JPH08146988A - Method and device for speech recognition - Google Patents

Abstract

PURPOSE: To improve the response from speech input to recognition result output without causing deterioration in recognition performance and to increase the number of categories without causing deterioration in response by using a specified representative pattern as retrieval key information and extracting other representative patterns to be narrowed down from a memory. CONSTITUTION: An input speech inputted from a speech input device 60 is digitized by a feature quantity extractor 61 and analyzed as specified to extract its speech feature quantity. A collator 11 collates the speech feature quantity with all representative patterns in a meaning dictionary 63 to select the representative pattern which matches the speech feature quantity most. Then a representative pattern mutual order table 10 is retrieved by using the representative pattern as retrieval key information and a representative pattern which is stored while related to the storage field of index information on the representative pattern is selected. Then the speech feature quantity of the input speech is collated with category patterns subordinate to the respective representative patterns and the category pattern which has the largest degree of matching is extracted and outputted as a recognition result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to speech recognition technology,
More specifically, the present invention relates to a technique for narrowing down representative patterns in a vocabulary dictionary used for matching with an input voice feature amount.

[0002]

2. Description of the Related Art FIG. 6 is a block diagram showing the main part of a conventional general speech recognition apparatus. In this type of voice recognition device, a voice taken in from a voice input device 60 such as a microphone is digitized by, for example, a feature quantity extractor 61, and required analysis is performed to extract the voice feature quantity. The collator 62 collates the voice feature amount sent from the feature amount extractor 61 with the vocabulary (category) pattern stored in the vocabulary dictionary 63, and recognizes the category that best matches the voice feature amount of the input voice. Output as a result.

As described above, since the speech recognition apparatus collates the input speech feature amount with all the category patterns in the vocabulary dictionary 63, when there are many category patterns stored in the vocabulary dictionary 63, the collation processing is performed. It takes a lot of time. Therefore, when creating the vocabulary dictionary 63, all categories are grouped in advance for each similar pattern to define a representative pattern that represents the characteristics of each group, and at the time of matching, first, the input speech feature amount and all the representative patterns are matched. After narrowing down the number of representative patterns according to the degree of matching with the input voice feature amount,
Usually, the vocabulary belonging to the group of the representative pattern having a higher degree of matching is sequentially compared with the input speech feature amount, and the vocabulary having the best match is output as the recognition result.

For example, a case will be described in which all category patterns stored in the vocabulary dictionary 63 are grouped into seven similar pattern groups as shown in FIG. 5, and representative patterns D1 to D7 are defined for each group. To do. In the case of this example, the first pattern “Aichi” holds a plurality of “Aichi” average values and variance values of the feature vectors of the learning voice data as parameters, and the representative pattern D1.
Holds the group average value and variance value of the parameters from "Aichi" to "Adachi", which is the fourth pattern, as its parameters. The same applies to the relationship between each pattern after the fifth pattern and the representative patterns D2 to D7 of each group.

The processing procedure of the collator 62 when the input voice is "Aichi" will now be described with reference to FIG. In addition, S
Indicates each processing step. The collator 62 first detects the voice feature amount of "Aichi" and all the representative patterns D1 in the vocabulary dictionary 63.
~ D7 are collated (S201), sorting (sorting) processing is performed according to the degree of matching (S202), and the top three representative patterns D1, D3, D5 having the highest degree of matching are selected (S203). . After that, each representative pattern D1, D3
D5 subordinate category patterns "Aichi" ~ "Adachi", "Amagasa" ~ "Amami Oshima", "Akita"
~ Collation processing between "Akiho" and the voice feature amount of the input voice "Aichi" is performed (S204), the category pattern "Aichi" having the highest degree of matching is extracted (S205), and it is output as a recognition result (S206). ).

[0006]

As described above, conventionally, the input voice feature amount and all the representative patterns D in the vocabulary dictionary 63 are conventionally used.
When selecting the top three representative patterns D1, D3, D5 with a higher degree of matching from the matching results with 1 to D7, it is common to perform the sorting process shown in S202 of FIG. 7 as a pre-process. It was Therefore, when the number of representative patterns in the vocabulary dictionary 63 is large, the processing time for sorting becomes long and there is a problem that the response from voice input to output of the recognition result is significantly reduced. If the number of categories subordinate to each representative pattern is increased, the number of representative patterns is reduced and the response problem can be solved, but the recognition performance is deteriorated. On the other hand, in order to improve the recognition performance, it is necessary to increase the number of category patterns, which makes the above problem more remarkable.

Therefore, an object of the present invention is to improve the response from voice input to recognition result output without degrading the recognition performance, and to increase the number of categories without lowering the response. It is an object of the present invention to provide a method and a voice recognition device that realizes this method.

