JPS5978400A - Preparation system of candidate train - 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 performs recognition in units of recognition such as kana characters, monosyllables, words, etc., and then combines a plurality of consecutive recognition results to derive one candidate string. The present invention relates to an improvement of a candidate sequence creation method that employs a plurality of recognition results for each recognition unit and creates a plurality of candidate sequences so that candidate sequences with high recognition accuracy as a whole can be sequentially processed.

<Prior art> In conventional speech recognition devices, for example, the method of recognizing each syllable and matching that syllable string with a dictionary as one phrase (word) involves taking multiple recognition results for each syllable. , a character candidate string is created by combining these characters, and the character candidate string is sequentially checked against a dictionary. This method produces character candidate strings whose number is the combination of the length of the phrase and the number of candidates for each syllable.

Therefore, if dictionary matching were to be performed for all of these combinations, an extremely large amount of processing time would be required as the number of syllables and recognition candidates increased.

Purpose of the present invention The purpose of the present invention is to provide a candidate sequence creation method that eliminates the conventional Φ problem. By performing processing, a method for creating a candidate string is provided in which the amount of processing does not increase significantly even as the number of syllables and recognition candidates increase.

<Example> Hereinafter, a speech recognition device implementing the candidate sequence creation method of the present invention will be described in detail by giving an example.

FIG. 1 is a block diagram showing the configuration of a speech recognition device using a monosyllabic speech standard pattern to which the candidate string creation method of the present invention can be applied.

In FIG. 1, a syllable speech input applied to an input terminal 1 is inputted to a monosyllable recognition section 3 via a syllable speech identification section 2 at the next stage. This monosyllable recognition unit 3 is conventionally known. For example, the syllable speech recognition unit 2 classifies the syllable speech input from the input terminal into syllable units, and the acoustic processing/comparison unit 5 provides characteristics for each monosyllable. Extraction is performed, and the characteristic pattern for each single syllable is temporarily stored in a buffer memory within the processing unit 5. On the other hand, the storage device 6 stores a standard pattern Pi (i=I to N) for each monosyllable, and this standard pattern Pi is sequentially read out to form the input speech stored in the buffer memory in the processing section 5. Matching calculation processing with the input feature pattern is performed.

As a result of this matching calculation process, the most approximated one is selected as the first candidate, and the sequentially approximated ones are selected as the next candidates, and the results are stored as a syllable lattice in the candidate syllable memory 4 together with distance difference information indicating the degree of approximation. Ru.

Each syllable unit candidate recognized by the monosyllable recognition unit 3 and stored in the memory 4 as a syllable lattice is manually input to the candidate string creation unit 7.

As will be described later, the candidate sequence creation unit 7 includes a candidate sequence memory for storing candidate sequences, a final pointer memory for indicating the last candidate sequence number, and one of the candidate sequences stored in the candidate sequence memory. comprising a fixed position storage memory for storing information corresponding to the position of the least accurate candidate sequence among recognition units that are not number one, and a maximum value setting memory for instructing the maximum number of candidate sequences; First, the above recognition result (
Candidate syllables) A candidate string is created by arranging only the first recognition results stored in the memory and stored in the candidate string memory.First recognition results (candidate syllables) Highly accurate recognition of the recognition results stored in the memory. Create a new candidate column by combining the candidates with the contents of the candidate column memory, sort the candidate columns in descending order of overall accuracy, and update the contents of the fixed position memory and final pointer memory. However, if the content of the final pointer memory exceeds the value corresponding to the value in the maximum value setting memory, it is fixed at that value, and if the value in the fixed position memory memory exceeds the value corresponding to the value in the maximum value setting memory. If there are no more cases or syllable candidates, the process of creating a candidate string is terminated.

The candidate strings created in the candidate string creating section 7 and stored in the candidate string memory in the candidate string creating section 7 are sequentially output in descending order of accuracy, and are compared with the phrases stored in the dictionary 8 and the dictionary matching section 9. , Verified and matched phrase output unit 1
It is configured so that it is output as 0.

Next, the above-mentioned candidate sequence creation section 7 will be explained in detail.

FIG. 2 is a block diagram showing a specific configuration of the candidate sequence creation section 7, and the same parts as in FIG. 1 are indicated by the same symbols.

In FIG. 2, the candidate column creation unit 7 includes a candidate column memory 11 that stores candidate columns, a maximum value setting memory 12 that specifies the maximum number of candidate columns, and a final pointer memory 13 that specifies the last candidate column number. a maximum value setting memory 14 for storing information corresponding to a candidate sequence position having the lowest probability as a candidate sequence among those in which only one recognition unit in the candidate sequence stored in the candidate sequence memory is not ranked first; Each recognition unit (syllable)
The search point memo month 5 stores information indicating how many candidates in the position were used to create the candidate string, and only the first recognition result stored in the above recognition result (candidate syllable) memory 4. are lined up to create a candidate string, which is stored in the candidate string memory 11, and first, the recognition results (candidate syllables) are sequentially stored in the memory 4.
A new candidate string is created by combining the recognition candidates with high accuracy of the recognition results stored in the memory 11 with the contents of the candidate string memory 11, and the candidate strings are sorted in descending order of overall accuracy. The contents of the position memory memo IJ 14 and the final pointer memory 13 are updated, and if the contents of the final pointer memory 13 exceed the value corresponding to the value of the maximum value setting memory 12, the value is fixed and the final pointer memory 13 is updated. 14 exceeds the value corresponding to the value in the maximum value setting memory, or when there are no more syllable candidates, the central processing unit (CPU) is programmed to terminate the process of creating a candidate string.] It consists of 6.

