JP3092384B2 - Pattern matching device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition device.

[0002]

2. Description of the Related Art FIG. 11 is a block diagram showing a configuration of a conventional pattern matching apparatus described in, for example, "Digital Speech Processing" by Sadateru Furui, p160-161 (Tokai University Press).

In FIG. 1, reference numeral 61 denotes acoustic analysis means for acoustically analyzing a voice signal 66 at regular time intervals and converts it into characteristic parameters 67; 62, a standard pattern storage means for storing a standard pattern; Each time is output, a pattern matching unit that performs pattern matching with the standard pattern 68 of the standard pattern storage unit 62 and updates the cumulative distance value of each candidate stored in the distance value storage unit 64. This is a result output means for extracting and outputting the collation result 70 from the accumulated distance value stored in the means 64. 69a and 69b are cumulative distance values.

Hereinafter, a conventional pattern matching apparatus will be described as an example of a pattern matching apparatus that performs pattern matching by a dynamic programming method. FIG. 12 is a flowchart showing the processing of the pattern matching means in the conventional pattern matching device.
It is assumed that the voice input starts at time 1 and ends at time T.

In FIG. 12, t is time, j is the frame number of the standard pattern, n is the candidate number, J (n) is the number of frames of the standard pattern of candidate n, N is the number of candidates to be recognized, d
(n, t, j) is a local distance value of the frame j of the standard pattern of the candidate n at the time t, and is calculated using, for example, the Euclidean distance. G (n, t, j) is a cumulative distance value to the frame j of the standard pattern of the candidate n at the time t.

Prior to the processing, the standard pattern storage means 62 stores standard patterns of all candidates.
The pattern matching process for N candidates starts from step 101 and ends at step 114. First, step 102
In, initialization of the distance value storage means 64 is performed. FIG.
FIG. 9 is a diagram showing a state in which the distance value storage means 64 is initialized for a candidate n. In the figure, the symbol "-" indicates that the item has not been initialized and the value has not been determined.

In steps 103 to 113, processing relating to the frame t of the input pattern is shown. In the pattern matching, the pattern matching for all the candidates is repeated every time one frame of the input pattern is input, and the cumulative distance value g (n, t, j) in the distance value storage unit 64 is updated. In steps 104 to 111, the processing regarding the candidate n is described. In steps 105 to 109, processing for the frame j of the standard pattern of each candidate is described.

In step 106, the j-th frame of the standard pattern of the candidate n is read from the standard pattern storage means 62, and the local distance value d (n, t, j) at time t is calculated. Using the local distance value, a calculation by the dynamic programming is performed in step 107, and the cumulative distance value g (n,
t, j) is updated. At the time T when the voice input ends, the result output unit 65 outputs g (n, T, J (n)) as the pattern matching result 70 of the candidate n. It can be determined that the input audio signal is more likely to be in the order of smaller candidates in the pattern matching result 70.

[0009] The conventional pattern matching device is configured as described above, and the amount of calculation increases as the number of candidates increases. In particular, in the case of performing continuous speech recognition, when it is desired to obtain a matching result of a plurality of similar candidates, the number of candidates increases due to the combination of the standard patterns, so that there is a problem that the amount of calculation becomes extremely large. For example, in the case of continuous speech recognition such as "I go to school", standard sentences of words such as "I", "ha", "school", "to", "go" are connected to represent sentence candidates. I do. FIG. 4 shows an example in which a sentence to be recognized is represented by a rule for connecting words in this manner. FIG. 5 shows an example of a sentence candidate that can be represented by this rule.

In FIG. 5, if candidate 1 is uttered, candidate 2 is an example in which "ha" is dropped, candidate 3
Can be regarded as an example in which “ha” has been replaced with “mo”, and candidate 4 can be regarded as an example in which “tachi” has been inserted. In this way, a plurality of portions composed of a common standard pattern sequence such as "I" and "go to school" are generated by insertion, substitution, and omission. For a candidate whose first half consists of a sequence of common standard patterns, such as "I", recognition is left
In the case of proceeding to t-to-Right, pattern matching for the "I" part can be omitted because it is common to all candidates and the same cumulative distance value can be used.

