JP7214841B2 - THRESHOLD ADJUSTMENT DEVICE, THRESHOLD ADJUSTMENT METHOD, AND RECORDING MEDIUM - Google Patents

Description

Embodiments of the present invention relate to a threshold adjustment device, a threshold adjustment method, and a recording medium.

2. Description of the Related Art Conventionally, there is known a speech recognition apparatus that, when input speech includes any of a plurality of predetermined recognition target words, outputs the recognition target word as a recognition result for the input speech. In this speech recognition device, the responsiveness to each recognition target word can be controlled by a threshold value. For example, if the distance in the feature space between the acoustic feature value calculated from the input speech and the acoustic feature value of any one of a plurality of recognition target words is equal to or less than the threshold set for the recognition target word , the recognition target word is output as the recognition result for the input speech. In this case, by setting an appropriate threshold for each of the plurality of recognition target words, the recognition target words contained in the input speech can be correctly recognized.

However, it is difficult to set an appropriate threshold value for each of a plurality of recognition target words so that the recognition target words are included in the input speech and not reacted to other recognition target words or noise. There is a demand for a mechanism that supports such threshold adjustment.

Japanese Patent Application Laid-Open No. 2018-072599

The problem to be solved by the present invention is to provide a threshold adjustment device, a threshold adjustment method, and a recording medium that support threshold adjustment so that an appropriate threshold can be set for each of a plurality of predetermined recognition target words. It is to be.

A threshold adjustment device according to an embodiment includes a speech recognition unit, an evaluation unit, and a display control unit. The speech recognition unit performs speech recognition. The evaluation unit inputs, to the speech recognition unit, a threshold list whose elements are a plurality of threshold values individually set for a plurality of predetermined recognition target words, and evaluation speech, and the speech recognition unit An evaluation value representing recognition accuracy for each of the plurality of recognition target words by the speech recognition unit using the threshold list is calculated based on the recognition result of the evaluation speech to be output. The display control unit causes the display device to display a threshold adjustment screen for adjusting a threshold corresponding to an arbitrary recognition target word among the plurality of recognition target words. The threshold adjustment screen presents the evaluation value calculated for each of the plurality of recognition target words by the evaluation unit together with the ideal value calculated in advance for each of the plurality of recognition target words. Contains a list of accuracy screens.

According to the threshold adjustment device having the above configuration, threshold adjustment can be supported so that an appropriate threshold can be set for each of a plurality of predetermined recognition target words.

FIG. 1 is a block diagram illustrating a functional configuration example of a threshold adjustment device according to an embodiment; FIG. 2 is a diagram showing an example of a keyword table. FIG. 3 is a diagram showing an example of an evaluation data table. FIG. 4 is a diagram showing an example of a threshold list table. FIG. 5 is a diagram showing an example of an evaluation result table. FIG. 6 is a diagram showing an example of an ideal value table. FIG. 7 is a flowchart illustrating an operation example of the threshold adjustment device of the embodiment; FIG. 8 is a diagram showing an example of an accuracy list screen. FIG. 9 is a diagram showing an example of an erroneous recognition analysis screen. FIG. 10 is a diagram showing an example of an initial evaluation result screen. FIG. 11 is a block diagram showing a functional configuration example of a threshold adjustment device according to a modification. FIG. 12 is a diagram showing an example of the accuracy list screen of the modification. FIG. 13 is a block diagram illustrating a hardware configuration example of the threshold adjustment device.

Hereinafter, specific embodiments of the present invention will be described in detail with reference to the drawings. In the following embodiments, an example of application to voice-triggered speech recognition that reacts only to several tens of specific keywords (recognition target words) will be described.

FIG. 1 is a block diagram showing a functional configuration example of the threshold adjustment device of this embodiment. The threshold adjustment device of this embodiment includes a speech recognition unit 1, an evaluation unit 2, and a display control unit 3, as shown in FIG.

