JPH0652478B2 - Recognition device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an improvement in a recognition device, and more specifically, for example, a piece of information to be recognized, such as a segment of speech, such as a syllable, is converted into a phoneme, a kana, a syllable, a segment, or the like. The present invention relates to an improvement of a recognition device that recognizes subdivided unit elements.

<Prior Art> When recognizing a segment of speech such as a syllable in a more finely-divided unit such as a phoneme, a kana, or a syllable, generally, a segment of speech information or the like to be recognized that has been input is typically processed by acoustic processing. Then, a feature vector input pattern for each unit such as phonology or syllable is obtained, and the input pattern is matched with a standard pattern stored in advance, and the input information has a high degree of similarity as a candidate unit sequence. Then, the candidate unit sequence outputted is collated with the contents of the dictionary such as bunsetsu to recognize the delimiter information such as bunsetsu to the input information.

However, according to such a conventional method, all phonemes,
A standard pattern such as a syllable is matched with an input pattern to calculate a similarity, and the similarity is output as candidate syllables in order.

Therefore, for example, in the case of recognizing a single syllable unit including a jumble, it is necessary to perform matching between 100 or more types of standard patterns of single syllables and input patterns for each syllable unit, which requires a long processing time. Was becoming.

In addition, after that, it is necessary to perform dictionary matching processing on all candidate unit strings output from the ones with a high degree of similarity, the processing time becomes long, and the accuracy of recognizing a correct phrase etc. does not improve, resulting in In addition, the amount of processing required for overall recognition was enormous.

<Purpose> An object of the present invention is to provide a recognition device in which the above-mentioned conventional drawbacks are eliminated, and improve the accuracy of recognizing a segment of information to be recognized such as a correct phrase and different problems and fields. (EN) A recognition device capable of designating a process according to information of a type to be recognized and consequently reducing the amount of processing required for the entire recognition.

<Example> Hereinafter, an example in which the recognition apparatus of the present invention recognizes a segmented speech input such as a syllable with a more subdivided unit element such as a syllable will be described as an example.

According to the embodiment of the present invention, the recognition device recognizes information to be recognized such as a segment of speech such as a syllable with N subdivided unit elements such as phonemes, kana, and syllables. Transitions that store multiple types of different transition matrices that describe transition relationships that are connection relationships between (N + 1) characters (unit elements) for each topic or field for character (unit element) strings such as clauses or sentences Matrix storage means,
From a plurality of types of transition matrices stored in the transition matrix storage means, a transition matrix designation means for designating a desired transition matrix according to the content (type, field, etc.) of the character string to be recognized, and the transition matrix designation means. A syllable (unit element) group that does not transition from any candidate syllable (unit element) before one syllable (unit element) based on the transition matrix specified by And a processing unit that performs recognition processing such as excluding or referring to the transition matrix for each candidate string when creating the candidate string, and excluding candidate strings that include combinations of syllables (unit elements) that do not transition, It is configured to reduce the processing amount in the next high-order dictionary matching.

First, prior to the description of 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 apparatus of the present invention will be described.

In general, a Japanese sentence can be represented by a syllable string corresponding to a kana character string when all the kana characters are expressed. For example, the phrase "Earth's" consists of four unit syllables called "chi", "kyu", "u", and "no". The connection relationship between two syllables (“chi” to “kyu”, “kyu” to “u”, “u” to “no”) in all Japanese, or in a particular field or topic When examined, there are syllable pairs that are not connected (transition; hereinafter, the term transition is used). For example, there is nothing other than "n" and "tsu" in front of the syllable of Payo. Also, "nya" does not come at the beginning of the word, and "he" (what happens to he) does not come at the end.

A transition matrix M (X, Y) as shown in FIG. 1 is created by describing the first-order transition relation of syllables forming such a syllable according to the following equation (1).

In FIG. 1, the transition matrix M (M, Y) describes the transition from the character X of the character string that is the unit element string to the next character Y, and when the number of unit elements (syllables) is N, ( N + 1) ×
This is an (N + 1) matrix and is stored in a ROM or the like in terms of hardware. The Y ₀ column indicates whether each unit element (1 to N) comes to the beginning of a node, and the X ₀ row contains data indicating whether each unit element (1 to N) comes to the beginning of a node. Written.

For example, FIG. 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 are represented by either binary values of 0 (non-transitional) and 1 (transitional), and 1
Stored in bits. In FIG. 2, all matrix elements other than the notation “1” are “0”, and the display thereof is omitted.

The creation of the transition matrix will now be explained in some detail.