For the candidate whose second half is composed of a series of common standard patterns "go to school", the same standard pattern is calculated. Are performed repeatedly. Here, a description will be given taking pattern matching of candidate 1 and candidate 2 as an example.

FIG. 6 schematically shows the contents of the distance value storage means in the case where the pattern matching between the candidate 1 and the candidate 2 shown in FIG. 5 is performed. In the figure, the vertical axis indicates the standard pattern, and the horizontal axis indicates the passage of time t. The solid line in the figure indicates the optimal path. The optimal path indicates an optimal combination of each frame of the input pattern and each frame of the standard pattern obtained by the dynamic programming. Since the standard pattern of "I" is used in common at the beginning of candidate 1 and candidate 2, pattern matching is the same operation, and the optimal path is the same until time t1. Since "ha" exists in the candidate 1, different pattern matching is performed from the time t1 to the time t2, and a different optimal path is followed. However, after time t2, the same optimal path is followed again. Therefore, after time t2, the difference between the accumulated distance values on the optimal path is constant. Only one of the pattern comparisons is performed, and the accumulated distance value of the remaining candidates can be obtained from this difference at time T.

[0013]

However, since the optimal path in the dynamic programming is found only after backtracking after all the pattern matchings are completed, it is detected that the optimal path follows the same path during the pattern matching. There is no means. Therefore, there has been a problem that the above-described redundant pattern matching calculation has to be performed.

As a method for avoiding such a problem, Shwa
rtz, A Comparizon of Several Apoximate Algorithm
s For Finding Multiple (N-BEST) Sentence Hypothse
s "(ICASSP-91.pp701-704) and Takao Watanabe," High-speed Method for Continuous Speech Recognition by Bundle Search "(Acoustic Society of Japan Spring 1991)
Bu, which is described in the season Research Workshop Proceedings p125-126)
An ndle search and the like have been proposed. However, these algorithms require a fairly large memory area for backtracking, and may not correctly calculate the lower-order pattern matching results, such as the Lattice N-Best algorithm or bundle search. There is.

A pattern matching device according to the present invention has been made in order to solve the above-described problems, and it is possible to obtain a pattern matching device having a small amount of calculation and a small amount of memory without deteriorating pattern matching performance. With the goal.

[0016]

Means for solving the problems Acoustic analysis means for acoustically analyzing a voice signal at regular time intervals and converting it into characteristic parameters,
A matching instruction means for storing matching candidates for performing pattern matching, and each time a characteristic parameter is input from the acoustic analysis means, a standard pattern corresponding to the candidate specified by the matching instruction means is referenced from the standard pattern storing means. Pattern matching means for performing pattern matching and calculating a cumulative distance value, distance value storing means for storing the cumulative distance value of each candidate based on the calculated cumulative distance value, and a latter half of the same standard pattern. If the cumulative distance value of the corresponding part in the latter half of the candidate is kept within a predetermined distance for a predetermined time width, these are detected as integrated candidates, and one of these integrated candidates is detected. Integrated candidate detecting means for removing the remaining integrated candidates from the matching target in the matching instruction means, and detecting the remaining integrated candidates by the integrated candidate detecting means. Integrated candidate storage means for storing the combination of the selected integrated candidates and the difference between the cumulative distance values thereof, and the distance value of the candidate removed by the matching instruction means, the cumulative distance value of the unremoved candidate, And a result output means for determining and outputting a collation result based on the distance value of each candidate obtained from the difference between the accumulated distance values of the candidate storage means.

In the integrated candidate detecting means, a difference between the minimum values of the cumulative distance values of the latter part is the same again after a certain time interval between the candidates composed of the same standard pattern in the latter part. First, these candidates are integrated candidates.

Further, best-f
By performing irst-search, the top N items are output.

