JPH04155398A - Voice recognition device - Google Patents

Abstract

PURPOSE:To prevent a fluctuation of rejection rate by a word and a fluctuation of rejection rate by a template group, by deciding the rejection parameter to each temperature or to a template group to be the recognition object, and rejecting by using the parameter. CONSTITUTION:When a speaker pushes the key of a keyboard 9, a register signal is transmitted to a controller 8, and a switch circuit 3 is converted to a register 4 side. The voice of the speaker is converted to an electric signal and delivered to a voice feature extractor 2. The extracted feature amount is delivered to the register 4 to produce a template. The data of all the templates are stored in a template memory 6. After that, the register 4 delivers a finishing signal to a controller 8, the controller 8 makes a recognition member 5 carry out the matching of the templates each other, and delivers the degree of similarity of them to a rejection parameter deciding member 7. In the deciding member 7, the rejection parameter is determined as to every template.

Description

[Detailed description of the invention] Industrial applications The present invention relates to a speech recognition device.

BACKGROUND ART Conventionally, when a speech recognition device performs a rejection, the rejection rate varies depending on the word to be uttered with a fixed rejection parameter, and the rejection rate may be very high or low depending on the word. For example, a rejection rate of 2% for one word group may be 10% for another word group.

Furthermore, a similar phenomenon may occur when a new word is additionally registered.

Problems to be Solved by the Invention When performing rejection as described above, if the recognition target word changes, if the same rejection parameter is used to reject the recognition result, a certain group of templates may be rejected properly. Even if the parameter value is recognized, the rejection rate may be too high or too low when recognized using other template groups. cannot be prevented.

Means for Solving the Problem Therefore, in order to solve such problems, claim 1
In the described invention, a voice feature extraction unit is provided for extracting the features of input voice, a registration unit is provided for creating a template using the feature amount of the input voice extracted by the voice feature extraction unit, and the template is registered in the registration unit. A template storage unit that stores the templates stored in the template is provided, and a recognition unit that performs matching between templates and calculates the degree of similarity using the template group stored in the template storage unit. A reject parameter determination unit is provided that determines a reject parameter for each template in accordance with the determined degree of similarity.

In the invention as set forth in claim 2, a voice feature extraction section is provided for extracting the features of the input voice, and a registration section is provided for creating a template using the feature amount of the input voice extracted by the voice feature extraction section. A template storage unit is provided to store the templates registered in the template storage unit, and a recognition unit is provided to match each template using the template group stored in the template storage unit and calculate the degree of similarity. A reject parameter determining unit is provided that determines a reject parameter for each combination of templates in accordance with the degree of similarity determined by the recognition unit.

In the invention as set forth in claim 3, a voice feature extraction section is provided for extracting the features of the input voice, and a registration section is provided for creating a template using the feature amount of the input voice extracted by the voice feature extraction section. A template storage unit is provided to store the templates registered in the template storage unit, and a recognition unit is provided to match each template using the template group stored in the template storage unit and calculate the degree of similarity. A reject parameter determining unit is provided which determines a reject parameter according to the degree of similarity determined by the recognition unit and updates the value of the reject parameter each time the template is added thereafter.

In the invention as set forth in claim 4, a voice feature extraction section is provided for extracting the features of the input voice, and a registration section is provided for creating a template using the feature amount of the input voice extracted by the voice feature extraction section. A template reject parameter storage unit is provided to store the template and reject parameters registered in the template reject parameter storage unit, and the reject parameter and the template stored in the template reject parameter storage unit are read out, and the template group is used to distinguish between the templates. A recognition unit that performs matching and calculates the degree of similarity is provided, and a reject parameter is determined according to the degree of similarity determined by the recognition unit at the time of registration, and the value of the reject parameter is determined every time the template is added thereafter. We have provided a reject parameter determination section that updates the parameters.

According to the invention described in claim 1, a rejection parameter is determined for each template to be recognized, and rejection is performed using this parameter, thereby making it possible to prevent fluctuations in the rejection rate depending on words.

