JPH04280299A - Speech recognition device - Google Patents

Abstract

PURPOSE:To lighten an user's burden concerning the registration of a specific dictionary by automatically performing the registration of the specific dictionary upon recognition operation. CONSTITUTION:The feature level of the voice to be input upon recognizing operation is temporality stored in a disctionary buffer 10, and its word number is stored in a word number buffer 11 refleting it to the voice stored in the dictionary buffer 10 by pushing a key in a recogniged result inputting part 12 when the recognized result to the voice is correct. When the same word numbers are stored in no less than the predetermined number (Th1) in the word number buffer 11, the similarity measure of respective characteristic quantities together with the voice stored in the dictionary buffer 10 corresponding to the word number is calculated, and subsequently either feature level out of them stored in the dictionary buffer 10 is stored, as a specific dictionary, in a specific dictionary part 4, when the similarity measure is greater than the predetermined threshold value (Th2).

Description

[Detailed description of the invention]

0001

[Technical Field] The present invention relates to a speech recognition device, more specifically,
This invention relates to improving the recognition rate of a speaker-independent speech recognition device.

0002

2. Description of the Related Art Speech recognition devices are being put into practical use as input devices for various control devices, personal computers, word processors, and the like. Generally, these devices mainly recognize discretely uttered words that represent control commands or control objects. Speech recognition devices that implement these functions can be roughly divided into speaker-specific and speaker-independent systems. The specific speaker method is a method in which the user using the recognition device registers the word speech to be recognized in his or her own voice, and the selection of vocabulary is also free to some extent. Furthermore, since the recognition dictionary can be configured with almost the same voice as the voice input during recognition, a relatively high recognition rate can be obtained. However, this method requires the user's voice registration, and as the number of words increases, the burden increases. On the other hand, the speaker-independent method is a method in which a speech dictionary of words to be recognized is prepared in advance by the provider (manufacturer) of the recognition device. Since the user uses this dictionary for recognition, there is no need for voice registration. Manufacturers record a large number of voices and compose a dictionary based on the average voice. However, with this method, it is difficult to obtain a relatively high recognition rate due to speaker-specific habits and dialects. A method that combines the ease of use of a non-specific speech recognition device, which requires no registration effort, with the high recognition rate of a specific speaker speech recognition device, is to have both a non-specific dictionary and a specific dictionary. was considered. The above method is basically a speaker-independent speech recognition device, but by selecting a word that is difficult to recognize while the user is using it and registering it as a speaker-specific dictionary, it is possible to recognize the word. The idea is to increase the rate.

FIG. 2 is a block diagram illustrating an example of a conventional speech recognition device as described above. During recognition, feature quantities of input speech are extracted by a feature extraction unit 1, and similarity calculation unit 2 extracts features of input speech. sent to. On the other hand, the feature amounts registered in the unspecified dictionary 3 and the specific dictionary 4 are sequentially sent to the similarity calculation unit 2, where the similarity with the previous input voice is calculated. Once the similarity calculations have been completed, the one with the highest similarity is output from the result output unit 5 as the recognition result. When the recognition device is first used, no words are registered in the specific dictionary 4, and the user can register words that are difficult to recognize in the specific dictionary 4. This can be done by setting the control unit 7 to the registration mode by inputting the key switch 6, specifying the number of the word to be registered, and then uttering the word to register it in the specific dictionary 4. Specific examples include, for example, Japanese Patent Application No. 1-261097.
No., Japanese Patent Application No. 1-261098, Japanese Patent Application No. 1-28679
0 etc., so the explanation here will be omitted.

[0004] As mentioned above, the recognition rate of speech registered in a specific dictionary has been improved. By inheriting the word number and other attributes (such as the actions that occur upon recognition) from the corresponding word in the unspecified dictionary, it can be used without any problems. By this method, it is possible to obtain a better speech recognition device that has both the ease of use of an unspecified speech recognition device, which requires no registration effort, and the high recognition rate of a speaker-specific speech recognition device. However, as mentioned above, registering a specific dictionary requires an action called "registration," and the user must memorize "words that are difficult to recognize" by themselves. It still falls short of the ease of use inherent in unspecified voice recognition devices.