[0019]

In the present invention, candidates having the same accumulated distance value are detected as integrated candidates in the latter half by the integrated candidate storage means, and only one candidate is left as a representative candidate.
For the other candidates , only the relative distance difference is stored in the integrated candidate storage means, and the distance value of the candidate whose pattern matching has been stopped is finally stored in the integrated candidate storage means as the cumulative distance value of the representative candidate. Since the difference is obtained by adding the difference between the distance values, the same pattern matching can be stopped, and the amount of calculation of the overlapping pattern calculation for the same standard pattern can be reduced.

[0020] The integrated candidate detecting means may further include a case where the difference between the minimum values of the cumulative distance values of the latter part is the same again after a certain time interval between the candidates composed of the same standard pattern in the latter part. By making these candidates as integrated candidates, the integrated candidate detection process is simplified.

Further, best-f
By performing the irst-search, the top N candidates having the small top N distance values are output.

[0022]

【Example】

Embodiment 1 FIG. Hereinafter, an embodiment of the invention will be described with reference to FIG. In the figure, reference numeral 31 denotes sound analysis means for sound-analyzing the audio signal 39 at fixed time intervals and converts it into characteristic parameters; 32, standard pattern storage means for storing standard patterns; and 33, distance value storage for storing the cumulative distance value of each candidate. Means 34 is a matching instruction means for storing candidates for pattern matching.

Each time a characteristic parameter 40 is input from the acoustic analysis means 31, the reference numeral 35 refers to the standard pattern 41 of the standard pattern storage means 32 and is designated by a collation instruction 42 from the collation instruction means 34. This is a pattern matching unit that performs pattern matching with a candidate and outputs a cumulative distance value 43a to the distance value storage unit 33 for updating.

The reference numeral 36 designates an integrated candidate in the case where the corresponding cumulative distance value 43c in the latter half maintains the same difference for a certain time width between candidates having the same standard pattern in the latter half. , And selects one of the integration candidates as the one to be left, and removes the remaining integration candidates from the matching target in the matching instruction means 34 by the removal instruction 44.

Reference numeral 37 denotes an integrated candidate storage means for storing the difference between the combination of the integration candidates detected by the integration candidate detection means 36 and the cumulative distance value, and 38 denotes the cumulative distance value of the candidate removed by the collation instruction means 34. Is finally determined from the difference between the cumulative distance value 43b of the candidate remaining in the matching instruction means 34 and the cumulative distance value between the integrated candidates stored in the integrated candidate storage means 37 to determine the matching result 46 of each candidate. And a result output means for outputting the result.

Hereinafter, a pattern matching apparatus according to the present invention will be described with an example of a pattern matching apparatus that performs pattern matching by a dynamic programming method as in the conventional pattern matching apparatus. FIG. 2 is a flowchart showing the processing of the pattern matching means in the pattern matching device of the present invention. In FIG. 2, t, j, n, T, J (n), N, d (n, t, j), g (n,
Each symbol of (t, j) has the same meaning as that used in the description of the conventional pattern matching device.

M (n) is the collation instruction 42 of the collation instruction means 34
M (n) = 1 indicates that the pattern matching for the candidate n is designated, and m (n) = 0 indicates that the pattern matching is not performed, that is, it is removed from the matching target. Prior to the processing, it is assumed that standard patterns of all candidates are stored in the standard pattern storage unit 32. The pattern matching process for N candidates starts from step 1 and ends at step 18. First, in step 2, as shown in FIG. 3, the storage contents of the distance value storage means 33 are initialized, and in step 3, the initialization is performed in the collation instruction means 34 so that all m (n) is set to 1. It is set so that all candidates are subjected to pattern matching.

Steps 4 to 17 show an operation in which the processing for the frame t of the input pattern corresponding to the time t is performed until the time T when the voice input ends. Among them, every time one frame of the input pattern is input, pattern matching of the one frame with the standard patterns of all the candidates is performed, and the cumulative distance value g (n, t, j) of the distance value storage means 33 for each candidate is checked. ) Will be updated.

