JP2003323196A - Voice recognition system, voice recognition method, and voice recognition program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition system in which a plurality of voice recognition processing sections having different reliability references for the outputs is complementarily combined. <P>SOLUTION: Voice recognition processing sections 131 and 132 conduct voice recognition for input voice data 111 in accordance with intrinsic recognition systems and output voice recognition results and reliabilities indicating the sureness of the results. A voice recognition result evaluating means 133 converts the outputted reliabilities into an evaluation value, that is a common scale, employing an evaluation function 134 that is prepared for every voice recognition processing section, selects an optimum voice recognition result on the basis of the evaluation value obtained by the conversion and outputs the result. The function 134 includes a reference evaluation function corresponding to the section 131 and an evaluation function corresponding to the section 132. The latter evaluation function provides a higher evaluation value than the reference evaluation function for the reliability range in which the voice recognition result is closer to a correct solution as compared with the voice recognition result of the section 131. For the range of reliability that is not close to the correct solution, the latter evaluation function provides an evaluation value that is less than the reference evaluation function. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition technique, and more particularly to a speech recognition system, a speech recognition method and a speech recognition program which improve a recognition rate by using a plurality of speech recognition processing units.

[0002]

[Prior Art] SmartVoice (http://www.ne
c.co.jp/press/ja/0107/2301-01.html) and ViaVoice (ht
tp: //www-6.ibm.com/jp/voiceland/main/function.htm
l) and other voice recognition systems are sold as products. However, since these conventional speech recognition systems use a single speech recognition method, it is possible to use multiple people's voices with different utterance methods or environments in which voices of different languages such as Japanese and English are mixed. Then, there is a problem that the recognition rate becomes high or low depending on the time and the case. Other documents describing a voice recognition system that uses a single voice recognition method include JP-A-8-248981 and JP-A-2001-175276.

Therefore, in order to adapt to a wide variety of voice input environments and improve the recognition rate, there is provided a voice recognition system in which a plurality of voice recognition means are complementarily used.
It is proposed in 8-202388 (hereinafter referred to as Reference 1). The conventional speech recognition system described in this document 1 performs speech recognition according to a unique recognition method for each input speech, and the reliability indicating the speech recognition result and its certainty (distance (distance) in document 1 and A plurality of voice recognition processing units for respectively outputting) and a recognition result integration unit for collectively evaluating the plurality of voice recognition results by a predetermined standard and outputting a final result. Here, as a predetermined criterion, the past hit rate of each speech recognition processing unit (the ratio of the number of correct output times to the total number of past recognitions, which is referred to as reliability in Reference 1) is used, and each speech recognition is performed. The reliability is corrected by multiplying the reliability output from the processing unit by the past hit rate of the voice recognition processing unit,
The final result is selected based on the result of comparison of the plurality of corrected reliabilities.

According to this voice recognition system, for example,
As a template to be stored in advance for recognition in a plurality of voice recognition processing units, different voicing methods for the same word (by gender, age, region, etc.),
And if you use templates created in different formats, you can improve the recognition rate by integrating each recognition result.
In addition, it is said that the number of recognized vocabularies can be increased without increasing the calculation cost by using templates created for different words.

[0005]

[Problems to be Solved by the Invention] However, Document 1
The voice recognition system described in 1 has a problem that it is difficult to combine arbitrary voice recognition processing units having different recognition methods. The reason is that the plurality of speech recognition processing units to be combined need to be premised that "the reliability of output is generated with completely the same reference and scale". That is, in Document 1, when there are two voice recognition processing units A and B, it is estimated that the recognition reliability is the same when both of them output the same reliability (distance in Document 1). This is because, in the case of a plurality of speech recognition processing units having different reliability generation criteria to be output, even if the reliability is the same, the objective certainty of both recognition results is not always the same. For this reason, the recognition methods are greatly different, and as a result, it is not possible to combine arbitrary speech recognition processing units having different output reliability generation standards and scales. Note that in Document 1, the reliability output from the voice recognition processing unit is corrected, and the process of comparing the corrected reliability is performed, but the correction is performed on information unique to each voice recognition processing unit. Since the correction is based on the above, it is not possible to complementarily combine a plurality of voice recognition processing units having different reliability generation standards even if such correction is performed.

An object of the present invention is to provide a voice recognition system and a voice recognition method which can complemently combine a plurality of voice recognition processing units having different standards and scales for outputting reliability. It is in.

[0007]

According to a first speech recognition system of the present invention, speech recognition is performed on an input speech according to a unique recognition method, and a speech recognition result and a reliability indicating its certainty are respectively obtained. The plurality of voice recognition processing units to output and the reliability output from each of the plurality of voice recognition processing units are shared by the plurality of voice recognition processing units by using an evaluation function prepared for each voice recognition processing unit. A voice recognition result evaluation means for converting the evaluation value as a scale and selecting and outputting a voice recognition result from the plurality of voice recognition results based on a comparison result of the plurality of evaluation values obtained by the conversion. ing. Here, the evaluation function includes, for example, a reference evaluation function corresponding to any one voice recognition processing unit of the plurality of voice recognition processing units and an evaluation function corresponding to another voice recognition processing unit. In the evaluation function corresponding to the voice recognition processing unit of, even if the reliability value is the same, the voice recognition result is closer to the correct answer range than the voice recognition result of the one voice recognition processing unit. Is a higher evaluation value than the reference evaluation function, and an evaluation value lower than the reference evaluation function is obtained in the range of reliability not close to the correct answer.