First, when creating the transition matrix, the transition row example memory is set to "0".
Initially set [M (X, Y) = 0].

Then the string However, when I is the number of characters in the column, the following equation (1) According to the string Write the character transition relation of the above into the transition matrix M (X, Y). Similarly, the creation of the transition matrix (first order) in which the transition relationship is written for all the character strings to be recognized 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, since the bit position of (X, Y) = (e, ku) is “1”, there is a transition from “e” to “ku”, and (X, Y) Since the bit position of (Y) = (e, ke) is “0”, it means that there is no transition from “e” to “ke”.

The above is a first-order transition, but a second-order transition, and generally M
Similarly, the M-th order transition matrix expanded to the next can be created according to the following equation (2).

Mth-order transition matrix: M (X ₁ , X ₂ , X ₃ , ..., X _M , Y), (N + 1) ^{M + 1} dimensional M (a _iM , a _{i- (M-1)} , ..., a _i ) = 1, (i = 1 to I + 1) ... (2) In the embodiment of the present invention, a plurality of this transition matrix is provided for each type of recognition target, topic, field, etc., and a specific transition matrix can be selected as necessary to execute the recognition processing. .

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, reference numeral 1 is a transition matrix designating means, which is connected to a central processing unit (CPU) and is constituted by a selection key provided on the operation surface or a selection input means by voice. Further, 3 is an input unit for inputting voice information to be recognized, 4 is an amplification unit, 5 is an acoustic processing unit, and 61, 6
2, ..., 6K are transition matrix storage means for storing different types of transition matrices, and 7 is a recognition processing unit.

In the above configuration, the transition matrix storage means 61, 62,
.., 6K have different types of transition matrices created from sentences in different fields (for example, science, literature, economy, etc.), and the voice information input to the input unit 3 now is, for example, science. If it is related, the transition matrix designating means 1 is operated to select and designate a transition matrix storing means (for example, M1) created from a scientific text or the like, and the selection and designation is stored in the memory M1. The recognition processing operation is performed by the recognition processing unit 7 using the transition matrix.

Next, the recognition operation using the transition matrix selected and designated according to the type (field) of information to be recognized as described above will be described.

FIG. 5 is a detailed block diagram of the acoustic processing unit 5 and the recognition processing unit 7 shown in FIG.

In FIG. 5, the voice information input to the phrase voice input unit 21 is input to the next acoustic distance / comparison unit 22. The acoustic processing / comparing unit 22 is a conventionally known one except for the processing unit using the transition matrix memory 26. For example, the phrase speech signal input to the phrase speech input unit 21 is the acoustic processing unit 22.
The feature extraction process is performed for each single syllable, and the feature pattern for each single syllable is temporarily stored in the buffer in the processing unit 22. On the other hand, the storage device 23 stores a standard pattern P _i (i = 1 to N) for each monosyllabic, and the standard pattern P _i is sequentially read and stored in a buffer in the processing unit in the processing / comparing unit 22. A matching calculation of the stored input voice with the input feature pattern is performed.

According to the conventional technique, the matching calculation processing of the standard pattern and the input feature pattern is performed for all the standard patterns, but according to the present embodiment, the information stored in the transition matrix memory 26 is described later. The matching with the standard pattern of the syllable that can be connected to the syllable recognized as a candidate (the syllable that may come first in the first case) is calculated, and the closest match is the first candidate. , Which are successively approximated are selected as the next candidates,
The result is stored in the candidate syllable memory 24. That is, when the syllable lattice is generated, a syllable group that does not transition from any candidate syllable one syllable before is processed to be excluded from the recognition contrast.

The transition matrix memory 26 corresponds to one memory (M ₁ ) of the transition matrix storage means 61, 62, ..., 6K designated by the transition matrix designation means 1.

The time series of the plurality of candidate syllables stored in the candidate syllable lattice memory 24 is input to the candidate sequence output unit 27 including the candidate sequence creation unit 25 and the transition matrix memory 26, and the candidate sequence output unit 27 specifies a specific sequence. By referring to the contents of the transition matrix memory 26 corresponding to a topic, a field, etc., a candidate sequence including syllable transitions that cannot be transitioned is excluded, and only transitionable candidate sequences are created in order of high reliability combination. Columns and dictionaries 28
The dictionary matching unit 29 matches the phrase stored in
If they match, the result is output to the phrase output unit 30.

Next, the syllable recognition process using the transition matrix M (X, Y) will be described with reference to the block diagram of the candidate syllable creation process 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. The transition matrix information described above is stored in the memory 26, and the syllable standard pattern is stored in the memory 23.