[0008]

According to a speech recognition method provided by the present invention, a plurality of categories are grouped for each similar pattern, representative patterns representing the characteristics of each group are stored in a vocabulary dictionary, and input. When collating the voice feature amount with the plurality of categories, after narrowing down a predetermined number of representative patterns having a high degree of matching by collating the input voice feature amount with all the representative patterns stored in the vocabulary dictionary, In the method of comparing the category belonging to the group of these predetermined representative patterns with the input speech feature amount and outputting the category with the highest degree of matching as the recognition result, the feature amount of each representative pattern stored in the vocabulary dictionary. Of the other representative patterns with respect to each other, the index information of the other representative patterns whose matching degree with the one representative pattern exceeds a predetermined value is A first step of storing in a memory in association with index information; a second step of identifying a representative pattern whose feature amount best matches the input voice feature amount and its index information; Using the index information as search key information, the index information of other representative patterns related to the memory is extracted,
The third step of collating the category in the vocabulary dictionary corresponding to the extracted representative pattern index information with the input speech feature amount is included. The index information of the representative pattern is, for example, the address of the representative pattern in the vocabulary dictionary, the feature amount data of the representative pattern, and the like, and any type can be used as long as the representative pattern can be uniquely derived.

In the first step, for example, a step of deriving a degree of matching with another representative pattern when one representative pattern is used as a pseudo voice feature amount, and the derived degree of matching exceeds a predetermined value. Sorting the index information of the other representative pattern in descending order of matching, and storing the index information of the one representative pattern in the memory,
In the third step, a step of extracting index information of other related representative patterns from the memory in descending order of matching degree with the specified representative pattern, and the vocabulary corresponding to the extracted representative pattern index information, respectively. And a step of sequentially matching the category in the dictionary with the input voice feature amount.

Further, the voice recognition device provided by the present invention is
Input voice feature amount extraction means for extracting the feature amount of the input voice to be recognized, a vocabulary dictionary storing representative patterns representing the features of each group by grouping a plurality of categories into similar patterns, and the extracted input A collator that collates the voice feature quantity with the representative patterns and categories in the vocabulary dictionary stepwise and outputs the category with the highest degree of coincidence as a recognition result, and the characteristics of the individual representative patterns stored in the vocabulary dictionary. Based on the mutual matching results of the feature quantities of all other representative patterns with respect to the amount, the index information of the other representative patterns whose matching degree with one representative pattern exceeds a predetermined value is the index of the one representative pattern in order of matching degree. It has a memory stored in association with information. In this configuration, the collator collates the input speech feature amount with the feature amount of the representative pattern stored in the vocabulary dictionary, and specifies the best matching representative pattern and its index information. Second means for extracting index information of another related representative pattern from the memory using the index information of the specified representative pattern as search key information, and a category in the vocabulary dictionary corresponding to the extracted index information of the representative pattern A third means for collating with the input voice feature amount is included.

[0011]

According to the present invention, the feature amounts of all other representative patterns with respect to the feature amounts of the individual representative patterns stored in the vocabulary dictionary are collated with each other in advance, and the degree of matching with one representative pattern has a predetermined value. The index information of the other representative patterns that have been exceeded is stored in the memory in association with the index information of the one representative pattern. At that time, preferably, the degree of matching with one of the other representative patterns when one representative pattern is used as the pseudo voice feature amount is derived, and the index information of the other representative pattern whose derived degree of matching exceeds a predetermined value is matched. It is sorted in descending order of frequency and stored together with the index information of the one representative pattern. By doing so, there is an advantage that the same result as the conventional sorting process can be obtained in a very short time.

At the time of collation, the feature quantity of the input speech to be recognized is extracted by the input speech feature quantity extraction means, and the input speech feature quantity and the vocabulary dictionary are collated by the first means of the collator to obtain the best match. One representative pattern and its index information are specified. After that, the second means of the collator searches the memory using the specified representative pattern as the search key information, and extracts the index information of other related representative patterns.
As described above, when other related representative patterns are stored in the matching degree order, the index information is extracted in that order. The third means of the collator collates a category subordinate to the group of representative patterns in the vocabulary dictionary uniquely derived from the extracted representative pattern index information with the input speech feature amount. As a result, one category with the highest degree of matching is output as the recognition result.

[0013]

Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a configuration diagram of a main part of a voice recognition device according to an embodiment of the present invention, in which elements having the same functions as those of the conventional device shown in FIG. Also, for convenience, the vocabulary dictionary will be described with the contents shown in FIG.
In this embodiment, the representative pattern mutual rank table 10 is provided in the memory, and the configuration of the collator 11 is partially changed accordingly.

The representative pattern mutual rank table 10 is
It is created in advance by the following procedure as a pre-stage of the voice recognition process.