The second voice recognition system of the present invention, in addition to the configuration of the voice recognition system described above, further presents to the user the voice recognition results obtained by the plurality of voice recognition processing units for sample voice data. Means, means for displaying the reference evaluation function on a graph of reliability and evaluation value, and an evaluation value to be given to the reliability obtained by the other speech recognition processing unit on the graph from the user And a means for creating the evaluation function for the other speech recognition processing unit based on a set of plot points on the graph.

The first voice input method of the present invention is a)
In parallel, a plurality of voice recognition processing units perform voice recognition on an input voice according to a unique recognition method, and output a voice recognition result and a reliability indicating its certainty, respectively, and b) the plurality of voice recognition processing units. Converting the reliability output from each of the speech recognition processing units into an evaluation value, which is a scale common to the plurality of speech recognition processing units, using an evaluation function prepared for each of the speech recognition processing units; ) A step of selecting and outputting a voice recognition result from the plurality of voice recognition results based on a comparison result of the plurality of evaluation values obtained by the conversion. Here, the evaluation function includes, for example, a reference evaluation function corresponding to any one voice recognition processing unit of the plurality of voice recognition processing units and an evaluation function corresponding to another voice recognition processing unit. In the evaluation function corresponding to the voice recognition processing unit of, even if the reliability value is the same, the voice recognition result is closer to the correct answer range than the voice recognition result of the one voice recognition processing unit. Is a higher evaluation value than the reference evaluation function, and an evaluation value lower than the reference evaluation function is obtained in the range of reliability not close to the correct answer.

The second speech recognition method of the present invention further includes d) a step of generating an evaluation function corresponding to the other speech recognition processing section, and the step d is d-1) for the sample speech data. Presenting the voice recognition results obtained by the plurality of voice recognition processing units to the user, d-2)
A step of displaying the reference evaluation function on a graph of reliability and an evaluation value; and d-3) an evaluation value to be given to the reliability obtained by the other speech recognition processing unit on the graph. Plotting according to the instructions from d.
4) repeating the steps d-1 to d-3 by the number of prepared sample voice data, and d-5) the step for the other voice recognition processing unit based on the set of plot points on the graph. And a step of creating an evaluation function.

[0011]

In the first voice recognition system and the voice recognition method of the present invention, the reliability output from each of the plurality of voice recognition processing units is determined by using the evaluation function prepared for each voice recognition processing unit. The speech recognition processing unit converts the evaluation value to a common scale and selects the optimum speech recognition result based on the comparison result of the converted multiple evaluation values. Even when combining multiple speech recognition processing units with different output reliability standards and scales,
A voice recognition system and method for selecting and outputting a voice recognition result closer to a correct answer can be obtained.

In the second speech recognition system and speech recognition method of the present invention, the user can easily create a necessary evaluation function from the sample voice data and the reference evaluation function.

[0013]

BEST MODE FOR CARRYING OUT THE INVENTION Next, a first embodiment of the present invention will be described in detail with reference to the drawings.

Referring to the block diagram of FIG. 1, a voice recognition system according to a first embodiment of the present invention includes a voice data input device 101, a data processing device 102 operated by program control, and a character data output device 103. A mode switch 104, a display device 105, and an input device 106.
It consists of and.

The mode switch 104 is composed of a switch for switching the operation mode of the voice recognition system between the voice recognition mode and the evaluation function generation mode. The output of the mode switch 104 is input to the data processing device 102 and the character data output device 103.

The voice data input device 101 inputs the input voice data 111 input by the user in the voice recognition mode and the sample voice data 112 input by the user in the evaluation function generation mode to perform data processing. Audio data input means 121 for outputting to the device 102
It is configured to include. The voice data input means 121 has a microphone that collects voice, an amplifier that amplifies its output, an AD converter that converts its output into a digital signal, and the like.

The data processing device 102 includes two voice recognition processing units 131 and 132 and a voice recognition result evaluation unit 133.
An evaluation function 134, an evaluation function creating means 135, and a voice recognition result plotting means 136.

The voice recognition processing units 131 and 132 perform voice recognition on the voices input from the voice data input unit 121 according to their own recognition methods, and respectively give the voice recognition result and the reliability indicating its certainty. Has a function to output. Here, the voice recognition result is character data. Further, the reliability is a numerical value whose maximum value is 100%, for example. The voice recognition processing units 131 and 132 perform the same operation in both the voice recognition mode and the evaluation function generation mode, except that the input voice data 111 and the sample voice data 112 are different from each other.