[0005]

[Purpose] The present invention has been made in view of the above-mentioned circumstances, and in particular, in order to reduce the burden on users regarding registration of specific dictionaries, it is an object of the present invention to automatically register such specific dictionaries at the time of recognition operation. It was done for that purpose.

[0006]

[Structure] In order to achieve the above objects, the present invention provides (1)
A feature extraction unit extracts the features of the input voice; a specific dictionary unit stores the extracted voice features as a dictionary for specific speakers; and a specific dictionary unit stores the extracted voice features as a dictionary for specific speakers; a non-specific dictionary section that stores the information as a dictionary; a similarity calculation section that calculates the degree of similarity between the extracted feature quantity of the voice and a plurality of dictionaries stored in the specific dictionary section and the non-specific dictionary section; and a result selection unit that selects one or more dictionaries as recognition results in descending order of similarity, the recognition result input unit receiving an input as to whether or not the recognition result is correct; A dictionary buffer that temporarily stores a plurality of voice feature values input during a recognition operation; The same word number is stored in the word number buffer (Th1).
), the degree of similarity between the features of the voices stored in the dictionary buffer corresponding to the word number is calculated, and when the degree of similarity is greater than or equal to a predetermined threshold (Th2), the The present invention is characterized in that any of the feature quantities stored in the plurality of dictionary buffers is stored as a specific dictionary in the specific dictionary section, and further, (2) in the above (1), the feature quantity is stored in the plurality of dictionary buffers as a specific dictionary. The feature is that the average value or median value of the stored feature quantities is stored as a specific dictionary in the specific dictionary section, and further, the above (1) or (
In 2), if the similarity when the input voice is correct (3) is higher than a predetermined threshold (Th3), the word number is not stored in the word number buffer and the contents of the dictionary buffer stored at that time are stored. or (4)
If the similarity when the input voice is correct is higher than a predetermined threshold (Th4), the word number is not stored in the word number buffer and the contents of the single or multiple dictionary buffers corresponding to the word number are stored. It is characterized by clearing. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining one embodiment of the present invention. In the figure, parts having the same functions as those of the prior art shown in FIG. 2 are designated by the same reference numerals as in FIG. is attached. In FIG. 1, in the initial state, the specific dictionary 4, dictionary buffer 10, and word number buffer 11 are all empty. At this time, this speech recognition device functions purely as a speaker-independent speech recognition device. When speech for recognition is input, its feature quantity is extracted by the feature extraction unit 1. The extracted feature amounts are sent to the similarity calculation section 2. Further, this feature quantity is stored in the dictionary buffer 10. The similarity calculation unit 2 calculates the similarity between the dictionary in the unspecified dictionary 3 and the previous feature amount,
A recognition result (word number, etc.) corresponding to the result is outputted by the result output unit 5. Here, a correct/incorrect response to the current recognition result is input by the user through the recognition result input unit 12 (here, a key switch is used) having a correct/incorrect key. If this recognition result is correct and the similarity at this time is less than a predetermined threshold, the word number of the correct answer is transferred to the dictionary buffer.
It is stored in the word number buffer 11 in association with 0. Even if the answer is correct, if the degree of similarity is sufficiently high (higher than a predetermined threshold), the unspecified dictionary 3 will be able to sufficiently recognize it, so the buffer at this time is cleared. In this way, feature quantities are stored in the dictionary buffer 10 each time recognition is performed. After this,
When the same word number stored in the word number buffer 11 appears a predetermined number (Th1), the similarity between the voice features stored in the dictionary buffer 10 corresponding to the word number is calculated, and the similarity is calculated. When the degree of similarity is equal to or greater than a predetermined threshold (Th2), one of the feature amounts (or average value, median value, etc.) stored in the plurality of dictionary buffers is stored in the specific dictionary section 4 as a specific dictionary. Once a feature quantity is registered in the specific dictionary 4, it becomes a recognition target from the next time onwards. By repeating recognition as described above, only words that are difficult to recognize (do not have a high degree of similarity) in the non-specific dictionary 3 can be automatically registered in the specific dictionary 4. The reason why the similarity of multiple utterances is calculated and the one that is equal to or higher than a certain threshold is adopted is to prevent sudden noises or voices that are not in good condition from being registered in a specific dictionary. Although one embodiment of the present invention has been described above, the details of parts other than those related to the present invention, such as the operation of the control system and the recognition device, are omitted because they are well known in known examples. Furthermore, regarding the user's input of correct/incorrect answers, there are various possibilities, such as determining the correct answer when there is no input of an incorrect answer, or determining the answer as incorrect by canceling the action that occurs after recognition, but these do not limit the present invention. It's not a thing. Furthermore, there are no limitations whatsoever on the voice features or the recognition algorithm.