(1) Each representative pattern, for example, the representative pattern D1 of the first group is set as a pseudo voice feature amount, and the degree of matching between this pseudo voice feature amount and all other representative patterns D2 to D7 is calculated. (2) Next, as a result of the calculation, the one whose matching degree exceeds a predetermined value is selected. This predetermined value may be arbitrarily adjusted according to the type or use of the recognition target vocabulary. It should be noted that this selection step can be omitted in some cases. (3) Next, the above calculation results or selection results are sorted in descending order of matching degree, and a pattern in the group in which the matching degree of the representative pattern D1 with the actual input voice feature amount is maximum is sorted. The representative pattern to be collated is stored in the reference column of the index of the representative pattern D1. (4) The above process is repeated for the other representative patterns D2 to D7.

FIG. 2 shows the contents of the representative pattern mutual rank table 10 thus created. In FIG.
Reference numeral 20 is a storage field for storing index information of the representative pattern that maximizes the degree of matching with the input voice feature amount, and 21 is for this field 20.
And is provided for storing index information of upper P (natural number) representative patterns for which subordinate patterns are to be collated, in order of matching degree.

In the upper part of the illustrated example, there are two other representative patterns whose degree of coincidence with the representative pattern D1 of the first group exceeds a predetermined value, and among these representative patterns, the representative pattern of the third group. D3 is the representative pattern D of the 5th group
It shows that the degree of agreement was higher than 5. Further, the content of the next stage shows that, when the representative pattern D2 of the second group is set as the pseudo voice feature amount F, there is only one representative pattern whose matching degree exceeds the predetermined value (representative pattern D3). There is. That is, according to the representative pattern mutual ranking table 10, if one representative pattern can be specified, another representative pattern that matches the specified representative pattern can be immediately specified. In addition, in FIG.
Although the identification information D1 to D7 of the respective representative patterns shown in (4) are used as the index information, the index information is not limited to these identification information as described above.

Next, with reference to FIG. 3, description will be given of the contents of the recognition processing when the input voice is "Aichi" in the voice recognition apparatus of the present embodiment having the above-mentioned configuration. In addition, in FIG.
Indicates each processing step.

The voice of "Aichi" taken in from the voice input device 60 shown in FIG. 1 is digitized by the feature quantity extractor 61, and the required analysis is performed to extract the voice feature quantity. The processing up to this point is similar to that of the conventional apparatus. As shown in FIG. 3, the collator 11 of the present embodiment collates the speech feature amount of "Aichi" with all the representative patterns D1 to D7 in the vocabulary dictionary 63 (S101), and the speech feature amount of "Aichi". A representative pattern D1 that most closely matches with is selected (S102: first means). This processing can be easily realized by a known pattern recognition technique.

After that, the representative pattern mutual rank table 10 is searched by using the representative pattern D1 as search key information, and the representative patterns D3 and D5 stored in association with the storage column 20 of the index information of the representative pattern D1 are selected. (S103: second means). After that, each representative pattern D
Category pattern "Aichi" subordinate to 1, D3, D5 ~
The matching process is performed with "Adachi", "Amagasa" to "Amami Oshima", "Akita" to "Akiho" and the voice feature amount of the input voice "Aichi" (S104: third means), and the highest degree of matching is obtained. Extract the category pattern "Aichi" (S1
05) It is output as a recognition result (S106).

FIG. 4 is an actual measurement chart showing the degree of coincidence between the result of selecting the top P representative patterns according to this embodiment in the word recognition experiment and the result of selecting the top P'representative patterns in the conventional sorting process. . The dictionary structure in this case is
Number of vocabulary patterns m = 2,000, number of representative patterns Q = 64
Has become. In the figure, the vertical axis represents the degree of coincidence [%], and the horizontal axis represents the number k [pieces] of vocabulary patterns to be checked after the representative pattern is selected.

As shown in FIG. 4, when P representative upper representative patterns are selected so that the number k of patterns of the vocabulary to be collated is almost 600, the upper representative patterns by the conventional method and the method of this embodiment are selected. The degree of agreement was about 75%. In addition, the recognition performance at this time was 89% in both of the speech recognition devices by both methods, and there was no difference.

As described above, in the speech recognition apparatus of this embodiment, the representative pattern mutual rank table 10 is searched in place of the conventional sorting processing, so that the P high-ranking items having a high degree of matching with the input speech feature amount are searched. Since the representative pattern is selected, it is possible to reduce the calculation amount for the matching process without degrading the recognition performance. Moreover, since the representative pattern mutual rank table 10 can be used as an index of the vocabulary dictionary 63 as described above, the response of the voice recognition process itself does not change even if a large number of vocabulary and representative patterns are stored in the vocabulary dictionary 63. The recognition performance can be expected to improve significantly. Further, even if the vocabulary and the representative pattern are additionally stored in the vocabulary dictionary 63 afterwards, it can be dealt with only by partially changing the representative pattern mutual rank table 10, so that there is versatility, which is advantageous in terms of operation cost.