In the case of the present embodiment, the voice recognition processing section 13
1 is composed of a voice recognition unit 137 and a reliability calculation unit 139. In general, various techniques such as a knowledge engineering technique, a pattern matching technique, a stochastic statistical technique, and a neural network technique have been proposed or put into practical use as the voice recognition technique, and even if the same technique is used. There are many varieties. Any of these methods can be used for the voice recognition means 137. Voice recognition means 13
When the input voice is given, 7 generates some recognition result candidates together with their reliability by internal processing, and outputs the candidate with the highest reliability and its reliability. Reliability calculation means 1
39 is the case where the voice recognition means 137 outputs the reliability in units such as words that form the input voice,
It is a means for calculating the reliability of the entire input voice from the reliability of each structural unit. For example, the average value of the reliability of all the constituent units is set as the reliability of the entire input voice. Voice recognition means 13
When 7 itself outputs the reliability of the entire input speech, the reliability calculation means 139 causes the speech recognition means 137.
The reliability output from can be configured as is, or can be omitted.

The other speech recognition processing section 132 is also composed of a speech recognition section 138 and a reliability calculation section 140 similar to the speech recognition section 137 and the reliability calculation section 139 in the speech recognition processing section 131. However, the recognition method of the voice recognition means 138 is different from the recognition method of the voice recognition means 137. Here, the difference in the recognition method is not limited to the case where the speech recognition methods themselves such as the knowledge engineering method, the pattern matching method, the stochastic statistical method, and the neural network method are different. The recognition method is used, but it also includes the case where the standard patterns stored in advance for performing the recognition process are different.

The operation of the voice recognition result evaluation means 133 differs depending on the mode. In the voice recognition mode, the voice recognition results and the reliability are input from the voice recognition processing units 131 and 132, and the respective reliability is calculated by using the evaluation function 134 prepared for each voice recognition processing unit. It is converted into an evaluation value which is a common scale for relatively comparing, and one speech recognition result is selected from the plurality of speech recognition results based on the comparison result of each evaluation value obtained by this conversion. It is output to the character data output device 103. In the evaluation function generation mode, the voice recognition result and the reliability output from the voice recognition processing units 131 and 132 are directly output to the character data output device 103.

The evaluation function 134 is the speech recognition processing unit 131.
Evaluation function fa () for speech recognition processing unit 132
Composed of fb () and. In the case of the present embodiment, either one is fixedly set in advance and the other is created in the evaluation function generation mode. An evaluation function fixedly set in advance is called a reference evaluation function. Either fa () or fb () may be used as the reference evaluation function, but in the following description, the voice recognition processing unit 1
The fa () for 31 is used as the reference evaluation function. FIG. 2 shows an example of fa () in a graph format. The horizontal axis represents the reliability and the vertical axis represents the evaluation value, and the line segment indicated by reference numeral 201 corresponds to fa (). In this case, the reference evaluation function fa () is a linear function having the reliability as a variable. According to this fa (), for example, the reliability of 65% is the value of 85
Is converted to the evaluation value of. The reference evaluation function is not limited to the linear function. In the evaluation function generation mode, the remaining evaluation function fb () is generated in relation to the reference evaluation function fa (). The voice recognition result plotting unit 136 and the evaluation function creating unit 135 are involved in this generation, and their functions will be described later.

The character data output device 103 has a character data output means 151. The operation of the character data output means 151 differs depending on the mode. In voice recognition mode,
The character data of the voice recognition result output from the voice recognition result evaluation unit 133 is output as the recognition result 113. The output destination is not particularly limited, and may be the display device 105, or a recording medium or line not shown. On the other hand, in the evaluation function generation mode, the voice recognition processing unit 13
Both of the voice recognition results obtained in Nos. 1 and 132 and the reliability thereof are displayed on the display device 105 and presented to the user. More specifically, in this embodiment, the following operation is performed.

First, the reference evaluation function fa () is displayed in the graph display section of the display device 105 in the graph format as shown in FIG. Next, when the voice recognition results and the reliability obtained by the voice recognition processing units 131 and 132 with respect to a certain sample voice data are passed from the voice recognition result evaluation unit 133, the character data of each voice recognition result is sent to the user. Is displayed in the recognition result column of the display device 105 for presentation to the user, and the evaluation value converted by the reference evaluation function fa () is converted into the evaluation value by a graph display. Present on the graph of the reference evaluation function fa () displayed in the column,
Furthermore, the reliability of the voice recognition processing unit 132 is presented on the same graph. For example, as shown in FIG. 3, the voice recognition processing unit 1
If the reliability obtained in step 31 is 65%, for example, then 65 on the horizontal axis
The reference evaluation function 201 is parallel to the vertical axis from the position of the% scale.
By drawing a line segment 202 reaching up to, and drawing a line segment 203 that passes through the intersection of this line segment 202 and the reference evaluation function 201 and is parallel to the horizontal axis, the reliability of the voice recognition processing unit 131 is determined by the reference evaluation function fa (). It is shown to the user that it will be converted into the evaluation value of value 85 by converting with. If the reliability obtained by the voice recognition processing unit 132 is 50%, as shown in FIG.
Straight line 204 parallel to the vertical axis from the position of the 50% scale on the horizontal axis
By drawing, the reliability of the voice recognition processing unit 132 is presented on the graph.

