JPH02210500A - Standard pattern registering system - Google Patents

Abstract

PURPOSE:To prevent the registration of the voice with which a block detection is wrong at the time of voice registration and to improve a recognition rate by superposing utterance plural times and registering the same as standard patterns only when the results of collations respectively indicate the prescribed similarity or above and the fluctuations thereof are within a prescribed range. CONSTITUTION:The voices inputted from a microphone 1 are subjected to extraction of characteristics quantity in a characteristic quantity extracting section 2. The voices of the first utterance are preserved as they are in an input buffer 4. The voices of the 1st time and the 2nd time are collated in a pattern collating section 3 and whether the prescribed similarity or above is satisfied or not is checked. The voices of the 2nd time and the 3rd time are collated in the pattern collating section 3 when the voices of the 3rd time are inputted. Whether the prescribed similarity or above is satisfied or not is then checked. The fluctuations in the respective similarities of the collations executed twice are investigated if the above-mentioned similarity is satisfied. The utterances of the three times registered in the input buffer are superposed on each other and are registered as the standard patterns in a standard pattern memory section 5 if the fluctuations are within the prescribed range. The registration of the wrong voices is prevented in this way.

Description

[Detailed Description of the Invention] The present invention relates to a standard pattern registration method, more specifically, a standard pattern registration method for registering a more accurate standard pattern in a speech recognition device for a specific speaker. In the standard pattern registration method of the recognition device, there is no effective way to check whether the section of the voice uttered for registration has been accurately detected, so the section of speech that is erroneously detected is registered as the standard pattern. There was a case. To solve this problem, in a speech recognition device that performs Mta by uttering a voice multiple times, when creating a standard pattern, 2
A method has been proposed in which the word length uttered for the first time and subsequent times is compared with the word length uttered for the first time, and if the comparison result exceeds a certain range, the word length is not registered as a standard pattern, thereby increasing the accuracy of the standard pattern.

However, in the case of word length, even the same word varies greatly from utterance to utterance, and therefore, if the range is made too tight, there will be a large number of rejections during standard pattern registration, and there is a possibility that smooth registration will not be possible. On the other hand, if you loosen the range, short noises may be added, or even if the beginning or end of a word is missing, it may be recorded as it is.

The present invention has been made in view of the above-mentioned circumstances.
In speech recognition devices that register standard patterns by uttering them multiple times, this technology was developed to prevent the registration of voices with incorrect segment detection during voice registration, create accurate standard patterns, and improve recognition rates. It is.

1M In order to achieve the above object, the present invention registers the speech to be recognized in advance, and when the speech is inputted by the speech input means, performs recognition by pattern matching the input speech with the registered speech. In a voice registration method of a voice recognition device, voices uttered multiple times for registration are each stored in an input cover sofa, and at the same time, adjacent utterances are each pattern-matched, and the matching results each indicate a predetermined degree of similarity or higher; Moreover, only when the variation is within a predetermined range, the plurality of utterances are superimposed and registered as a standard pattern. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention,
Figure 2 is a flow diagram, in which 1 is a microphone, 2
3 is a feature extraction unit, 3 is a pattern matching unit, 4 is an input cover sofa, 5 is a standard pattern storage unit, and 6 is a result output unit. In the present invention, the voice to be recognized is registered in advance, and the voice is input as a voice input. In the voice registration method of the voice recognition device, which performs recognition by pattern matching the input voice with the registered voice when the input voice is input by means of
) is stored in a manual buffer, and at the same time, patterns are matched between adjacent utterances such as the first and second, second and third, etc., N-1 and N-th times, and the corresponding Only when the matching results show a predetermined degree of similarity or higher and the variation is within a predetermined range, the plurality of utterances are superimposed and registered as a standard pattern. When registering a standard pattern,
Any method may be used as long as the voice is uttered multiple times and registered, but here, the description will be made assuming that the voice is uttered three times and then registered. In the present invention, when registering the mark 7 ((at the time of pattern registration), first, the feature quantity of the voice inputted from the voice input means (microphone) 1 is extracted by the feature quantity extracting section 2, and is stored in the input buffer 4.

The 1st audio is saved as is in the human buffer 4,
When the second input is made, it is stored in the input buffer 4, and the pattern matching section 3 compares the first and second voices to check whether they satisfy a predetermined degree of similarity or higher. If the conditions are not met, it will be determined that either the first or second time the section was detected incorrectly.
Cancel and start over from the first utterance. If the conditions are satisfied, the third utterance is prompted and the audio is captured. When the third voice is input, it is stored in the input buffer 4, and the second and third voices are compared in the pattern matching section 3 to check whether the similarity exceeds a predetermined degree. If the conditions are not met, cancel the third utterance and re-enter only the third utterance,
Check again. If the similarity is satisfied, the variation in similarity between the two comparisons is checked to see if it is within a predetermined range, and if it is outside the range, registration is performed again. If the variation is within a predetermined range, the three utterances stored in the manual buffer are superimposed and recorded in the standard pattern storage section 5 as a standard pattern.

As is clear from the above description, according to the present invention, in a speech recognition device that registers a standard pattern by uttering it multiple times, it is possible to prevent the registration of speech with incorrect section detection during speech registration, and to create an accurate standard pattern. Creating patterns and improving the recognition rate becomes an IIF function.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention.
FIG. 2 is a flow diagram thereof. 1...Microphone, 2. Feature extraction section, 3. Pattern matching section, 4. Input buffer, 5. Standard pattern storage section, 6. Result output section. Patent applicant Ricoh Co., Ltd.

Claims (1)

[Claims] 1. In the voice registration method of a voice recognition device, the voice to be recognized is registered in advance, and when the voice is input by a voice input means, recognition is performed by pattern matching the input voice with the registered voice. When each voice uttered multiple times is stored in an input buffer, each adjacent utterance is pattern-matched, and each of the matching results shows a predetermined degree of similarity or higher, and the variation is within a predetermined range. A standard pattern registration method characterized in that only the plurality of utterances are superimposed and registered as a standard pattern.