A speech recognition unit 1 uses a speech recognition model 10 to perform speech recognition processing on an input speech. The speech recognition model 10 used in this embodiment includes an acoustic model for analyzing sound characteristics and a group of keywords predetermined as recognition objects of the speech recognition model 10 . A keyword group included in the speech recognition model 10 is registered in the keyword table 20 .

A threshold list is input to the speech recognition unit 1 in order to control the responsiveness of the speech recognition model 10 to individual keywords to be recognized. The threshold list is a list whose elements are a plurality of thresholds individually set for each keyword. In this embodiment, the distance in the feature space is used as an index of the similarity between the acoustic features. Since 1 outputs the keyword, the higher the threshold is set, the more likely it is to respond to the corresponding keyword. Here, the value range of the threshold is 0 to 1, and the distance in the feature space is normalized to a value of 0 to 1 and compared with the threshold. The threshold list is registered in the threshold list table 40. FIG.

The speech recognition unit 1 calculates, for example, an acoustic feature quantity from an input speech using an acoustic model, and selects an acoustic feature quantity that is closest in the feature space to the acoustic feature quantity of the input speech among the keywords to be recognized. Identify your keywords. Then, the speech recognition unit 1 recognizes the input speech when the distance in the feature space between the acoustic feature quantity of the input speech and the acoustic feature quantity of the specified keyword is equal to or less than a threshold value set for the specified keyword. As a result, the specified keyword is output.

The evaluation unit 2 inputs to the speech recognition unit 1 the threshold list registered as needed in the threshold list table 40 and the speech of the evaluation data registered in the evaluation data table 30 (evaluation speech), Based on the recognition result of the evaluation speech output by the speech recognition unit 1, an evaluation value representing the recognition accuracy for each keyword by the speech recognition unit 1 using the threshold list is calculated. The evaluation result by the evaluation unit 2 is recorded in the evaluation result table 50. FIG.

Also, the evaluation unit 2 calculates in advance an ideal evaluation value (ideal value) of each keyword through an initial evaluation described later. The ideal value for each keyword calculated in advance by the evaluation unit 2 is recorded in the ideal value table 60 .

The display control unit 3 causes an arbitrary display device to display a threshold adjustment screen for adjusting a threshold corresponding to an arbitrary keyword among the keyword group to be recognized by the speech recognition model 10 . The threshold adjustment screen includes an accuracy list screen 70 (see FIG. 8), an erroneous recognition analysis screen 80 (see FIG. 9), and an initial evaluation result screen 90 (see FIG. 10), which will be described later.

FIG. 2 is a diagram showing an example of the keyword table 20. As shown in FIG. The keyword table 20 is a table in which each keyword to be recognized by the speech recognition model 10 is registered. The keyword table 20, as shown in FIG. 2, includes an ID uniquely assigned to each keyword and the text of each keyword.

FIG. 3 is a diagram showing an example of the evaluation data table 30. As shown in FIG. The evaluation data table 30 is a table in which evaluation data used for evaluation by the evaluation unit 2 is registered. Evaluation data consists of speech and keywords contained in the speech. In other words, evaluation data is a combination of an evaluation speech containing a keyword predetermined as a recognition target of the speech recognition model 10 and a correct keyword to be recognized from the evaluation speech. The evaluation data table 30, as shown in FIG. 3, includes an ID uniquely assigned to each evaluation data, an audio file name, and a keyword ID. The keyword ID indicates the ID of the keyword included in the voice.

FIG. 4 is a diagram showing an example of the threshold list table 40. As shown in FIG. The threshold list table 40 is a table in which the threshold list input to the speech recognition unit 1 is registered. The threshold list table 40, as shown in FIG. 4, includes an ID uniquely assigned to each threshold list and a threshold list. As described above, the threshold list is a list whose elements are thresholds set individually for each keyword to be recognized by the speech recognition model 10 .