The recognition result is stored in the candidate syllable lattice 24 as shown in the following table, but when recognizing the i-th syllable, the following processing is executed.

However, J _(i) : i-th syllable candidate number Sij: j-th syllable i candidate syllable number Now, the previous syllable candidate is X = {S _{i-1, j} } j = 1 to J (i-1) Number of combinations: J If (i-1) (S _{l, j} = 0 when _l = 0), the immediately preceding plural number (J (i-1)
Number) of candidate syllables, the transition matrix is summed, and the syllable in which the obtained row m (Y) is 0 is designated as not 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 non-transitionable is excluded, and the next similarity comparison process is performed. Done,
The candidate syllable of the i-th syllable is output and written in the candidate syllable lattice 7. However, when i = 1 (syllable at the beginning of the syllable), the 0th row M
A syllable group designated by (0, Y) as a transition impossible is excluded, and the similarity comparison process is performed.

By repeating the above, the creation of the candidate syllable lattice of the one-sentence voice is completed.

Now, when "national" is input as one syllable voice, the acoustic processing unit 22 performs feature extraction for each syllable, and the feature pattern for each syllable. Are stored in the input pattern time series buffer 31. Next, the process shifts to the candidate syllable creation process using the transition matrix, and first the first
Syllable feature pattern Is read into the input pattern buffer 32, and then the process proceeds to step n3 to specify the row of the transition matrix by the previous candidate syllable group according to the equation (3). Step n4 for the first case
At M, M (0, Y) in the 0th line is designated, the content is temporarily stored in the buffer 33, and the occurrence syllable is designated in step n5.

Next, in step n6, the input pattern buffer 32
First syllable stored in This feature pattern is loaded. And the standard pattern stored in the syllable standard pattern memory 23 are compared with the standard pattern designated by the buffer 33 of the standard pattern and sequentially read out to the standard pattern buffer 34 (step n7). Based on this, the candidate syllable is output (step n8), and the result is written in the candidate syllable lattice 24. In this embodiment, the first syllable candidates are "KO", "GO",
“BO” is stored.

Next, returning to step n2, the second syllable feature pattern Is input to the buffer 32, the process proceeds to step n3,
Candidate Syllables Based on the first candidate syllable of Lattice24, "K
M (S of each row corresponding to "O", "GO", "BO"
₁ , 1 to ₃ , Y) are specified, and in step n4, the sum (OR) of the transition matrices is created and the result is stored in the buffer 3
3 is temporarily stored, and the occurrence syllable is designated in step n5.

Next, the process proceeds to step n6, and the same steps n6 to
n9 is executed to store the second candidate syllables “KU” and “GU” in the memory 24.

The above operation is repeated to complete the creation of the one-syllable candidate syllable lattice.

As described above, the candidate examples are stored in the candidate syllable lattice 24. However, an example of the case of the conventional method without using the transition matrix and an example of the case of this method are shown in the following table for the input voice "Kuniwa". Show.

As is clear from the above example, it can be seen that the correct character string is higher in the candidate string in this method.

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

Note that the M-th order transition matrix is created according to the above equation (2), and the syllable designation from the previous candidate syllable (up to M syllables before) is given by the following equation.
It can be done by (4).

That is, in the case of extension to the Mth-order transition matrix M (X ₁ , X ₂ , ..., X _M , Y), the preceding syllable candidate sequence is Number of combinations: J (iM) ・ J (i- (M-1))… J (i-1) (Sl _{, j} = 0 when l0) It will be done by.

It should be noted that if the order of M is increased, the limitation of the generated syllable becomes stronger and the effect becomes larger.

Next, the operation of creating a candidate syllable string using the transition matrix executed by the candidate string output unit 27 will be described with reference to the processing block diagram of creating the candidate string using the transition matrix shown in FIG.

Based on the contents of the candidate syllable lattice memory 24, which stores the time series of the plurality of candidate syllables output from the acoustic processing / comparing unit 22 shown in FIG. The candidate strings are created in descending order, and the results are temporarily stored in the candidate syllable string buffer 42. The candidate syllable sequence stored in the candidate syllable sequence buffer 42 is the transition matrix: M stored in the memory 26 in the transition matrix reference unit 43.
With reference to (X, Y), the determination unit 44 determines whether transition is possible or not according to the following equation (5), and only the possible candidate strings are stored in the candidate syllable string output buffer 46 via the candidate string writing unit 45. I will remember.

Now the i-th candidate syllable string However, when a _{i is} the i-th syllable number I is the number of syllables in the sequence, the candidate sequence negation using the transition matrix M (X, Y) by the determination unit 44 is It is done when any one of the above is established.