The processing from step 8 to step 11 is performed for a certain candidate n for the frame t input at the time t.
2 shows a process of performing pattern matching with the standard pattern of No. and updating the cumulative distance value g (n, t, j) for the candidate. That is, in step 8, the j-th frame of the standard pattern of the candidate n is read from the standard pattern storage unit 32, and the input frame t and the candidate n
Local distance value d (n,
t, j), and based on this, the calculation by the dynamic programming is performed in step 9 to obtain the cumulative distance value g (n, t, j), and the corresponding value of the distance value storage means 33 shown in FIG. Update the contents of the location.

In step 10, j is increased, and the processing from step 8 is repeated until it is determined in step 11 that the processing has been performed the same number of times as the number of all frames of the standard pattern of the candidate n. Thus, the frame t input at the time t
Is compared with all frames of the standard pattern of the candidate n, and the cumulative distance value g (n, t,
j) is updated (one vertical column corresponding to t in FIG. 3 is updated).

After that, detection and processing of integration candidates consisting of steps 12 and 13 are performed, which will be described later. By repeating the processing from step 8 to step 13 while increasing the candidate number n in step 14, the same processing is performed for another candidate, and it is determined in step 15 that the processing has been performed for all candidate numbers N. Repeat until As a result, the matching of the frame at the time t with respect to all the candidates is performed. That is, the storage contents as shown in FIG. 3 are set in the distance value storage means 33 for each candidate, and the cumulative distance values g (n, t, j) for all the candidates are updated.

Here, in step 12, an integration candidate is detected by the integration candidate detection means 36. This is to detect candidates that can be integrated, with the latter half being composed of the same pattern sequence. When an integrated candidate is detected in step 12, if the candidate is not a representative candidate representing the integrated candidate in step 13, the matching instruction contents m (n) of the candidate in the matching instruction means 34
Is set to 0 to remove the candidate from the collation target, and store the combination of the integration candidates and the difference between their cumulative distance values in the integration candidate storage means 37.

In the process of updating the cumulative distance value g (n, t, j) in each of the candidates, the collation instruction content m (n)
Is set to 0, it is determined in step 6 that the candidate has been removed from the collation target.
The pattern matching calculation up to 3 is not performed. As a result, the amount of calculation of the latter half part which is duplicated among a plurality of candidates can be reduced.

The detection of the integration candidate by the integration candidate detection means 36 in the above-mentioned step 13 is performed as follows. For example, as shown in FIG.
Frame js (2) in the latter half of (1) and candidate 2
It is assumed that the latter half is expressed by the same standard pattern. In other words, the frames from candidate frames js (1) to J (1) and the candidate 2 from frames js (2) to J (2) are the candidates for the integrated candidate detection.

The frame length of this portion is defined as Jw. FIG. 7 shows the flow of the integration candidate detection process at time t3. In steps 73 to 76, corresponding portions of candidate 1 and candidate 2 (from js (1) to J (1) of candidate 1 and candidate 2)
The difference e (t, j) between the accumulated distance values g (n, t, j) of js (2) to J (2) is calculated. This is calculated retroactively from t3, and is determined until it is determined in steps 77 and 78 that it has been retroactive from the time t3 by a predetermined time width Tw.

If the obtained values of e (t, j) are all the same during a predetermined time width Tw, the candidates 1 and 2 are regarded as integrated candidates. Then, for example, candidate 1 is left as a representative candidate, and candidate 2 is deleted from collation instruction means 34. That is, a removal instruction 44 is issued from the integration candidate detection means 36 to the collation instruction means 34, m (2) = 0 is set in the collation instruction means 34, and the candidate number 1 of the representative candidate and the deleted candidate number 2 Cumulative distance value difference d
1 is stored in the integration candidate storage means 37. Note that d1 = e
(t, j).