The candidate syllable memory 4 has each syllable 8 as shown in FIG.
Area A that stores a plurality of candidate syllables (recognition results) for each of 1 to Sn and the difference in distance (ratio) in recognition from the first candidate.
and D, and the recognized monosyllable identification number (for example, corresponding to a JIS code) is stored in area A.

In Figure 3, a r j is j of the i-th syllable.
In addition, daij represents the distance between the i-th input syllable and aij.

Further, in this embodiment, the first candidate distance difference region P1
~Pn is configured to be substituted as the search point memory 15.

Further, as shown in FIG. 4, the candidate sequence memory 11 is composed of an area for storing candidate sequences in descending order of accuracy and an area M for storing the total distance of the candidate sequences. The configuration is such that a value indicating which monosyllable candidate is used is stored.

Next, the operation of the apparatus configured as described above will be explained according to the processing flow shown in FIG.

The recognition operation starts at step nl in Fig. 5.

Hereinafter, step ni will be simply referred to as ni. ) The speech input to be recognized that is applied to the input terminal 1 is sent to the next stage syllable speech recognition section 2.
is input to the monosyllable recognition unit 3 via the monosyllable recognition unit 3, where monosyllable recognition is executed, and the result is stored in the recognition result memory 4 together with distance difference information from the first recognition result (n2, n3). This operation is performed in units of words or phrases, and when a word or phrase is detected (n4), the next step n
Move to 5.

The phrase to be recognized now is ``/ya/ma/ni/'', which is stored as a syllable lattice as shown in Table 1.

Next, the process moves to the candidate column creation operation, and is sequentially processed as follows.

(2) First, in step n5, the first candidate syllable in the recognition result memory 4 is input to the beginning of the candidate string memory 11 to create a first candidate string. This behavior causes
The contents of the candidate column memory 11 are as shown in Table 2.

(2) Next, the contents of the fixed position storage memory 14 and the final pointer memory 13 are set to the initial value "1". (n6).

Further, the contents of the maximum value setting memory 12 are set in advance to a value corresponding to the number of candidate columns desired to be created.

(2) Next, proceed to step n7, and select syllable candidates with high accuracy from the recognition results for the remaining unused monosyllables in the inner layer. That is, a syllable candidate with the smallest distance difference is selected from the remaining syllable lattices that have not been used so far, and the contents of the search point memory 15 corresponding to the selected syllable candidate are updated (n8). Next, put the recognition result (rank number of the selected syllable) in the position of the selected syllable, put the first candidate (0) in the position of the other syllables, and put the distance difference in the position of the sum of distances. Create column (nlO).

Next, the created candidate string is inserted and stored in the candidate string memory 11 at a position in ascending order of the sum of distances between the ranks indicated by the values of the fixed position storage memory 14 and the final pointer memory 13 (nll).

Next, the contents of the fixed position storage memory 14 are set to the position next to the insertion position (n12), and the contents of the final pointer memory 13 are updated and incremented by "1'' (n15). Note that if the value in the final pointer memory 13 exceeds the value in the maximum value setting memory 12, the value in the memory 13 is fixed to the maximum value in the memory 12.

Through this series of operations, the second syllable "ma" is selected as the most accurate syllable candidate, and the total distance is 2.
0 is created and the warping result is inserted into the second position of the candidate column memory 11.

As a result, the stored contents of the memories 11, 13, and 14 become as shown in FIG.

(2) Next, search the position of the candidate string memory 11 from the beginning to the river to find one where the position of the currently selected syllable is 0 (n16), and if there is such a candidate string (n17), move to step n18,
The difference in distance is added to the sum of the distances, and the candidate column memo IJ
Insert newly created candidate columns in ascending order of distance within I. Next, returning to step n15, the contents of the final pointer memory 13 are updated and incremented by "1'', and the operation (2) is repeated until the position indicated by the contents of the final pointer memory 13 is reached.

If the selected syllable position is not 0, the process returns to step n7 through steps n16 and n17, and the above-described operation (2) is repeated again.

■, The operations of ■ and ■ above are performed until the value of the fixed position storage memory 14 exceeds the value corresponding to the maximum value of the maximum value setting memory 12 (n18.n and then D, or until there are no syllable candidates (n9.n19) , n20).

Through the above-described operation, a candidate string is created without an exponential increase in processing as the number of syllables in a phrase or the number of candidates for each syllable increases.