The above-mentioned speech recognition result plotting means 136
Is a means for giving an evaluation point for the reliability of the evaluation function fb () to be automatically created, and one point on the straight line 204 drawn as described above is input to the input device 1
It has a function of plotting according to the user's instruction from 06.
Which part on the straight line 204 is designated by the user as a plot point is determined by comparing the voice recognition results of the voice recognition processing units 131 and 132 displayed on the display device 105.
Specifically, for example, it is performed as follows.

As shown in FIG. 3, the voice recognition processing unit 13
The reliability obtained this time with 1 and 132 is 65% and 50%, respectively.
If the voice recognition result of the voice recognition processing unit 132 is closer to the correct answer than the voice recognition result of the voice recognition processing unit 131, the reliability of 65% of the voice recognition processing unit 131 is set to the reference evaluation function fa (). Since it is necessary to give the speech recognition processing unit 132 an evaluation score higher than the evaluation score 85 obtained by converting
As illustrated in the plot point 205 of FIG. 3, a portion of the straight line 204 slightly above the line segment 203 is plotted. On the contrary, if the voice recognition result of the voice recognition processing unit 131 is closer to the correct answer than the voice recognition result of the voice recognition processing unit 132, the original reliability of 65%, 50% itself indicates the relationship correctly. Therefore, the intersection of the straight line 204 and the reference evaluation function 201 is plotted as illustrated in the plot point 206 of FIG. If there is no difference in the recognition result, the straight line 204
And the intersection of the line segment 203 is plotted.

As shown in FIG. 4, the voice recognition processing unit 13
The reliability obtained this time with 1 and 132 is 65% and 75%, respectively.
If the voice recognition result of the voice recognition processing unit 132 is closer to the correct answer than the voice recognition result of the voice recognition processing unit 131, the reliability of 65% of the voice recognition processing unit 131 is set to the reference evaluation function fa (). Since it is necessary to give the speech recognition processing unit 132 an evaluation score higher than the evaluation score 85 obtained by converting
For example, the straight line 2 as illustrated in the plot point 207 of FIG.
The intersection of 04 and the reference evaluation function 201 is plotted. On the contrary, if the voice recognition result of the voice recognition processing unit 131 is closer to the correct answer than the voice recognition result of the voice recognition processing unit 132, the line segment of the straight line 204 is exemplified as the plot point 208 of FIG. The portion slightly below 203 is plotted. If there is no difference in the recognition result, the intersection of the straight line 204 and the line segment 203 is plotted.

As shown in FIG. 5, the voice recognition processing unit 13
If the reliability obtained this time in Nos. 1 and 132 is the same 65% and the voice recognition result of the voice recognition processing unit 132 is closer to the correct answer than the voice recognition result of the voice recognition processing unit 131, The straight line 20 as illustrated in the plot point 209
The portion slightly above the line segment 203 of 4 is plotted. On the contrary, if the voice recognition result of the voice recognition processing unit 131 is closer to the correct answer than the voice recognition result of the voice recognition processing unit 132, the line segment of the straight line 204 is exemplified as the plot point 210 of FIG. The portion slightly below 203 is plotted. If there is no difference in the recognition result, the intersection of the straight line 204 and the line segment 203 is plotted.

The above-described plotting work is carried out for each sample voice data. The sample voice data is prepared in a sufficient number of samples so as to cover a wide variety of input voice environments. When the above-described plotting work is completed for all sample voice data, plot points as many as the sample voice data are obtained. The evaluation function creating means 135 generates an evaluation function fb () for the voice recognition processing unit 132 from the set of these plot points. The evaluation function fb () only needs to be a function that approximates the set of plot points well, and its specific creation method may be arbitrary. For example, for every 1% of the reliability scale on the horizontal axis, for all plot points on a straight line drawn from the reliability parallel to the vertical axis, the line segment drawn from that point on the horizontal axis intersects the vertical axis. It is possible to use a method in which the evaluation value is read, the average value is obtained, and a function representing a curve passing through the average value for each reliability at every 1% is represented by fb (). The evaluation function fb () thus created is registered in the evaluation function 134 and used in the voice recognition mode.

In the above description, a graph of the reference evaluation function is displayed on the display device 105, the reliability obtained by the voice recognition processing unit 131 and the evaluation value corresponding thereto are presented on the graph, and the voice recognition processing unit 132. Although the character data output unit 151 performs the presentation of the reliability obtained in the above step on the graph, these processes may be performed by the voice recognition result plotting unit 136. Further, the user may calculate the reliability obtained by the voice recognition processing unit 131 and the evaluation value corresponding to the reliability.

Next, the operation of the speech recognition system of this embodiment in the evaluation function generation mode will be described along the flow of the flowchart of FIG. The processing shown in FIG. 6 is executed when the mode switch 104 is switched to the evaluation function generation mode.

First, the user creates a reference evaluation function fa () that serves as a reference for evaluation of reliability in advance and sets it in the evaluation function 134 (step S101). Next, the sample voice data 11 is used by the user as voice data to be voice-recognized.
2 is voice data input means 1 of the voice data input device 101
21 is input, the sample voice data 112 input by the voice data input means 121 is input to the data processing device 102 (step S102). In the data processing device 102, the sample voice data 112 received from the voice data input means 121 is sent to the voice recognition processing unit 131 and the voice recognition processing unit 132.