FIG. 8 schematically shows the storage contents of the integration candidate storage means 37. Each row indicates a combination of integration candidates, the first column indicates a representative candidate, the second column indicates a removed candidate, and the third column indicates a cumulative distance difference. In FIG. 8, the combination of the combined candidates of candidate 1 and candidate 2 obtained in the above description is included in the first row, and in this example, the combination is sequentially stacked upward from the sixth row input earlier. It is shown in the format.

When the time T at which the voice input ends is reached,
For the candidate 1 not removed by the collation instruction means 34 in the result output means 38, the cumulative distance value 43 (b) from the distance value storage means 33 is referred to, and the cumulative distance value g (1, T, J
Let (1)) be the final distance value of candidate 1. For the candidate 2 removed by the collation instruction means 34, the candidate number 1 of the representative candidate for the candidate 2 is obtained from the integrated candidate storage means 37, and the cumulative distance value 43b of the candidate 1 obtained from the distance value storage means 33; The difference d1 of the cumulative distance value of the candidate 2 with respect to the representative candidate 1 stored in the integrated candidate storage means 37 is added to obtain the final distance value of the candidate 2 as g (1, T, J (1)) + d1. Is determined.

For example, when the integrated candidate storage means 37 has the contents as shown in FIG. 8, the distance value of the candidate 3 is g (1, T, J (1)) + d2 and the distance value of the candidate 4 is g (1 , T, J (1)) + d1 + d3 Distance value of candidate 5 is g (1, T, J (1)) + d1 + d4 Distance value of candidate 6 is g (1, T, J (1)) + d2 + d5 Distance value of candidate 7 Is obtained as g (1, T, J (1)) + d2 + d6.

In the input audio signal, it can be determined that a candidate having a smaller final distance value thus obtained is more likely.

In the above description of the integrated candidate detecting process, the difference e (t, j) between the cumulative distance values g (n, t, j) of the candidate 1 and the candidate 2 is calculated every time t advances. Although the calculation has been performed, the difference e (t, j) may be stored and used at the next time.

In addition, the case where the difference e (t, j) is the same value during the predetermined time width Tw has been described as the judgment of the integration candidate. It may be assumed that there is. As above, the latter half is the same
By setting the candidates as integrated candidates, duplicate operations in the latter half of the same operation result are reduced.

It is preferable that the detection of the integration candidate be performed every frame every time the characteristic parameter is output from the acoustic analysis means 31, but the integrated detection processing in the integration candidate detection means 36 is output from the acoustic analysis means 31. It may be thinned out once for several frames of the feature parameter to be performed. In this case, the same effect can be obtained.

The pattern matching apparatus according to the present invention can be used in combination with the beam search method, in which case the effect of reducing the amount of calculation can be further increased. Here, as described above, in the pattern matching device according to the present invention, the same effect can be obtained even if any of the candidates detected as integrated candidates is set as the representative candidate. However, when used together with the beam search, a better pattern matching device can be obtained by performing pattern matching while leaving the smaller distance value as a representative candidate so as not to be censored.

In the above description, the case where all the accumulated distance values g (n, t, j) for the width of the time T are stored in the distance value storage means 33 has been described. Only a capacity for storing the accumulated distance value corresponding to one frame of the input feature parameter can be prepared, and pattern matching can be performed by repeatedly using the capacity, and the same effect can be obtained even with a small capacity. Similarly, a table e (t, j) for calculating the difference between the accumulated distance values for detecting the integration candidate may be prepared for the width Tw and used repeatedly, and the memory amount can be reduced.

Embodiment 2 FIG. An embodiment of the invention according to claim 2 will be described. The configuration diagram of the pattern matching device according to the present invention is the same as that of FIG.
The other components have the same functions as those of the first embodiment, and a description thereof will not be repeated.

In the first embodiment, the processing amount of the integration candidate detecting means 36 is large, but in the second embodiment, means for detecting the integration candidate using a simpler method is shown. Here, detection of an integration candidate is performed as follows. As in the first embodiment, as shown in FIG. 6, when the second half of the candidate 1 after the frame js (1) and the second half of the candidate 2 after the frame js (2) are expressed by the same standard pattern sequence. I do.