To specifically show the state of processing after the state shown in Figure 6, in the process of ■, '' is selected for '' in the third syllable that is the next closest to 'ma', and as shown in Figure 7. Memory I+, 1
3.14 memory contents change.

Next, proceeding to the processing step (■), the fourth candidate string as shown in Figure 8 is created, and returning to the processing step (■), a new monosyllable candidate "gya" is selected, and 5th as shown in
A candidate string for the 6th to 8th place is created as shown in FIG.

The following steps are executed in the same manner.

At this time, the ranking of candidate columns before the position indicated by the contents of the determined position storage memory 14 is determined at this point.
The insertion target is from the ranking position indicated by the contents of the fixed/fixed position storage memory 14 onwards. Therefore, the process can be ended when the contents of the fixed position storage memory 14 exceed the storage contents of the maximum value setting memory I2.

The candidate string created by the above-described processing operations is compared with the contents of the dictionary 8 by the matching unit 9, and the matching items are outputted to the output unit IO as the phrase (word) recognition results.

In addition, if the candidate string stored in the candidate string memo IJ 11 is stored as a value indicating the number of the candidate for each syllable, the identification number of the monosyllable such as JIS code is stored depending on the stored contents of the candidate syllable memory 4. After being converted into columns, a comparison with the dictionary 8 will be performed.

Effect> As described above, according to the present invention, there is a recognition result memory that stores distance information corresponding to the difference between a recognition result in a recognition unit and the first recognition accuracy, and a recognition result memory that stores a candidate sequence. A column memory and a fixed position storage memory that stores information corresponding to the candidate column position with the lowest probability among those in which only one recognition unit in the candidate column stored in the candidate column memory is not ranked first. First, a candidate sequence is created based on the first recognition result stored in the recognition result memory and stored in the candidate sequence memory, and then the recognition results sequentially stored in the recognition result memory are A new candidate sequence is created by combining the highly accurate recognition candidates with the contents of the recognition result memory described in 2. Since the candidate columns are sorted in descending order of accuracy after the candidate column position, the candidate columns can be sorted without causing an exponential increase in processing as the number of syllables in a phrase or the number of candidates for each syllable increases. You can create columns.

In the above embodiments, the process for recognition of monosyllabic speech was explained as an example, but the candidate string creation method of the present invention can also be applied to other word speech 2 handwritten kana character recognition, etc. Needless to say.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of a speech recognition device to which the present invention is applied, FIG. 2 is a block diagram showing the configuration of main parts of the device implementing the present invention, and FIG. 3 is a configuration example of a candidate syllable memory. 4 is a diagram showing a configuration example of a candidate column memory, FIG. 5 is a diagram showing a flowchart of a candidate column creation process of the present invention, and FIGS. 6 to 10 are a diagram showing a candidate column creation process of the present invention, respectively. FIG. 3 is a diagram showing the storage state of a candidate column memory at each stage. 3... Single syllable recognition section, 4... Candidate syllable (recognition result) memory, 7... Candidate string creation section, 11... Candidate string memory, 12... Maximum value setting memory, 13... Final pointer memory, 14... Determined position storage memory, 16...
Central processing unit. Agent Patent Attorney Aihiko Fuku (and 2 others) Figure 1 SI? a J l'(l 5/7 No. 9 (S No. 10 No. 1 1) Indication of the case Patent application 1986-190642 2. Method for creating a list of candidates for the name of the invention 3 Person making the amendment Relationship with the case Patent applicant director Tadahiro Nagasekimoto 4 Agent Address = 545 22-22 Nagaike-cho, Abeno-ku, Osaka City
No. 6, Subject of amendment (1) Detailed explanation of the invention column 7 of the specification, Contents of amendment (1) Changed the statement "for each single syllable" on page 4, line 8 of the specification to "single single syllable" (2) Maximum value setting memo IJI4 on page 7, line 3 of the same book.
The description of J has been corrected to "Fixed position memory 14". (3) Figure 3 in the figure has been corrected as shown in the attached sheet. Above Figure 3 1

Claims (1)

[Claims] 1. In a candidate sequence creation device that creates a plurality of candidate sequences by employing a plurality of recognition results from each recognition unit after performing recognition in a predetermined recognition unit, and a recognition result memory that stores distance information corresponding to the difference between the result and its recognition accuracy. A candidate sequence memory that stores candidate sequences; and information corresponding to the candidate sequence position with the lowest probability as a candidate sequence among those in which only one recognition unit in the candidate sequence stored in the candidate sequence memory is not ranked first. and a fixed position storage memory for storing, first creating a candidate sequence based on the first recognition result stored in the recognition result memory and storing it in the candidate sequence memory, and then sequentially storing it in the recognition result memory. A new candidate sequence is created by combining the stored recognition results with the contents of the recognition result memory, and in response to the contents of the confirmed position storage memory, the recognition results are stored in the recognition result memory. 1. A candidate string creation method, characterized in that candidate strings are sorted in descending order of accuracy after the candidate string position.