The voice recognition processing unit 131 is a voice recognition means 1
37, the received sample voice data 112 is voice-recognized by a recognition method unique to the voice recognition processing unit 131 to output a voice recognition result A, and the reliability calculation means 139 outputs the reliability a of the voice recognition result. (Step S10
3). Similarly, in the voice recognition processing unit 132, the voice recognition unit 138 performs voice recognition on the received sample voice data 112 by a recognition method specific to the voice recognition processing unit 132, outputs a voice recognition result B, and the reliability calculation unit 140. Due to
The reliability b of the voice recognition result is output (step S1).
04).

Next, the voice recognition result evaluation unit 133 outputs the voice recognition results A and B and the reliability a and b received from the voice recognition processing units 131 and 132 to the character data output unit 151 of the character data output device 103. (Step S10
5), the character data output unit 151 displays the character data of both voice recognition results in the recognition result column of the display device 105, and the character recognition unit 131 of the voice recognition processing unit 131 uses the method described with reference to FIG. Reliability a and evaluation value fa corresponding to it
(a), the reliability b of the voice recognition processing unit 132 is presented on a graph (step S106).

Next, the user compares the two voice recognition results displayed in the recognition result column of the display device 105, judges which one is closer to the correct answer, and based on the judgment result, FIG.
Through the method described with reference to FIG. 5, the evaluation value fb (b) corresponding to the current reliability b of the voice recognition processing unit 132 is plotted on the graph by using the voice recognition result plotting unit 136 (step S107).

Next, the user confirms the completion of the sampling process, and if not completed (step S108 returns N
o), input the next sample voice data (step S)
102). As a result, the same processing as the above-mentioned processing is repeated. If all the sample voice data have been processed (Yes in step S108), the evaluation function creating means 135 is activated by an instruction from the input device 106. The evaluation function creating unit 135 uses the voice recognition processing unit 13 from the set of evaluation values fb (b) plotted by the user on the graph of the reliability and the evaluation value by the voice recognition result plotting unit 136.
The evaluation function fb () for 2 is generated and set in the evaluation function 134 (step S109).

FIG. 7 shows an example of a graph plotted for all sample voice data and an example of the evaluation function fb () generated from the set of the plots. In the figure, each of the small dots is a plot point, and the curve shown by the broken line is an example of the evaluation function fb () generated from the set of plot points. In this example, when the reliability is divided into two areas of high and low, even if the reliability is the same in the high area, the evaluation function
An evaluation function fb () is generated such that an evaluation value higher than that of fa () is obtained and an evaluation value lower than that of the evaluation function fa () is obtained in the low region even with the same reliability.

Next, the operation in the voice recognition mode of the voice recognition system of the present embodiment will be described along the flow of the flowchart of FIG. The process shown in FIG. 8 is executed when the mode switch 104 is switched to the voice recognition mode.

When the user inputs the input voice data 111 to be voice-recognized by using the voice data input means 121 of the voice data input device 101, the voice data input means 121 outputs the input voice data 111 to the data processing device 102.
(Step S201). Data processing device 10
In 2, the input voice data 111 received from the voice data input means 121 is sent to the voice recognition processing unit 131 and the voice recognition processing unit 132.

The voice recognition processing section 131 comprises the voice recognition means 1
37, the received input voice data 111 is voice-recognized by a recognition method unique to the voice recognition processing unit 131, and a voice recognition result A is output, and the reliability calculation means 139 outputs the reliability a of the voice recognition result. (Step S20
2). Similarly, in the voice recognition processing unit 132, the voice recognition unit 138 performs voice recognition on the received input voice data 111 by a recognition method unique to the voice recognition processing unit 132, outputs a voice recognition result B, and the reliability calculation unit 140. Thus, the reliability b of the voice recognition result is output (step S20).
3).

Next, the voice recognition result evaluation means 133 uses the reliability a received from the voice recognition processing unit 131 as the evaluation function fa.
() Is used to convert to the evaluation value fa (a), and the reliability b received from the speech recognition processing unit 132 is converted to the evaluation value fb (b) using the evaluation function fb () (step S204). If the evaluation value fa (a) of the voice recognition processing unit 131 is higher than the evaluation value fb (b) of the voice recognition processing unit 132 (Yes in step S205), the voice recognition result A of the voice recognition processing unit 131 is set. It is selected and output to the character data output device 103 (step S206). If not, that is, if the evaluation value fa (a) of the voice recognition processing unit 131 is not higher than the evaluation value fb (b) of the voice recognition processing unit 132 (step S).
No in 205), the voice recognition result B of the voice recognition processing unit 132 is selected and output to the character data output device 103 (step S207). The character data output unit 151 of the character data output device 103 recognizes the character data of the voice recognition result passed from the voice recognition result evaluation unit 133 as the recognition result 113.
Is output (step S208).