[0008]

[Effect] As is clear from the above explanation, claims 1 and 2
According to the invention, since the speech at the time of recognition can be automatically registered as a specific dictionary, the burden on the user regarding registration of the specific dictionary can be reduced. According to the third and fourth aspects of the invention, since registration in the specific dictionary is performed only for words that are unspecified and difficult to recognize, the specific dictionary is not required and an increase in capacity can be prevented.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram for explaining an embodiment of a speech recognition device according to the present invention.

FIG. 2 is a configuration diagram for explaining an example of a conventional speech recognition device.

[Explanation of symbols]

1...Feature extraction unit, 2...Similarity calculation unit, 3...Unspecified dictionary,
4...Specific dictionary, 5...Result output section, 6...Key switch, 7
...Control unit, 10...Dictionary buffer, 11...Word number buffer, 12...Recognition result input unit.

Claims (4)

[Claims] Claim 1: A feature extraction unit that extracts features of input speech; a specific dictionary unit that stores extracted feature quantities of a plurality of speech sounds as a dictionary for a specific speaker; Similarity calculation that calculates the degree of similarity between the unspecified dictionary section stored as a dictionary for unspecified speakers, the feature amount of the extracted voice, and a plurality of dictionaries stored in the specified dictionary section and the unspecified dictionary section. and a result selection unit that selects one or more dictionaries as recognition results in descending order of the calculated similarity, in which a recognition result is input as to whether or not the recognition result is correct. a result input unit; a dictionary buffer that temporarily stores a plurality of feature quantities of a voice input during a recognition operation; and when a recognition result for the voice is correct, the word number is stored in the dictionary buffer. It is equipped with a word number buffer that stores a plurality of word numbers in association with speech, and when a predetermined (Th1) number or more of the same word number is stored in the word number buffer, the speech stored in the dictionary buffer corresponding to the word number is , and when the similarity is greater than or equal to a predetermined (Th2) threshold, one of the feature values stored in the plurality of dictionary buffers is sent to the specific dictionary section as a specific dictionary. A speech recognition device characterized by memory. 2. The speech recognition device according to claim 1, wherein an average value or a median value of the feature quantities stored in the plurality of dictionary buffers is stored as a specific dictionary in a specific dictionary section. Device. 3. In the speech recognition device according to claim 1 or 2, the degree of similarity when the input speech is correct is a predetermined (
Th3) If it is higher than the threshold value, the speech recognition device does not store the word number in the word number buffer and clears the contents of the dictionary buffer stored at that time. 4. In the speech recognition device according to claim 1 or 2, the degree of similarity when the input speech is correct is a predetermined (
Th4) If the threshold value is higher than the threshold value, the speech recognition device clears the contents of one or more dictionary buffers corresponding to the word number without storing the word number in the word number buffer.