JPS6342279B2 - - Google Patents

Description

[Detailed description of the invention]

<Technical Field> The present invention relates to the improvement of a recognition device, and more specifically, the present invention relates to the improvement of a recognition device, and more specifically, the present invention relates to the improvement of a recognition device, and more specifically, the present invention relates to the improvement of a recognition device, and more specifically, the present invention relates to the improvement of a recognition device. This invention relates to improvements in recognition devices. <Prior art> When recognizing one segment of speech, such as a phrase, in more subdivided units such as phonemes, kana, syllables, etc., conventionally, the input speech information of one segment to be recognized is generally processed through acoustic processing, for example. to obtain a feature vector input pattern for each unit such as phoneme, syllable, etc., and to match this input pattern with a pre-stored standard pattern, to select the input information as a candidate unit sequence with a high degree of similarity. The output candidate unit string is compared with the contents of a dictionary of phrases, etc., to recognize one section of information, such as a phrase, for the input information. However, according to such conventional methods, the similarity is calculated by matching the input pattern with standard patterns of all phonemes, syllables, etc., and the similarity is output as candidate syllables etc. in descending order of similarity. ing. Therefore, for example, when recognizing monosyllables including syllables, it is necessary to match the input pattern with more than 100 standard patterns of monosyllables for each syllable, which requires a large amount of processing time. It had become a thing. In addition, it is necessary to perform dictionary matching processing on all of the candidate unit sequences output from those with high similarity, which increases the processing time and does not improve the accuracy of recognizing correct phrases, etc. The amount of processing required for overall recognition was enormous. <Purpose> The present invention aims to provide a recognition device that eliminates the above-mentioned conventional drawbacks, and improves the accuracy of recognizing a section of information such as a correct phrase, and as a result improves overall recognition. The present invention provides a recognition device that can reduce the amount of processing required. <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 string of characters (unit elements) such as a bunsetsu or a sentence, and by this transition matrix creation means. Based on the created transition matrix, when generating a syllable (unit element) latex, a syllable (unit element) that does not transition from any candidate syllable (unit element) one syllable (unit element) before.
Processing means that performs recognition processing such as removing groups from recognition targets, and/or referring to a transition matrix for each candidate sequence when creating candidate sequences, and excluding candidate sequences that include combinations of syllables (unit elements) that do not transition. and,
It is configured to reduce the amount of processing during the next high-level dictionary comparison. First, prior to describing the embodiments of the present invention, a transition matrix showing a transition relationship, which is a connection relationship between unit elements used in the recognition device of the present invention, will be described. In general, if a Japanese sentence is expressed entirely in kana characters, it can be expressed as a syllable string corresponding to the kana character string. For example, the phrase ``earth'' is ``chi'', ``kiyu'', and ``u''.
It is made up of four monosyllable unit elements called "no". The connection relationship between two syllables (“chi” to “kiyu”, “kiyu” to “u”, “u” to “no”) can be studied in all of Japanese or in specific fields.
If you look at sentences etc. in the topic, there are syllable pairs that do not connect (transition: hereinafter we will use the term transition). For example, the syllables in the line “pa” are preceded by “n” and “tsu.”
Nothing else will come. Also, “niya” does not come at the beginning of the word,
“He” (pronounced “he”) does not come at the end of the word. The first-order transition relationship of the syllables constituting such a bunsetsu is described according to the following equation (1), and a transition matrix M(X, Y) as shown in FIG. 1 is created. In Figure 1, the transition matrix M(X, Y) describes the transition from character X to the next character Y in a character string that is a unit element string, and the unit elements (syllables) are N.
, it is a (N+1)×(N+1) matrix,
In terms of hardware, it is stored in ROM etc. In addition, the _Y0 column indicates whether each unit element (1 to N) comes at the beginning of the clause, and the _X0 row contains data indicating whether each unit element (1 to N) comes at the end of the clause. written. For example, FIG. 2 shows an example in which the transition of the character string "red" is written in a transition matrix. The elements of the transition matrix are expressed as either 0 (transition not possible) or 1 (transition possible), and are stored as 1 bit. In FIG. 2, all matrix elements other than the notation "1" are "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 ₁ , a ₂ , a ₃ , ...a _I ) However, when I is the number of characters in the string, the following formula (1) M (0, a ₁ ) = 1, (i = 1) M(a _i-1 , a _i )=1, (i=2~I) M(a _I ,0)=1, (i=I+1) ...According to (1), the characters of the string 〓 The transition relationship is expressed as a transition matrix M
Write to (X, Y). Similarly, write transition relationships for all character strings to be recognized and transition matrix (1
Complete the creation of the following). An example of a concrete transition matrix (first order) M(X, Y) created in this way is shown in FIG.
As is clear from this figure 3, for example (X, Y)
Since the bit position of = (E, KU) is “1”,
There is a transition from “e” to “ku”, and (X,
Since the bit position of Y)=(E, KE) is "0", this indicates that there is no transition from "E" to "KE". The above is a first-order transition, but the second-order transition, and moreover, the M-order transition matrix that is generally extended to the M-order is also expressed by the following formula.
It can be created in accordance with (2). M-order transition matrix: M(X ₁ , X ₂ , X ₃ , ..., X _M ,