First, for each candidate n, the cumulative distance value g (n, t, t) at the time t in the latter half part js (n) to J (n) is obtained.
The minimum cumulative distance value gmin (n, t) which is the minimum value of j) is obtained.

Gmin (n, t) = min (g (n, t, j)), j
Is from js (n) to J (n)

The minimum cumulative distance value g between candidates in which the latter half is composed of the same standard pattern
If the difference e of min (n, t) is equal at intervals of the time width Tw, it is determined that these candidates are integration candidates. That is, the following conditions are satisfied for candidate 1 and candidate 2.

Gmin (1, t) -gmin (2, t) = gmin (1, t)
t-Tw) -gmin (2, t-Tw)

At this time, the difference e between the minimum cumulative distance values is determined as the difference between the cumulative distance values of candidate 1 and candidate 2. According to this method, the same effect as in the first embodiment can be obtained, and the processing amount of the integration candidate detecting unit 36 can be reduced.

It is desirable that the integration candidate is detected every frame every time the characteristic parameter is output from the acoustic analysis means. However, the integration detection processing may be thinned out once every several frames. To play. Further, it can be used together with the beam search method, and the effect of reducing the amount of calculation can be further increased. Further, the same effect can be obtained by selecting any of the candidates detected as the integration candidates as the representative candidate. But,
When used together with the beam search, a better pattern matching device can be obtained by performing pattern matching while leaving smaller distance values as representative candidates so as not to be censored.

As in the first embodiment, it is possible to prepare the capacity of the distance value storage means 33 for one frame and perform pattern matching, and the same effect is obtained. In addition, a table gmin (n, t) for calculating the difference between the accumulated distance values for detecting the integration candidate may be actually prepared for the width Tw and used repeatedly.
The required memory amount can be reduced.

Embodiment 3 FIG. A third embodiment of the present invention will be described. The configuration diagram of the pattern matching device according to the present invention is the same as that of FIG. 1, and the other components except for the result output means 38 have the same functions as those of the first embodiment, so that the description is omitted here.

Here, a method for acquiring N candidates with a small final distance value from the contents of the integrated candidate storage means 37 using the best-first-search method will be described. After the voice input is completed, it is assumed that the result as shown in FIG. 8 is obtained in the integration candidate storage means 37, and that the differences d1 to d6 of the accumulated distance values are as shown in FIG.

First, a stack area is prepared. Hereinafter, the operation of the stack will be described with reference to FIG. First, candidates not removed by the collation instruction means 34 are put on a stack and sorted by distance value (number of repetitions: 0). Next, the candidate having the smallest distance value is extracted from the stack. In this case, since there is only one candidate, candidate 1 is extracted. Next, candidate 2 which is a candidate with candidate 1 as a representative candidate
As shown in the first embodiment, the distance values are obtained from the differences d1 and d2 of the cumulative distance values, and the candidates are put into the stack in ascending order and sorted by the distance values (the number of repetitions 1).

Next, the candidate 2 having the smallest distance value in the stack is taken out of the stack. Similarly, the same process is repeated for a candidate whose candidate is candidate 2. By repeating such processing N times, N candidates with small distance values can be obtained.

The same effect can be obtained even when there are a plurality of candidates which have not been removed by the collation instruction means 34.

[0060]

As described above, according to the present invention, candidates having the same cumulative distance value in the latter half are detected as integrated candidates by the integrated candidate detecting means, and only one candidate is left as a representative candidate , while other candidates are left as representative candidates. Is stored only in the integrated candidate storage means, and the distance value of the candidate for which pattern matching has been stopped is finally calculated as the cumulative distance value of the representative candidate by the cumulative distance value of the integrated candidate storage means. Since it is obtained by addition, the same pattern collation can be stopped, the amount of operation of the duplicate pattern operation for the same standard pattern can be reduced, and an effective device can be obtained.

