flow diagram for creating candidate syllables using a transition matrix, and Figure 9 is a process for creating a candidate string using a transition matrix. It is a block diagram. 1... Floppy disk device, 3... Keyboard, 4... Changeover switch means, 5... Central processing unit (
CPLI), 8... Transition matrix memory, 9... Recognition processing unit, II... Transition matrix creation instruction function key 0 Agent Patent attorney Yoshihiko Fukushi (and 2 others) a81 Medium ■ 74 Figure 5 Figure 6D x Y y, Xxy Xt Xxy Y Figure 7

Claims (1)

[Scope of Claims] 1. In a recognition device that recognizes one section of information to be recognized using N unit elements that are further subdivided, for a predetermined unit false sequence to be recognized, (N+I) unit elements are recognized. The present invention is characterized by comprising: a transition matrix creating means for creating a transition matrix that describes a transition relationship that is a connection relation; and a processing means for performing recognition processing on the transition matrix created by the transition matrix creating means. recognition device. 2. The recognition device according to claim 1, wherein the information to be recognized in one section is speech information in units of words or phrases, and the unit element string is character strings in units of words or phrases.