Y), (N+1) ^M+1 dimension M(a _iM , a _i-(M-1) , ..., a _i ) =1, (i=1~I+1) ...(2) However, l0 l>I When a _l =0, the embodiment of the present invention establishes a transition relationship that is a connection relationship between (N+1) unit elements for a predetermined unit element sequence to be recognized so that this transition matrix can be created arbitrarily. The recognition device is provided with transition matrix creation means for creating a transition matrix that describes the transition matrix. Next, embodiments 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, reference numeral 1 denotes a floppy disk device, into which a storage medium storing character strings such as phrases or sentences to be recognized is attached. 2 is a character code input terminal, into which a character code of a character string is input from an external device. Further, 3 is a keyboard device, 4 is a changeover switch means, 5 is a central processing unit (CPU), 6 is a character buffer, 7 is a character counter,
8 is a transition matrix memory, 9 is a recognition processing unit, 10 is an input terminal to which speech information to be recognized is input, and 11 is a function key for inputting an instruction signal for creating a transition matrix to the CPU. In the above configuration, when writing desired transition matrix information into the transition matrix memory 8, first operate the function key 11 to instruct the CPU to create a transition matrix, and then switch the switch means 4.
is operated to select the keyboard 3, floppy disk device 1, or other input means, and input the character string to be recognized as a unit. The character string input from the above input method is CPU5
Under the control of , transition matrix information is written into the transition matrix memory 8 according to the process 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, the process moves to a specific transition matrix creation operation (n2), and 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 6 is
It is shifted into the transition determination means consisting of single-digit character buffers 6X, 6Y and single-character delay device 6D shown in Figure a, and the (X, Y) address of the transition matrix memory 8 is specified to the buffers 6X and 6Y in accordance with the contents. At the same time, "1" is written to that address location.
Therefore, by the first shift operation, the first character code is input to the buffer 6Y, and the
=0, Y=a ₁ address position is specified, and "1" is written to that address position. With the next shift operation, the first character code is input to buffer 6X, and the second character code is input to buffer 6Y.
is input, the address position of X=a ₁ and Y=a ₂ in the memory 8 is specified, "1" is written to that address position, and the transition relationship from a ₁ to a ₂ is written.
Thereafter, similar operations are performed corresponding to the contents stored in the character counter 7, and writing of the transition relationship for one character string is completed. Thereafter, similar operations are performed for all character strings to be recognized, completing the creation of the transition matrix (n5, n6). When adding a new word to be recognized, steps n3 to n5 are executed except for the initial setting operation of step n1 in FIG. 5, 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 above formula (2) can be used similarly.
Accordingly, it can be created by the transition determination means shown in FIGS. 6b and 6c. Next, a recognition operation using the transition matrix created as described above will be explained. FIG. 7 is a detailed 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 21 is input to the next stage acoustic processing/comparison section 22. This acoustic processing/comparison unit 22 is a conventionally known part except for a processing part using a transition matrix memory 26 (corresponding to the memory 8 in FIG. 5).
For example, the phrase speech signal input to the phrase speech input section 21 is subjected to feature extraction processing for each monosyllable by the acoustic processing section 22, and the feature pattern for each monosyllable is temporarily stored in a buffer in the processing section 22. . On the other hand, the storage device 23 stores a standard pattern P _i for each single syllable.
(i=1 to N) are stored, and the standard patterns P _i are sequentially read out and the processing/comparison section 22 performs matching calculations with the input feature pattern of the input voice stored in the buffer in the processing section. It will be done. According to the prior art, this matching calculation process between the standard pattern and the input feature pattern was performed for all standard patterns, but according to the embodiment of the present invention, as will be described later, the transition matrix memory 2
Based on the information stored in step 6, the matching with the standard pattern of syllables that can be connected to the syllable previously recognized as a candidate (in the first case, a syllable that may come at the beginning) is calculated, and the most approximate pattern is calculated. Things come first
The candidates and successively approximated ones are selected as the next candidates, and the results are stored in the candidate syllable memory 24. That is, when generating a syllable latex, syllable groups that do not transition from any candidate syllable one syllable before are processed to be excluded from recognition targets. The time series of the plurality of candidate syllables stored in the candidate syllable latex memory 24 is outputted to a candidate sequence output unit 27 consisting of a candidate sequence creation unit 25 and a transition matrix memory 26.
The candidate string output unit 27 refers to the contents of the transition matrix memory 26, excludes candidate strings that include syllable transitions that cannot be transitioned, and selects only transitionable candidate strings in the order of combinations with high reliability. This candidate string created and stored in the dictionary 28 is compared with the clauses stored in the dictionary 28 by the dictionary collation unit 29, and if they match, the result is output to the clause output unit 30. 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 latex buffer 24 in chronological order. Further, the above-mentioned 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 candidate syllable latex 24 as shown in the following table, and when the i-th syllable is to be recognized, the following processing is executed.