In addition, the integrated candidate detecting means may use the same standard pattern in the latter half part for the same pattern again if the difference in the minimum value of the accumulated distance value in the latter half part is the same again after a certain time interval. By making these candidates as integrated candidates, the integrated candidate detection process is simplified.

Further, best-f
By performing the irst-search, the top N candidates having small top N distance values are easily output.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing a pattern matching device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing the operation of the pattern matching device according to one embodiment of the present invention.

FIG. 3 is an explanatory diagram showing storage contents of a distance value storage means of an embodiment of the present invention and a conventional pattern matching device.

FIG. 4 is an explanatory diagram showing an example of expressing a sentence by connecting words;

FIG. 5 is an explanatory diagram showing an example of a sentence candidate.

FIG. 6 is an explanatory diagram showing an optimal path.

FIG. 7 is a flowchart showing the operation of the integrated candidate detecting means of the pattern matching device according to one embodiment of the present invention.

FIG. 8 is an explanatory diagram showing storage contents of an integrated candidate storage means of the pattern matching device according to one embodiment of the present invention.

FIG. 9 is an explanatory diagram showing a difference between accumulated distance values stored in an integrated candidate storage unit corresponding to each candidate;

FIG. 10 is an explanatory diagram illustrating an operation of a result output unit of the pattern matching device according to the third embodiment of the present invention.

FIG. 11 is a configuration diagram showing a conventional pattern matching device.

FIG. 12 is a flowchart showing the operation of a conventional pattern matching device.

[Explanation of symbols]

 Reference Signs List 31 acoustic analysis means 32 standard pattern storage means 33 distance value storage means 34 collation instruction means 35 pattern collation means 36 integrated candidate detection means 37 integrated candidate storage means 38 result output means 39 voice signal 40 feature parameter 41 standard pattern 42 collation instruction 43a 43b, 43c Cumulative distance value 44 Removal instruction 45a, 45b Integration candidate information 46 Matching result

──────────────────────────────────────────────────の Continued on the front page (58) Fields surveyed (Int. Cl. ⁷ , DB name) G10L 15/08 G10L 15/10 G10L 15/12 G10L 15/14 JICST file (JOIS)

Claims (3)

(57) [Claims] 1. A pattern matching apparatus for performing pattern matching of a recognition target candidate obtained by connecting a pre-stored standard pattern with an externally input audio signal, wherein the audio signal is subjected to acoustic analysis at regular intervals. Acoustic analysis means for converting to a feature parameter, standard pattern storage means for storing a standard pattern, matching instruction means for storing candidates to be matched for pattern matching, and each time a feature parameter is input from the sound analysis means A pattern matching unit that calculates a cumulative distance value by performing pattern matching by referring to a standard pattern corresponding to a candidate designated by the matching instruction unit from the standard pattern storage unit; Distance value storage means for storing the cumulative distance value of each candidate based on the distance value; If the cumulative distance value of the corresponding part in the latter half part keeps a difference within a predetermined value for a predetermined time width,
Detect these as an integrated candidate, leaving one of these integration candidates, and integrated candidate detecting means for removing the remaining integrated candidate from the verification target in the comparison instruction means, the integration <br/> case candidate detecting means The integrated candidate storage means for storing the combination of the integrated candidates detected in and the difference of the cumulative distance value thereof, and the distance value of the candidate removed by the collation instruction means, the cumulative distance value of the unremoved candidate, And a result output means for determining and outputting a matching result based on the distance value of each candidate obtained from the difference between the accumulated distance values of the integrated candidate storage means. 2. The integrated candidate detecting means according to claim 1, wherein said second half portion has a difference in the minimum value of the accumulated distance value in the second half portion again with a certain time width between candidates having the same standard pattern. 2. The pattern matching apparatus according to claim 1, wherein in such a case, these candidates are regarded as integration candidates. 3. The best-first-sea method in the result output means.
The pattern matching apparatus according to claim 1 or 2, wherein the N top candidates having a small distance value are output by performing rch.