According to the second aspect of the invention, a rejection parameter is determined for each combination of templates to be recognized, and rejection is performed using this parameter, thereby making it possible to prevent variations in the rejection rate due to words.

According to the third and fourth aspects of the invention, a rejection parameter is determined based on a group of templates to be recognized, and rejection is performed using this parameter, thereby making it possible to prevent variations in the rejection rate due to the group of templates.

Embodiment An embodiment of the invention set forth in claim 1 will be described based on FIG. First, the overall structure will be described. A microphone 1 that converts voice into an electrical signal is connected to a voice feature extractor 2 that extracts features of input voice. This audio feature extraction section 2 is connected to a switch circuit 3. This switch circuit 3 is connected to a registration section 4 and a recognition section 5 so that switching can be performed. The registration unit 4
A template can be created using the feature amount of the input speech extracted by the speech feature extraction section 2. A template storage unit 6 that stores the registered template is connected to the registration unit 4.

Further, this template storage section 6 is connected to the recognition section 5. The recognition unit 5 has the function of matching templates using the template group stored in the template storage unit 6 and calculating the degree of similarity between the templates. Further, the recognition section 5 is connected to a reject parameter determination section 7. The reject parameter determining section 7 has the function of determining a reject parameter for each template according to the degree of similarity determined by the recognizing section 5.

Furthermore, as a configuration of other parts, a control section 8 is provided which is connected to each of the above-mentioned sections. Connected to this control section 8 are a keyboard 9 for switching between registration and recognition, and a recognition display section 10 for displaying recognition results.

In such a configuration, first, the time of template registration will be described. When the speaker presses a predetermined key on the keyboard 9, a signal to perform registration is transmitted to the control unit 8. When the control section 8 receives the signal, it switches the switch circuit 3 to the registration section 4 side. Next, the voice uttered by the speaker into the microphone 1 is converted into an electrical signal and sent to the voice feature extraction section 2. Here, the extracted feature amount is sent to the registration unit 4, and a template is created thereby. In this case, once templates for the number of words set by the speaker using the keyboard 9 are created, the data of all the templates is sent to the template storage section 6 and stored therein. Thereafter, the registration section 4 sends an end signal to the control section 8, and when the control section 8 receives the signal, it causes the recognition section 5 to start matching the templates. The recognition unit 5 performs matching between templates using the template group created by the registration unit 4, and sends the degree of similarity to the rejection parameter determination unit 7. This is performed for all combinations of templates, and based on all of the similarity data, the reject parameter determining unit 7 determines reject parameters for each template.

For example, for template A whose similarity with any template is X or more, the reject parameter is determined to be X, and if the similarity with any template is y or more and X or less. For template B, the reject parameter is determined to be Y.

Next, the time of recognition will be described. When the speaker presses a predetermined key on the keyboard 9, a signal to perform recognition is transmitted to the control unit 8. When the control section 8 receives the signal, it switches the switch circuit 3 to the recognition section 5 side. Next, the voice generated by the speaker into the microphone l is converted into an electrical signal and sent to the voice feature extraction section 2. The feature amount extracted here is sent to the recognition unit 5 and matched with the template group sent from the template storage unit 6. As a result, the item with the highest degree of similarity is
-M3 recognition candidate J, with the second largest one as the "second recognition candidate", and the degree of similarity, difference, ratio, etc. with that template, and the template of the first recognition candidate determined at the time of template registration. It is determined whether or not to reject using the rejection parameters for the target, and if it is not rejected, the "first recognition candidate" is set as the recognition result, and if it is rejected, a signal indicating that it has been rejected is sent to the recognition display unit 10. send.

As described above, by determining a rejection parameter for each template to be recognized and rejecting using this parameter, it is possible to prevent fluctuations in the rejection rate depending on words.

Next, an embodiment of the invention according to claim 2 will be described. Note that explanations of the same parts as those of the circuit (see FIG. 1) described in the embodiment of the invention according to claim 1 described above will be omitted, and the same parts will be denoted by the same reference numerals.