[Table] Now _{, the} previous syllable candidates are: ), the immediately preceding multiple pieces (J
The sum of each row of the transition matrix is calculated for (i-1) candidate syllables, and the syllables whose row m(Y) obtained is 0 are designated as non-transitionable. m(Y)=VM(S _i-1,j ,Y) ...(3) =M(S _i-1,1 ,Y)+M(S _i-1,2 ,Y) +...+M(S _{i -1,J(i-1)} , Y) In this equation (3), m(Y) = 0, and the syllable group designated as impossible to transition is excluded and the next similarity comparison process is performed. Output the candidate syllable of the i syllable and write it in the candidate syllable latex 7. However, when i=1 (syllable at the beginning of a clause), the syllable group designated as non-transitionable by the 0th row M(0, Y) is excluded from the similarity comparison process. By repeating the above steps, the creation of candidate syllable latisses for one sentence of speech is completed. Now, when "Kokumin wa" is inputted as a single syllable speech, the acoustic processing unit 22 extracts features for each syllable, and the feature pattern 〓 _i for each syllable is stored in the input pattern time series buffer 31. Next, we move on to the candidate syllable creation process using the transition matrix. First, the feature pattern of the first syllable = ₁ is read into the input pattern buffer 32, and then we move to step n3, where the previous candidate syllable group is used to create the formula (3). ) to specify the rows of the transition matrix. In the first case, M(0, Y) in the 0th line is specified in step n4, its contents are temporarily stored in the buffer 33, and the occurring syllable is specified in step n5. Next, the process moves to step n6, where the characteristic pattern of the first syllable 〓 ₁ stored in the input pattern buffer 32 is loaded, and this characteristic pattern 〓 ₁ and the buffer 33 of the standard patterns stored in the syllable standard pattern memory 23 are loaded. A similarity comparison is made between the standard patterns designated as the occurring syllables and sequentially read out to the standard pattern buffer 34 (step n7), and candidate syllables are output based on the results (step n8). ), and the result is written in the candidate syllable latex 24. In this embodiment, "KO", "GO", and "BO" are stored as first syllable candidates. Next, the process returns to step n2, where the second syllable feature pattern 〓 ₂ is input to the buffer 32, and the process proceeds to step n3, where "KO", "GO", M (S _{l,1 to 3} , Y) of each row corresponding to "BO" is specified, and in step n4, the sum (OR) of the transition matrices is created, the result is temporarily stored in the buffer 33, and the result is temporarily stored in the buffer 33. The specification of the occurring syllable of n5 is completed. Next, move to step n6, and execute similar steps n6 to n9 to produce the second candidate syllables “KU”, “GU”.
is stored in the memory 24. By repeating the above operations, the creation of candidate syllable latisses for one phrase is completed. As described above, candidate strings are stored in the candidate syllable lattice 24. Examples of the conventional method and the present method when no transition matrix is used are shown in the table below for the input speech "Kokuminwa". show.