The threshold list registered in the threshold list table 40 includes the initial evaluation threshold list used in the initial evaluation by the evaluation unit 2, the initial threshold list obtained by the initial evaluation by the evaluation unit 2, and the threshold adjustment screen. There is an adjusted threshold list. In the example of FIG. 4, the threshold lists with IDs "1" and "2" are threshold lists for initial evaluation, the threshold list with ID "19" is the initial threshold list, and the threshold list with ID "20" is adjusted. It is assumed to be a post-threshold list.

The initial evaluation threshold list is a threshold list in which a common threshold is set for each keyword to be recognized by the speech recognition model 10, that is, a threshold list whose element values are all the same. A plurality of initial evaluation lists having different element values common to each keyword are registered in the threshold list table 40 .

The initial threshold list is a threshold list that uses the thresholds of the initial evaluation threshold list that were used when the ideal values were obtained by the initial evaluation as the thresholds that are individually set for each keyword. The initial threshold list is generated based on the results of the initial evaluation by the evaluation unit 2 and registered in the threshold list table 40 .

The post-adjustment threshold list is a threshold list obtained by the user adjusting the threshold corresponding to an arbitrary recognition target word using the threshold adjustment screen. The post-adjustment threshold list is generated as needed according to the operation using the threshold adjustment screen and registered in the threshold list table 40 .

FIG. 5 is a diagram showing an example of the evaluation result table 50. As shown in FIG. The evaluation result table 50 is a table in which evaluation results by the evaluation unit 2 are recorded. The evaluation result table 50 includes, as shown in FIG. 5, an ID uniquely assigned to each evaluation result, a threshold list ID, a keyword ID, an accuracy rate, a rejection rate, and an incorrectly recognized keyword. .

The threshold list ID is the ID of the threshold list input to the speech recognition unit 1 at the time of evaluation. The keyword ID is the ID of the keyword to be evaluated. Here, the recognition accuracy by the speech recognition unit 1 for each keyword is evaluated with respect to one threshold list, and the evaluation result for each combination of the threshold list and the keyword is recorded in one entry of the evaluation result table 50. .

The accuracy rate is defined as the correct answer when the recognition result of the speech recognition unit 1 for the evaluation speech containing the keyword to be evaluated is the keyword, and %). The rejection rate is calculated by (number of rejections/total number of data for evaluation including keywords to be evaluated)×100 (%) when the recognition result of the speech recognition unit 1 is no keyword. These accuracy rate and rejection rate are recorded in the evaluation result table 50 as evaluation values representing the accuracy of keyword recognition by the speech recognition unit 1 using the threshold list.

In addition to the above-mentioned correct answer and rejection, there is a case where the recognition result by the speech recognition unit 1 is a different keyword from the correct answer, that is, there is a case where another keyword is erroneously recognized. When the speech recognition unit 1 incorrectly recognizes another keyword that is different from the correct answer, the evaluation unit 2 counts the number of times of misrecognition for each misrecognized keyword. The set is recorded as an array in the erroneously recognized keyword of the evaluation result table 50 .

FIG. 6 is a diagram showing an example of the ideal value table 60. As shown in FIG. The ideal value table 60 is a table in which the ideal value for each keyword calculated by the evaluation unit 2 by the initial evaluation is recorded. The ideal value table 60 includes an ID uniquely assigned to each keyword, and the correct answer rate and rejection rate, which are ideal values for each keyword. As for the accuracy rate, the highest value among the accuracy rates obtained by the initial evaluation is recorded. The rejection rate is recorded as the rejection rate in the initial evaluation with the highest correct rate.

Next, the operation of the threshold adjustment device of this embodiment will be described along the flowchart of FIG. FIG. 7 is a flow chart showing an operation example of the threshold adjustment device of this embodiment.

First, data registration for the threshold adjustment device is performed (step S101). The data registered here are the speech recognition model 10 and evaluation data. Registration of the speech recognition model 10 is performed, for example, by uploading a JSON file containing the file name of the model and the character strings of each keyword to be recognized. The model file may be stored in the device in advance, or may be uploaded separately. The registered speech recognition model 10 is used by the speech recognition unit 1 to perform speech recognition processing. Each keyword to be recognized included in the speech recognition model 10 is registered in the keyword table 20 .