9 to 12, when the evaluation functions fa () and fb () are 201 in FIG. 2 and 301 in FIG. 7, respectively, the reliability a of the voice recognition processing unit 131 and its evaluation value fa
(a), reliability b of the voice recognition processing unit 132, and its evaluation value fb
The combination of (b) is shown. In FIG. 9, the reliability a is the reliability b.
The case where the evaluation value fa (a) is higher than the evaluation value fb (b) is also higher. In this case, the voice recognition result evaluation unit 133 selects the voice recognition result A as the voice recognition result. Figure 10
Indicates that the reliability a is lower than the reliability b and the evaluation value fa (a)
Also shows the case where the evaluation value is lower than the evaluation value fb (b). In this case, the voice recognition result evaluation unit 133 selects the voice recognition result B as the voice recognition result. In the case of FIG. 9 and FIG. 10, there is no particular difference from the case where the voice recognition result is selected with high or low reliability instead of the evaluation value. On the other hand, as shown in FIG. 11, the reliability a is higher than the reliability b, but the evaluation value fa (a) is equal to the evaluation value fb.
If it is lower than (b), the voice recognition result evaluation unit 133 selects the voice recognition result B as the voice recognition result. As shown in FIG. 12, when the reliability a is lower than the reliability b, but the evaluation value fa (a) is higher than the evaluation value fb (b), the voice recognition result evaluation unit 133 outputs the voice recognition result as a voice recognition result. The recognition result A is selected. In this way, in the cases of FIG. 11 and FIG. 12, the opposite results to the case where the voice recognition result is selected with high and low reliability are obtained.

In the above embodiment, the voice recognition processing section 1
The 32 evaluation function fb () was experimentally created based on the result of actual voice recognition using the reference evaluation function fa () and sample voice data, but the characteristics of the voice recognition processing unit to be combined are estimated. If possible, an evaluation function can be created in advance based on the estimation. Hereinafter, this point will be described.

The voice recognition means 13 of the voice recognition processing unit 131
Reference numeral 7 is a voice recognition means that performs voice recognition by phrase analysis and does not perform recognition processing by semantic analysis.
It is assumed that 9 calculates the reliability of the voice recognition result voice-recognized by the voice recognition means 137. As the reliability calculation method in the reliability calculation means 139, for example, a method in which the average of the reliability of each word included in the speech recognition result is used as the reliability of the speech recognition result can be used. That is,
For example, if the voice recognition result of "Today is a nice weather" is obtained when the input voice "Konichiha (wa) Itenkidesune" is recognized, "Today", "Ha", ", ",""Good","weather","is","ne","." When the reliability is expressed in terms of appearance frequency, "70%""60%""75%" respectively
"35%""50%""40%""30%""80%"
Then, the reliability of the speech recognition result of the entire input speech is set to "55%" of the average.

On the other hand, it is assumed that the voice recognition means 138 of the voice recognition processing section 132 is a voice recognition means for performing voice recognition by phrase analysis and semantic analysis. In addition, in the case of performing the semantic analysis, generally, the reliability of the entire input voice recognized by the continuous voice recognition is output.
For example, the similarity of the voice recognition result output by 38 is output as it is as the reliability of the voice recognition result of the entire input voice.

Voice recognition means 1 for performing voice recognition by phrase analysis
In the case of 37, since the arbitrary phrase of the character data of the voice recognition result is not influenced by the preceding and following phrases, the evaluation function fa () that obtains the evaluation value from the reliability of the voice recognition result should be almost a straight line. Considered optimal. For example, when the input voice "Konichiha (wa) Itenkidesune" is recognized, the phrases "today" and "ha" before and after the phrase "good"
Even if ",""weather","is","ne","." Can be expected to have a fairly high degree of reliability, the reliability of the phrase "good" is not affected, so the reliability of the speech recognition result of the entire input speech data is It is appropriate to apply a value such that the reliability of all words monotonically increases when the reliability of each word such as the average value of all words monotonically increases. Further, if the reliability of all words, which is the reliability of the voice recognition result, monotonically increases, the recognition rate, which is the rate at which the user recognizes the voice recognition result as the correct answer, also monotonically increases. Therefore, as indicated by the line segment 401 on the graph in FIG. 13, the evaluation function fa () that evaluates the reliability of the voice recognition unit 137 that performs voice recognition by phrase analysis is a linear function.

On the other hand, when using the voice recognition means 138 for performing voice recognition by semantic analysis, in the case of the above-mentioned example sentence, the voice recognition result shows "Today is good (good) weather."
"Today is good weather."
Since the number of combinations of voice recognition results is larger than that in the case of performing voice recognition by phrase analysis, the reliability of the overall voice recognition result is lowered. Further, when the words before and after the speech are recognized to some extent, the recognition rate of the entire input voice data is increased by utilizing the words. On the contrary, if the surrounding words are not recognized, the recognition rate may be low. For example,
If the words "today", "ha", ",""weather","is","ne" and "." Before and after the word "good" can be recognized to some extent, the input voice data can be input even if the reliability is not so high. The recognition rate of "Today's weather is good" is high in the voice recognition result of the entire voice data. Therefore, FIG.
As indicated by the curve 402 on the graph of, the evaluation function fb () that evaluates the reliability of the voice recognition unit 138 that performs voice recognition by semantic analysis is the voice recognition unit 1 that performs voice recognition by phrase analysis.
Compared to the evaluation function fa () for evaluating the reliability of 37, the evaluation value becomes high when the reliability is high, and the evaluation value becomes low when the reliability is low.