【table】

【table】

【table】

[Table] 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-th 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 using the following equation (4). That is, M-order transition matrix M (X ₁ , X ₂ ,..., X _M , Y)
In the case of expansion, _the previous syllable _candidate sequence is {X _{1 , X 2} , ... _, 1~J(i-M) j ₂ =1~J(i-(M-1)) 〓 j _M =1~J(i-1) Number of combinations: J(i-M)・J(i- (M-1))
...J(i-1) (S _l,j = 0 when l0), the syllable specification is m(Y) = VM(S _iM,j1 , S _i-(M-1),j2 , ..., S _i-1,jM ,Y) ...(4) j ₁ =1~J(i-M) _j2 =1~J(i
−(M−1)) 〓 j _M =1 to J(i−1). 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 section 27, will be explained with reference to the process block diagram of creating a candidate string using a transition matrix shown in FIG. Based on the contents of the candidate syllable latex 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 strings are created in order, and the results are temporarily stored in the candidate syllable string buffer 42. The candidate syllable strings stored in the candidate syllable string buffer 42 are stored in the transition matrix reference section 4.
Transition matrix stored in memory 26 at 3: M
(X, Y), the determination unit 44 determines whether the transition is possible or not using the following equation (5), and outputs only possible candidate syllable strings via the candidate syllable string writing unit 45. I will store it in Batsuhua 46. Now, the j-th candidate syllable string is 〓 _j = (a ₁ , a ₂ , ... a _I ) However, when a _i is the i-th syllable number and I is the number of syllables in the string, the transition by the determination unit 44 Matrix M(X,
Candidate sequence negation using Y) is M(0, a ₁ )=0 (i=1) M(a _i-1 , a _i )=0 (i=2~1) M(a _I ,0)= 0 (i=I+1)...This is done when any one of (5) holds true. In this equation (5), candidate syllable strings containing non-transitionable syllable strings in which one of the conditions holds true are excluded, the same judgment is made for the next candidate syllable string, and only transitional candidate syllable strings are output. Batsuhua 4
6 is stored. Now, when "Kokuminwa" is inputted as a single sentence speech, candidate syllables as shown in the following table are stored in chronological order in the candidate syllable latex memory 24 through processing by the acoustic processing/comparison section 22.

[Table] Based on the syllable lattice stored in this memory 24, candidate strings are created in order of reliability, and the candidate strings created by referring to the transition matrix: M(X, Y) are transitionable ones. In this example, the candidate syllable string is output as follows.

[Table] According to the conventional method that does not refer to the transition matrix, "GOKUPINWA" is output as the candidate string with the highest reliability, but according to this method, the syllable transition of this candidate string, for example "KU", is output. Therefore, it is determined that "PI" cannot be transitioned using the transition matrix: M(X, Y), and is excluded from the subsequent dictionary matching process. 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-th 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(X ₁ , X ₂ , ..., X _M ,
Y), the j-th candidate column is 〓 _j = (a ₁ ,
a ₂ ,…, a _I ) then M(a _iM , a _i-(M-1) ,…, a _i ) =0 (i=1~I+1)…(6) (However, if l0, l>I If any one of the _following holds true, negation is achieved. 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. The recognition target of the above-mentioned recognition device is not limited to phrases.
It may be a syllable, a word, or a sentence, and the subdivided unit is not limited to a syllable, but may be a phoneme or a word. Also, character strings such as alphanumeric characters or
It may also be a character string in a programming language such as FORTRAN. 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 can be reduced. At the same time, transition matrices can be created in the recognition device according to the type, content, topic, field, etc. of the information to be recognized. It is possible to further increase the effect of recognition processing using matrices. In addition, in the present invention, a new transition matrix can be easily created by summing and synthesizing transition matrices of different types with the same degree, such as those created for sentences and clauses for each topic: M _i and M _j . M (M=
M _i ∪M _j ) can be created.

[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. FIG. 5 is a block diagram showing the configuration of an embodiment of a recognition device implementing the present invention. FIG. 5 is a processing flow diagram for creating a transition matrix according to the present invention. FIG. Figure 7 is a detailed block diagram of the recognition processing unit using a transition matrix, Figure 8 is a processing flow diagram of candidate syllable creation using a transition matrix, and Figure 9 is a processing block diagram of candidate string creation using a transition matrix. be. DESCRIPTION OF SYMBOLS 1...Floppy disk device, 3...Keyboard, 4...Switching means, 5...Central processing unit (CPU), 8...Transition matrix memory, 9...Recognition processing section, 11...Transition matrix creation Instruction function key.

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 sequence of unit elements to be recognized, (N+
1) A transition matrix memory that stores information on whether connections between unit elements are possible in the relationship between rows and columns, a means for inputting a character string to be recognized, and a code for the input character string. transition matrix creation means for sequentially reading the encoded information in the storage means character by character, designating a corresponding address in the transition matrix memory, and storing connectable information at that address; , processing means for performing recognition processing based on the transition matrix created by the transition matrix creation means.