JPS59170899A - Voice registration pattern learning system for voice recognition equipment - 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] Technique 9 - The present invention provides a dictionary learning method in a speech recognition device.In creating a dictionary for a speech recognition device, normally one word is registered by taking the average of several times. There is. However, in a so-called specific speaker recognition device that performs recognition using a dictionary corresponding to the user, it is known that erroneous recognition increases as the way of speaking changes as time passes after the dictionary is created. To prevent this, dictionary learning is necessary. Conventionally, as this dictionary learning method, (1) a method of rewriting the dictionary only for words that have been successfully recognized (Japanese Patent Laid-Open No. 56-81
(No. 899), (2) A method of rewriting the dictionary when the degree of similarity between the dictionary and the input speech becomes less than a certain value (Japanese Patent Laid-Open No. 56-5
(No. 1800), (3) A method in which the dictionary is updated only by the operator's operation for incorrectly recognized words (Japanese Patent Laid-Open No. 15690/1983)
No. 0) etc. have been announced. However, the above (1)
In this method, everything is rewritten when recognition is successful, so the disadvantage is that it will also be rewritten if noise from ambient noise occurs during recognition utterance, or if two utterances are insufficient. There is, but this is the same as mentioned above (
The same applies to method 2), and there is a possibility that the dictionary may be corrupted in these methods. Furthermore, the method (3) has drawbacks such as the fact that errors in operation by the operator have a large effect on the dictionary. Note that the above-mentioned (Japanese Patent Application Laid-open No. 56-51800) describes a method of re-registering by taking the average of the dictionary and the input voice, but in this way, sudden fluctuations in the input voice cause the input voice to be added to the dictionary. Although it has the advantage of reducing the impact, it has the disadvantage that words that are used infrequently in the registered car language are not updated sufficiently.

The present invention has been made in order to eliminate the drawbacks of the above-mentioned prior art.In particular, when rewriting a dictionary based on input speech, the present invention reduces the influence of sudden fluctuations and also allows words that are used infrequently to be used. This paper attempts to provide a voice registration pattern learning method that can be updated sufficiently.

The structure of the present invention will be described below based on one embodiment.

In the present invention, when the input voice is correctly recognized by the recognition device, the average of the voice parameters and the parameters registered in the dictionary are newly registered in the dictionary. In this case, the dictionary parameters are registered with considerably greater weight than the input voice parameters. In addition, if the uttered voice is not recognized correctly, it is uttered again and the parameters of this input voice and the parameters of the dictionary are averaged and re-registered. The input speech is averaged by giving almost the same weight to the parameters of the dictionary, or by giving greater weight to the input speech. Therefore, as a result, words that are used less frequently and are likely to be misrecognized change significantly during updating, and words that are used more frequently are updated little by little.

The figure is a diagram showing an example of a configuration for achieving the present invention as described above, in which 51 is a microphone, 2 is a group of filters, 3 is a voice section cutting section, 4 is a changeover switch,
5 is a dictionary, 6 is a matching section, 7 is a maximum similarity selection section, 8 is a result display section, 9 is an averaging section, 10 is a weighting section, and 11 is a switch. Dictionary registration and recognition are as follows. Do it.

Registration of a dictionary> Turn the switch 4 to the dictionary side, decompose the voice input from the microphone 1 into a spectrum through the filter group 2, extract only the voice section through the voice section cutout section 3, and output it to the dictionary 95.
Register.

Recognition> Switch 1 is set to the recognition side, and the signal from which the speech section is cut out in the same manner as described above is compared with each single entry in the dictionary 5, and the degree of similarity is calculated. The word with the highest degree of similarity is output as the result. Looking at this display, the operator determines whether the recognition is correct or incorrect, and inputs it using the switch 11. This information is reflected in the weight, and in the event of a misrecognition, a re-speech is requested, and the word with a higher weight is averaged with the dictionary word and is re-registered in the dictionary. On the other hand, if the answer is correct, it is averaged with the dictionary word with a smaller weight than when it is incorrectly recognized.

As is clear from the above description, according to the present invention, it is possible to realize a speech recognition dictionary that can obtain a stable recognition rate without updating for a long time.

[Brief explanation of drawings]

The figure is a configuration diagram for explaining one embodiment of the present invention. 1...Microphone, 2...Group of filters, 3.
・Voice section cutting section, 4... changeover switch, 5
. . . Dictionary, 6. Collation unit, 7. Maximum similarity selection unit, 8. Result display unit, 9. Averaging unit, 10.
・Weighting is part, 11...switch.

Claims (1)

[Claims] In a speech recognition device that registers words in a dictionary for comparison, if the input speech is correctly recognized, the pattern in the dictionary and the pattern of the input speech are given a certain weight and averaged, and then the result is returned to the dictionary. If the same word is incorrectly recognized, the next time the same word is uttered and correctly recognized, the weight value is changed and averaged, and then the word is registered in the dictionary again. A method for learning voice registration patterns in devices.