In this equation (5), candidate syllable strings that include non-transitionable syllable strings for which any one of the conditions holds are excluded, the same determination is performed for the next candidate syllable string, and only transitionable candidate syllable strings are output. It is stored in the buffer 46. Now, when "national" is input as the one-syllable voice, the candidate syllables as shown in the following table are chronologically stored in the candidate syllable lattice memory 4 by the processing of the sound processing / comparing unit 2.

Based on the syllable lattice stored in the memory 24, candidate sequences are created in the order of high reliability, and the transition matrix · M (X, Y)
Only those candidate strings that can be transitioned are created and output in this example. In this example, candidate syllable strings are output as follows.

According to the conventional method that does not refer to the transition matrix, "GOKUPINWA" is output as the candidate sequence having the highest reliability, but according to this method, the transition of the syllables of this candidate sequence, for example, from "KU" to "PI If "is not transitionable, the transition matrix:
It is determined using 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 further to a general M-th order transition by the same method.

It should be noted that the M-th order transition matrix can be created according to the above equation (2), and the negation of the candidate syllable sequence can be performed by the following equation (6).

That is, in the case of the extension part to the Mth-order transition matrix: M (X ₁ , X ₂ , ..., X _M , Y), the j-th candidate sequence is Then, M (a ^iM , a _{i- (M-1)} , ..., a _i ) = 0 (i = 1 to I + 1)… (6) (However, if l0 and l> 1, a _l = 0) If any one of the above is established, the denial is made. Note that if the order of M is increased, the limitation of the candidate syllable string becomes stronger, and the effect becomes larger.

As described above, at the time of creating a candidate string, the matrix M is referred to for each candidate string, and the candidate string including a combination of syllables that do not transition is excluded.

The recognition target of the above-described recognition device is not limited to syllables, and may be syllables, words, and sentences, and the subdivided units are not limited to syllables, but may be phonemes and words.

It may be a character string such as an alphabet or a character string of a programming language such as FORTRAN language.

In general, the present invention can be applied to any character string having a transitional relationship of subdivided units forming a recognition target word.

<Effect> As described above, according to the present invention, since a correct candidate sequence can be extracted with high accuracy, the accuracy of recognizing a correct phrase or the like is increased, and as a result, the processing amount of high-order dictionary matching or the like is increased. The type of information that can be reduced,
Depending on the content, topic, field, etc., the transition matrix of topic, field, etc. can be arbitrarily selected and used as needed each time, so that the effect of the recognition process using the transition matrix is further increased. It is possible.

In the present invention, transition matrices of different kinds having different homogeneities such as those created for sentences and phrases for each topic; M _i , M
_A new transition matrix; M (M = M _i UM _i ) can be easily created by taking the sum of _j and combining them.

[Brief description of drawings]

FIG. 1 is a diagram showing a first-order transition matrix, FIG. 2 is a diagram showing a transition matrix in which transitions of character strings are written, FIG. 3 is a diagram showing an example of a transition matrix of phrase strings, and FIG. 4 is a book. FIG. 5 is a block diagram showing a configuration of an embodiment of a recognition device embodying the invention, FIG. 5 is a detailed block diagram of a recognition processing section using a transition matrix, and FIG. 6 is a processing flow chart of candidate syllable creation using a transition matrix. FIG. 7 is a processing block diagram of creating a candidate sequence using a transition matrix. 1 ... Transition matrix designating means, 2 ... Central processing unit (CPU),
61, 62, ..., 6K ... Transition matrix storage means, 7 ... Recognition processing unit.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (56) References Japanese Patent Laid-Open No. 56-29299 (JP, A) Kazuo Nakata, “Synthesis and Recognition of Speech” (Sho 55-7-20) General Electronic Publishing, P. 124-127 IEICE Pattern Recognition and Learning Study Material PRL75-55 (1975) P. 11-20 IEICE Technical Committee on Pattern Recognition and Learning Material PRL73-72 (November 1973) P. 12

Claims (1)

[Claims] 1. An apparatus for recognizing a single segment of information to be recognized by N more subdivided unit elements, wherein (N + 1) unit elements can be connected for a predetermined unit element sequence to be recognized. A plurality of different transition matrix memories that store information on whether or not there is a transition matrix memory designating unit that designates a predetermined transition matrix memory from among the plurality of types of transition matrix memories; and the transition matrix memory designating unit. A recognition device, comprising: a processing unit that performs recognition processing based on data of a designated transition matrix memory.