As described above, the voice recognition processing units 131 and 13
2 If both evaluation functions fa () and fb () can be estimated in advance,
The estimated and created evaluation functions fa () and fb () are assigned to the evaluation function 134
By setting to, the processing in the voice recognition mode can be started immediately. In that case, the evaluation function creation unit 135, the voice recognition result plotting unit 136, the mode switching unit 104, the display device 105, the respective elements of the input device 106, and the evaluation function generation included in the voice recognition result evaluation unit 133 and the character data output unit 151. By omitting the function in the mode, the voice recognition system as shown in FIG. 14 can be simplified.

In each of the above embodiments, the two voice recognition processing units are complementarily combined, but it is also possible to combine three or more voice recognition processing units. When two or more other evaluation functions are created from the reference evaluation function in the same manner as the evaluation function generation mode in the embodiment of FIG. 1, the voice recognition result of one voice recognition processing unit to which the reference evaluation function is assigned, The evaluation values are plotted by comparing each of the voice recognition results of the two or more voice recognition processing units. Therefore, the evaluation values for the number of other voice recognition processing units are plotted for one sample voice data. In order to distinguish which speech recognition processing unit each plot point corresponds to, a plurality of types of plotting points are assigned to other speech recognition processing units, and one to one is assigned to a set of plotting points of the same type. The evaluation function creating means 135 generates an evaluation function for each time.

In each of the above embodiments, the reliability calculation means 13 is provided in each of the voice recognition processing units 131 and 132.
Although 9, 140 are independently provided, the method of calculating the reliability is the same in the case of using a plurality of voice recognition means 137, 138 of the recognition method that uses the same method of voice recognition and has a different standard pattern. For example, as shown in FIG. 15, a plurality of voice recognition means 137 and 138 may use the same reliability calculation means 161.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above embodiments, and various other additions and changes can be made. Further, the voice recognition system of the present invention can be realized by a computer and a voice recognition program, as well as by hardware realizing the function of the voice recognition system. The speech recognition program is
It is provided by being recorded in a computer-readable recording medium such as a magnetic disk or a semiconductor memory, and is read by a computer at the time of starting the computer, and the operation of the computer is controlled, so that the computer in each of the above-described embodiments is controlled. Make it function as a voice recognition system.

[0052]

As described above, according to the present invention, even a plurality of speech recognition processing units having different output reliability generation standards and scales can be combined in a complementary manner. A voice recognition method is obtained.
Also, the user can easily create the evaluation function required for it from the sample voice data and the reference evaluation function.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a first exemplary embodiment of the present invention.

FIG. 2 is a diagram showing an example of a reference evaluation function used in the present invention.

FIG. 3 is an explanatory diagram of an evaluation point plotting operation performed by a user in an evaluation function generation mode.

FIG. 4 is an explanatory diagram of evaluation point plotting work performed by a user in an evaluation function generation mode.

FIG. 5 is an explanatory diagram of evaluation point plotting work performed by a user in an evaluation function generation mode.

FIG. 6 is a flowchart showing a processing example in an evaluation function generation mode according to the first embodiment of the present invention.

FIG. 7 is a diagram showing an example of a set of evaluation point plots executed for all sample voice data and an evaluation function created from the set.

FIG. 8 is a flowchart showing a processing example in a voice recognition mode according to the first embodiment of the present invention.

FIG. 9 is an operation explanatory diagram of the voice recognition result evaluation means in the voice recognition mode according to the first embodiment of the present invention.

FIG. 10 is an operation explanatory diagram of the voice recognition result evaluation means in the voice recognition mode according to the first embodiment of the present invention.

FIG. 11 is an operation explanatory diagram of the voice recognition result evaluation means in the voice recognition mode according to the first embodiment of the present invention.

FIG. 12 is an operation explanatory diagram of the voice recognition result evaluation means in the voice recognition mode according to the first embodiment of the present invention.

FIG. 13 is a diagram showing an example of an evaluation function used in the second embodiment of the present invention.

FIG. 14 is a block diagram showing a configuration of a voice recognition system according to a second embodiment of the present invention.

FIG. 15 is a block diagram showing a configuration in which reliability calculation means is shared by a plurality of voice recognition means.