Here, the internal configuration of the reject parameter determining section 7 is changed. That is, the reject parameter determining unit 7 determines a reject parameter for each combination of templates according to the degree of similarity determined by the recognizing unit 5.

Thereby, at the time of registration, the reject parameter determination unit 7 determines a reject parameter for each combination of templates. For example, the reject parameter is determined to be X for templates A and B whose similarity is greater than or equal to X, and templates A and C whose similarity is greater than or equal to y and less than or equal to X. , the reject parameter is determined to be Y.

In addition, during recognition, the rejection parameters determined by the combination of templates "first recognition candidate" and "second recognition candidate" and the degree of similarity, difference, ratio, etc. between each template are evaluated. If the recognition result is not rejected, the "first recognition candidate" is sent as the recognition result, and if the recognition result is rejected, a signal indicating the rejection is sent to the recognition display section ]O. do.

As described above, by determining a rejection parameter for each combination of templates to be recognized and rejecting using this parameter, it is possible to prevent fluctuations in the rejection rate depending on words.

Next, an embodiment of the invention according to claims 3 and 4 will be described based on FIG. Note that the description of the same parts as in the embodiment of the invention described in claim 1 (see FIG. 1) described above will be omitted, and the same parts will be denoted by the same reference numerals.

Here, the internal configurations of the reject parameter determination section 7, template reject parameter storage section 11, and recognition section 5 are changed. That is, the reject parameter determining unit 7 has the function of determining a reject parameter according to the degree of similarity determined by the recognizing unit 5, and updates the value of the reject parameter every time a template is added thereafter. be. Further, the template reject parameter storage section 11 has the function of not only storing templates registered in the registration section 4 but also storing reject parameters.

Further, in addition to reading out the template stored in the template rejection parameter storage unit 11, the recognition unit 5 also reads out the rejection parameters.

As a result, at the time of registration, the reject parameter determining unit 7 determines reject parameters based entirely on the similarity data and the number of templates. for example,
If the number of combinations with a similarity of X or more is y% or more of all combinations, the reject parameter is determined to be A, and if it is 2% or more, the reject parameter is determined to be B.

The reject parameters determined in this way are sent to the template reject parameter storage section 11 and stored together with the created template.

Furthermore, when creating a new template for a different word group or when adding a new word, the same operation as above is performed, and the reject parameters determined by the reject parameter determining unit 7 are created by the registering unit 4. The template is sent to the template reject parameter storage unit 11j together with all the templates that have been created, and is stored in an area different from the area where the template group is stored.

As mentioned above, when registering templates, the degree of similarity between each template is measured when all registrations are completed, and the reject parameter value is determined based on the distribution of the degree of similarity, the number of templates, etc. do. Then, when the registered template is stored in the template reject parameter storage section 11, the reject parameters are also stored at the same time. Furthermore, when loading (reading) the template group from the template reject parameter storage section 11, the reject parameters are also loaded at the same time. By operating in this way, it is possible to prevent the rejection rate from varying depending on the template group.

Effects of the Invention The invention as set forth in claim 1 provides a voice feature extraction unit that extracts features of input voice, and a registration unit that creates a template using the feature amount of the input voice extracted by the voice feature extraction unit. , a template storage unit that stores the templates registered in the registration unit; a recognition unit that matches each template using the template group stored in the template storage unit and calculates the degree of similarity; Since a reject parameter determination unit is provided, a reject parameter is determined for each template according to the degree of similarity determined by the recognition unit at the time of registration, and a reject parameter is determined for each template to be recognized. By rejecting using this parameter, it is possible to prevent fluctuations in the rejection rate depending on words.

The invention according to claim 2 provides a voice feature extraction section that extracts the features of the input voice, and a registration section that creates a template using the feature amount of the input voice extracted by the voice feature extraction section. A template storage unit is provided to store the templates registered in the template storage unit, and a recognition unit is provided to match each template using the template group stored in the template storage unit and calculate the degree of similarity. A reject parameter is determined for each combination of templates according to the degree of similarity determined by the recognition unit, and a reject parameter determination unit is provided, so that a reject parameter is determined for each combination of templates to be recognized. However, by rejecting using this parameter, it is possible to prevent fluctuations in the rejection rate depending on the word.