Registration of the evaluation data is performed by uploading a JSON file containing, for example, the file names of multiple evaluation voices and the character strings of keywords included in each evaluation voice. Each evaluation voice file may be stored in the device in advance, or may be uploaded separately. The file name and keyword ID of each uploaded evaluation voice are registered in the evaluation data table 30 .

When the data registration is completed, the evaluation unit 2 performs initial evaluation (step S102). In the initial evaluation, the evaluation unit 2 uses an initial evaluation threshold list that sets a common threshold for each keyword, and the speech (evaluation speech) of each evaluation data registered in the evaluation data table 30 for the initial evaluation. The threshold value in the threshold value list is repeatedly input to the speech recognition unit 1 while being changed at predetermined intervals. For example, the threshold values in the initial evaluation threshold list are changed from 0.1 to 0.95 in increments of 0.05. These initial evaluation threshold lists are sequentially input to the speech recognition unit 1 and registered in the threshold list table 40 .

The evaluation unit 2 evaluates the recognition accuracy for each keyword by the speech recognition unit 1 using the thresholds of the initial evaluation threshold list and the evaluation speech each time the speech recognition unit 1 receives the initial evaluation threshold list and the evaluation speech. The evaluation of each keyword corresponding to the thresholds in the initial evaluation threshold list is performed by comparing the recognition results output by the speech recognition unit 1 with the keywords included in the evaluation data, and calculating the accuracy rate/rejection rate (evaluation value) for each keyword. This is done by calculating That is, the evaluation unit 2 repeatedly inputs the initial evaluation speech and the evaluation speech with the threshold changed as described above to the speech recognition unit 1, and based on the recognition result of the evaluation speech output from the speech recognition unit 1, , the evaluation value for each keyword is repeatedly calculated. The evaluation results of the initial evaluation by the evaluation unit 2 are recorded in the evaluation result table 50 together with the ID of the threshold value list for initial evaluation registered in the threshold value list table 40 .

After carrying out the above initial evaluation, the evaluation unit 2 searches the evaluation result table 50 for the evaluation result with the best correct answer rate for each keyword, and uses the correct answer rate/rejection rate as ideal values for the keyword. It is recorded in the ideal value table 60 together with the ID. In addition, the evaluation unit 2 registers a threshold list whose elements are the threshold values for each keyword for which the best accuracy rate is obtained in the threshold value list table 40 as an initial threshold value list.

Next, the evaluation unit 2 performs evaluation using the initial threshold list (step S103). As in the initial evaluation, the evaluation unit 2 inputs the threshold list (here, the initial threshold list) and evaluation speech to the speech recognition unit 1, and based on the recognition results output by the speech recognition unit 1, correct answers for each keyword are determined. The rate/rejection rate (evaluation value) is calculated and the evaluation result is recorded in the evaluation result table 50 .

When the evaluation using the initial threshold list is completed, the display control unit 3 displays a threshold adjustment screen on an arbitrary display device such as the display device of the terminal used by the user, and the user's operation using this threshold adjustment screen The threshold value is adjusted accordingly (step S104).

First, the display control unit 3 causes the display device to display, for example, an accuracy list screen 70 as shown in FIG. 8 as the threshold adjustment screen. The accuracy list screen 70 presents the evaluation values calculated for each of the keywords to be recognized by the evaluation unit 2 together with the ideal values recorded in the ideal value table 60 .