[Explanation of symbols]

101 ... Voice data input device 102 ... Data processing device 103 ... Character data output device 104 ... Mode changer 105 ... Display device 106 ... Input device 111 ... Input voice data 112 ... Sample audio data 121 ... Voice data input means 131, 132 ... Voice recognition processing unit 133 ... Voice recognition result evaluation means 134 ... Evaluation function 135 ... Evaluation function creating means 136 ... Speech recognition result plotting means 137, 138 ... Voice recognition means 139, 140 ... Reliability calculation means

Claims (11)

[Claims] 1. A plurality of voice recognition processing units, each of which performs voice recognition on an input voice according to a unique recognition method, and outputs a voice recognition result and a reliability indicating its certainty, respectively, and the plurality of voice recognitions. The reliability output from each of the processing units is converted into an evaluation value which is a scale common to the plurality of speech recognition processing units by using an evaluation function prepared for each speech recognition processing unit, and is obtained by the conversion. A voice recognition system comprising: a voice recognition result evaluation means for selecting and outputting a voice recognition result from the plurality of voice recognition results based on a comparison result of a plurality of evaluation values. 2. The evaluation function includes a reference evaluation function corresponding to any one voice recognition processing unit of the plurality of voice recognition processing units and an evaluation function corresponding to another voice recognition processing unit, and In the evaluation function corresponding to the voice recognition processing unit of, even if the reliability value is the same, the voice recognition result is closer to the correct answer range than the voice recognition result of the one voice recognition processing unit. The speech recognition system according to claim 1, wherein is an evaluation function that obtains an evaluation value higher than the reference evaluation function, and an evaluation value that is lower than the reference evaluation function in the range of the reliability not close to the correct answer. 3. Means for presenting to the user the voice recognition results obtained by the plurality of voice recognition processing units for sample voice data, and means for displaying the reference evaluation function on a graph of reliability and evaluation values. A unit for plotting an evaluation value to be given to the reliability obtained by the other speech recognition processing unit on the graph according to an instruction from a user, and the other unit based on a set of plot points on the graph. The speech recognition system according to claim 2, further comprising means for creating the evaluation function for the speech recognition processing unit. 4. The voice recognition processing unit calculates voice recognition means for performing voice recognition on an input voice and outputting a voice recognition result, and reliability indicating the certainty of the voice recognition result of the voice recognition means. The speech recognition system according to claim 1, further comprising a reliability calculation means. 5. The voice recognition system according to claim 4, wherein the reliability calculation means is shared by a plurality of voice recognition processing units. 6. A) in parallel with a plurality of speech recognition processing units,
Voice recognition is performed on the input voice in accordance with its own recognition method, and a voice recognition result and a reliability indicating its certainty are respectively output; and b) output from each of the plurality of voice recognition processing units. A step of converting the reliability into an evaluation value which is a scale common to the plurality of speech recognition processing sections using an evaluation function prepared for each speech recognition processing section; and c) a plurality of evaluations obtained by the conversion. A voice recognition method comprising: selecting and outputting a voice recognition result from the plurality of voice recognition results based on a value comparison result. 7. The evaluation function includes a reference evaluation function corresponding to any one voice recognition processing unit of the plurality of voice recognition processing units and an evaluation function corresponding to another voice recognition processing unit, and In the evaluation function corresponding to the voice recognition processing unit of, even if the reliability value is the same, the voice recognition result is closer to the correct answer range than the voice recognition result of the one voice recognition processing unit. 7. The speech recognition method according to claim 6, wherein the evaluation value is higher than the reference evaluation function, and the evaluation value is lower than the reference evaluation function in the range of reliability not close to the correct answer. 8. The method further comprising the step of: d) generating an evaluation function corresponding to the other speech recognition processing section, the step d
D-1) presenting to the user the voice recognition results obtained by the plurality of voice recognition processing units with respect to the sample voice data, and d-2) the reference evaluation function on the graph of the reliability and the evaluation value. D-3) plotting an evaluation value to be given to the reliability obtained by the other speech recognition processing unit on the graph according to an instruction from the user, and d-4) Repeating the steps d-1 to d-3 for the number of prepared sample speech data, and d-5) the evaluation function for the other speech recognition processing unit based on a set of plot points on the graph. The method according to claim 7, further comprising the step of: 9. A plurality of voice recognition processing units, each of which performs voice recognition on an input voice according to a unique recognition method, and outputs a voice recognition result and a reliability indicating its certainty, respectively. The reliability output from each of the voice recognition processing units is converted into an evaluation value that is a common scale for the plurality of voice recognition processing units by using the evaluation function prepared for each voice recognition processing unit, and the conversion is performed. A voice recognition program that functions as a voice recognition result evaluation unit that selects and outputs a voice recognition result from the plurality of voice recognition results based on a comparison result of a plurality of obtained evaluation values. 10. The evaluation function includes a reference evaluation function corresponding to any one voice recognition processing unit of the plurality of voice recognition processing units and an evaluation function corresponding to another voice recognition processing unit, and In the evaluation function corresponding to the voice recognition processing unit of, even if the reliability value is the same, the voice recognition result is closer to the correct answer range than the voice recognition result of the one voice recognition processing unit. 10. The speech recognition program according to claim 9, which is an evaluation function that obtains an evaluation value higher than the reference evaluation function, and an evaluation value that is lower than the reference evaluation function in the range of reliability not close to the correct answer. . 11. The computer further comprises means for presenting to a user voice recognition results obtained by the plurality of voice recognition processing units for sample voice data, the reference evaluation function on a graph of reliability and evaluation value. Means for displaying, means for plotting an evaluation value given to the reliability obtained by the other speech recognition processing unit on the graph according to an instruction from a user, based on a set of plot points on the graph The speech recognition program according to claim 10, which is caused to function as a unit that creates the evaluation function for the other speech recognition processing unit.