[0024]

As is apparent from the above description, according to the voice recognition method of the present invention, after the one representative pattern that most matches the input voice feature amount is specified, the specified representative pattern is stored as search key information in the memory. Since other representative patterns to be narrowed down are extracted from, the response from voice input to recognition result output can be improved without degrading the recognition performance. In addition, the number of category patterns can be increased without lowering the response, which can also contribute to improvement in recognition performance.

Further, according to the speech recognition apparatus of the present invention, the information to be stored in the memory can be arbitrarily changed according to the number of category patterns and the number of representative patterns stored in the vocabulary dictionary without changing the structure of the collator. Since it is possible to improve the response from voice input to recognition result output and improve the recognition performance, the number of category patterns can be arbitrarily changed according to the application, and a more versatile device configuration can be achieved. Can be realized.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a main part of a voice recognition device according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of contents of a representative pattern mutual rank table included in the voice recognition device according to the present embodiment.

FIG. 3 is a schematic procedure diagram of voice recognition processing according to the present embodiment.

FIG. 4 is an actual measurement diagram showing the degree of coincidence between the representative pattern selection result according to the present example in the word recognition experiment and the representative pattern selection result in the conventional sorting process.

FIG. 5 is an explanatory diagram showing a configuration example of a vocabulary dictionary used in this type of voice recognition device.

FIG. 6 is a configuration diagram of a main part of a conventional voice recognition device.

FIG. 7 is a schematic procedure diagram of conventional speech recognition processing.

[Explanation of symbols]

 10 mutual pattern mutual rank table 11,62 collator 60 voice input device such as microphone 61 feature amount extractor 63 for extracting feature amount of input voice 63 vocabulary dictionary storing a plurality of representative patterns

Claims (3)

[Claims] 1. A plurality of vocabularies are grouped for each similar pattern, representative patterns representing the characteristics of each group are stored in a vocabulary dictionary, and when an input speech feature amount and the plurality of vocabularies are collated. , The input speech feature amount is compared with all the representative patterns stored in the vocabulary dictionary to narrow down a predetermined number of representative patterns having a high degree of matching, and the vocabulary belonging to the group of these predetermined representative patterns and the input In a voice recognition method of collating with a voice feature amount and outputting the vocabulary with the highest degree of matching as a recognition result, the feature amounts of all other representative patterns with respect to the feature amounts of the individual representative patterns stored in the vocabulary dictionary are Match each other,
A first step of storing index information of another representative pattern whose degree of matching with one representative pattern exceeds a predetermined value in a memory in association with the index information of the one representative pattern, and A second step of identifying a representative pattern that best matches the input voice feature amount and its index information, and extracting index information of other relevant representative patterns from the memory using the index information of the identified representative pattern as search key information. And a third step of collating the vocabulary in the vocabulary dictionary corresponding to the extracted representative pattern index information with the input speech feature amount, the speech recognition method. 2. The first step is a step of deriving a degree of agreement with another representative pattern when one representative pattern is used as a pseudo voice feature amount, and a step of deriving the degree of agreement that has been derived exceeds a predetermined value. Sorting the representative pattern index information in descending order of the degree of matching and storing the same in the memory together with the index information of the one representative pattern, and the third step includes another related representative from the memory. Extracting the index information of the patterns in descending order of the degree of matching with the specified representative pattern, and sequentially collating the vocabulary in the vocabulary dictionary corresponding to the extracted representative pattern index information with the input speech feature amount. The method according to claim 1, further comprising: 3. An input voice feature amount extraction means for extracting a feature amount of an input voice to be recognized, a vocabulary dictionary storing a representative pattern representing a feature of each group by grouping a plurality of vocabularies into similar patterns, And a collator that collates the extracted input speech feature amount with the representative pattern and the vocabulary in the vocabulary dictionary stepwise and outputs the vocabulary having the highest degree of coincidence as a recognition result. Index information of other representative patterns whose matching degree with one representative pattern exceeds a predetermined value, based on the mutual collation result of the characteristic amounts of all other representative patterns with respect to the characteristic amounts of each representative pattern stored in the dictionary. Is provided in association with the index information of the one representative pattern in the order of the degree of matching, and the collator is configured to store the input voice feature quantity and the representative pattern stored in the vocabulary dictionary. Means for identifying the most representative representative pattern and its index information by collating with the feature amount of the search target and index information of the other representative pattern related from the memory using the index information of the identified representative pattern as search key information. It is configured to include a second means for extracting information, and a third means for collating the vocabulary in the vocabulary dictionary corresponding to the extracted representative pattern index information with the input speech feature amount. And a voice recognition device.