In the accuracy list screen 70 shown in FIG. 8, a graph display 71 presents the evaluation value of each keyword in the threshold list (here, the initial threshold list) input to the speech recognition unit 1 together with the ideal value. In the graph display 71, each keyword to be recognized is arranged on the horizontal axis, and the evaluation value (correct answer rate/rejection rate) and the ideal value (correct answer rate/rejection rate) for each keyword are plotted at the corresponding positions on the vertical axis. ing. The white circles in the figure indicate the accuracy rate in the initial threshold list, the black circles indicate the accuracy rate as the ideal value, the white triangles in the figure indicate the rejection rate in the initial threshold list, and the black triangles indicate the ideal value. shows the rejection rate of The evaluation value for each keyword is acquired by searching the evaluation result table 50 using the ID of the initial threshold value list and the ID of the keyword as keys. The ideal value for each keyword is acquired by searching the ideal value table 60 using the keyword ID as a key.

For a given keyword, the evaluation value in the initial threshold list does not match the ideal value. This is due to interactions between keywords that occur due to setting thresholds. By referring to the graph display 71 of the accuracy list screen 70, the user can easily grasp the keyword (“on” in the example of FIG. 8) whose evaluation value in the initial threshold list is lower than the ideal value. can do.

When the user clicks and selects an arbitrary keyword on the accuracy list screen 70 and presses the "analysis of false positive detection" button 72, the threshold adjustment screen transitions to the false recognition analysis screen 80 shown in FIG. 9, for example. This misrecognition analysis screen 80 shows the number of times the speech recognition unit 1 has output another keyword as a recognition result for the evaluation speech containing the keyword selected on the accuracy list screen 70, that is, the speech recognition unit 1 is a screen that presents the number of misrecognitions by keyword 1 for each misrecognised keyword.

In the misrecognition analysis screen 80 shown in FIG. 9, a graph display 81 presents the number of misrecognitions for each misrecognised keyword. In the graph display 81, each keyword to be recognized is arranged on the horizontal axis, and the misrecognized keyword is represented by a bar graph extending to the corresponding position on the vertical axis for the number of misrecognition times. The number of incorrectly recognized keywords is obtained by searching the evaluation result table 50 using the ID of the keyword selected on the accuracy list screen 70 as a key. By referring to this erroneous recognition analysis screen 80, the user can easily grasp the keyword (“music” in the example of FIG. 9) for which the threshold value should be lowered in order to prevent erroneous recognition.

When the user presses a "return" button 82 on the misrecognition analysis screen 80, the threshold adjustment screen returns to the accuracy list screen 70 shown in FIG. Then, when the user clicks to select an arbitrary keyword on the accuracy list screen 70 and presses the "initial evaluation result" button 73, the threshold adjustment screen changes to an initial evaluation result screen 90 shown in FIG. 10, for example. This initial evaluation result screen 90 presents a list of evaluation values (correct answer rate/rejection rate) calculated by the evaluation unit 2 for each threshold in the initial evaluation threshold list used in the initial evaluation for the selected keyword. is.

In the initial evaluation result screen 90 shown in FIG. 10, a graph display 91 presents a list of evaluation values for each threshold in the initial evaluation threshold list used in the initial evaluation. In the graph display 91, the thresholds of each initial evaluation threshold list used in the initial evaluation are arranged on the horizontal axis, and the accuracy rate and the rejection rate corresponding to each threshold are represented by bar graphs extending to the corresponding positions on the vertical axis. there is The white bar graph in the figure indicates the accuracy rate for each threshold, and the black bar graph indicates the rejection rate for each threshold. The accuracy rate/rejection rate for each threshold is acquired by searching the evaluation result table 50 using the ID of the keyword and the ID of the threshold list for initial evaluation as keys. By referring to this initial evaluation result screen 90, the user can grasp the adjustable range of the threshold value. If the accuracy rate is sufficiently high and the rejection rate is sufficiently low (range of 0.45 to 0.75 in the example of FIG. 10), the threshold may be adjusted.

When the user clicks and selects the threshold he/she wants to set on the initial evaluation result screen 90 and presses the "Set Threshold" button 92, the threshold for the keyword selected on the accuracy list screen 70 is selected on the initial evaluation result screen 90. and the threshold value adjustment screen returns to the accuracy list screen 70 . When changing the threshold for another keyword grasped on the misrecognition analysis screen 80, the user selects the keyword on the accuracy list screen 70, selects the threshold on the initial evaluation result screen 90, and presses the "Set Threshold" button. By pressing 92, the threshold for that keyword can also be changed. That is, the user appropriately adjusts the threshold value set for each keyword to a desired value using the accuracy list screen 70, the misrecognition analysis screen 80, and the initial evaluation result screen 90 displayed as the threshold adjustment screen on the display device. be able to.