The invention according to claim 3 provides a voice feature extraction section that extracts the features of the input voice, a registration section that creates a template using the feature amount of the input voice extracted by the voice feature extraction section, and The section is provided with a template storage section that stores the registered templates, a recognition section that matches each template using the template group stored in the template storage section and calculates the degree of similarity; A reject parameter determination unit is provided which determines a reject parameter according to the degree of similarity determined by the recognition unit and updates the value of the reject parameter each time the template is added. By determining a rejection parameter based on a group of templates and rejecting using this parameter, it is possible to prevent fluctuations in the rejection rate due to the group of templates.

The invention according to claim 4 provides a voice feature extraction section that extracts the features of the input voice, and a registration section that creates a template using the feature amount of the input voice extracted by the voice feature extraction section. A template storage section is provided for storing the templates and reject parameters registered in the template storage section, the reject parameters and the templates stored in the template storage section are read out, and the templates are matched using a group of templates. A recognition unit that calculates the degree of similarity is provided, a reject parameter is determined according to the degree of similarity determined by the recognition unit at the time of registration, and the value of the reject parameter is updated every time the template is added thereafter. Since several reject parameter determining units are provided, reject parameters are determined according to the group of templates to be recognized, and rejection is performed using these parameters, thereby making it possible to prevent fluctuations in the rejection rate due to the group of templates.

[Brief explanation of drawings]

FIG. 1 is a circuit diagram showing an embodiment of the invention as claimed in claims 1 and 2, and FIG. 2 is a circuit diagram showing an embodiment of the invention as claimed in claims 3 and 4. 2... Audio feature extraction section, 4... Registration section, 5... Recognition section, 6... Template storage section, 7... Reject parameter storage section, 11... Template rejection parameter storage section 3 ,1 Figure

Claims (1)

[Claims] 1. A voice feature extraction unit for extracting features of input voice is provided,
A registration unit is provided for creating a template using the features of the input voice extracted by the audio feature extraction unit, a template storage unit is provided for storing the template registered in the registration unit, and the template is stored in the template storage unit. A recognition unit is provided that performs matching between each template using the template group and calculates the degree of similarity, and a rejection parameter is determined for each template according to the degree of similarity determined by the recognition unit at the time of registration. What is claimed is: 1. A speech recognition device comprising: a rejection parameter determination section for determining a rejection parameter. 2. Provide a voice feature extraction unit that extracts the features of input voice,
A registration unit is provided for creating a template using the features of the input voice extracted by the audio feature extraction unit, a template storage unit is provided for storing the template registered in the registration unit, and the template is stored in the template storage unit. A recognition unit is provided that performs matching between each template using the template group and calculates the degree of similarity, and a rejection parameter is set for each combination of templates according to the degree of similarity determined by the recognition unit at the time of registration. 1. A speech recognition device comprising a reject parameter determination unit that determines a rejection parameter. 3. Provide a voice feature extraction unit that extracts the features of input voice,
A registration unit is provided for creating a template using the features of the input voice extracted by the audio feature extraction unit, a template storage unit is provided for storing the template registered in the registration unit, and the template is stored in the template storage unit. A recognition unit is provided that performs matching between templates using the template group and calculates the degree of similarity, and a rejection parameter is determined according to the degree of similarity determined by the recognition unit at the time of registration. A speech recognition device comprising a reject parameter determining unit that updates the value of a reject parameter each time a template is added. 4. Provide a voice feature extraction unit that extracts the features of input voice,
A registration unit is provided for creating a template using the features of the input voice extracted by the audio feature extraction unit, and a template reject parameter storage unit is provided for storing the template and reject parameters registered in the registration unit. A recognition unit is provided that reads out the reject parameters and the templates stored in the template rejection parameter storage unit, performs matching between templates using a template group, and calculates the degree of similarity. A speech recognition device comprising: a reject parameter determining unit that determines a reject parameter according to the magnitude of the similarity, and updates the value of the reject parameter each time the template is added thereafter.