After performing the same operation for all the keywords for which the threshold is to be changed, when the user presses the "re-evaluate" button 74 on the accuracy list screen 70 (step S105: No), a new threshold reflecting the changed threshold is set. The list is registered in the threshold list table 40 as an adjusted threshold list. In addition, the operation flow of the threshold adjustment device returns to step S103, the evaluation unit 2 performs the evaluation again using the adjusted threshold list, and the evaluation result is recorded in the evaluation result table 50. FIG. After that, an accuracy list screen 70 presenting the evaluation value of each keyword in the post-adjustment threshold list together with the ideal value is displayed. At this time, the evaluation value of each keyword in the initial threshold list may also be presented without being erased. In this case, it is desirable that the evaluation values in the initial threshold list and the evaluation values in the adjusted threshold list can be clearly distinguished by, for example, different colors.

The user repeats the above operation until an appropriate evaluation result is obtained for each keyword to be recognized, and after confirming that an appropriate evaluation result has been obtained for each keyword, clicks the "end button" 75 on the accuracy list screen 70. is pressed (step S105: Yes). This completes a series of operations of the threshold adjustment device. At this time, the latest adjusted threshold list may be distributed to the specified external distribution destination as the optimal threshold list corresponding to the speech recognition model 10 registered in step S101. Also, the latest adjusted threshold list may be stored inside the threshold adjustment apparatus as an optimal threshold list corresponding to the speech recognition model 10 registered in step S101, and may be accessed from the outside as necessary. .

As described above in detail with specific examples, the threshold adjustment device of the present embodiment uses thresholds set individually for each of a plurality of keywords predetermined as recognition targets as elements. An evaluation value representing the recognition accuracy when using the threshold list is calculated for each keyword, and an accuracy list screen 70 presenting the calculated evaluation value together with the ideal value is displayed on the display device. Therefore, by referring to the accuracy list screen 70, the user can easily grasp the keyword whose threshold value should be changed.

Further, the threshold adjusting apparatus of the present embodiment displays an erroneous recognition analysis screen 80 and an initial evaluation result screen 90 on the display device when an arbitrary keyword is selected on the accuracy list screen 70 . Therefore, by referring to the erroneous recognition analysis screen 80, the user can easily grasp other keywords that are likely to be erroneously recognized with respect to the selected keyword, and by referring to the initial evaluation result screen 90, the selected keyword can be identified. The range in which the threshold can be changed can be easily grasped, and the threshold can be appropriately changed.

In this manner, the threshold adjustment device of the present embodiment can support threshold adjustment so that an appropriate threshold can be set for each of a plurality of keywords predetermined as recognition targets.

<Modification 1>
The threshold adjustment device described above may be configured to have a function of automatically adjusting the threshold. FIG. 11 is a block diagram showing a functional configuration example of the threshold adjustment device of this modification. The threshold adjustment device of this modification has a configuration in which an automatic adjustment unit 4 is added to the configuration shown in FIG.

In this modification, when the evaluation using the initial threshold list is completed, the display control unit 3 first causes the display device to display an accuracy list screen 70 as shown in FIG. 12 as the threshold adjustment screen. This accuracy list screen 70 has a configuration in which an "automatic adjustment" button 76 is added to the accuracy list screen 70 shown in FIG. When the user presses an "automatic adjustment" button 76 on the accuracy list screen 70, the automatic adjustment section 4 is activated.

After being activated, the automatic adjustment unit 4 first selects a keyword (first recognition target word) whose threshold is to be adjusted based on the difference between the evaluation value and the ideal value presented on the accuracy list screen 70. . For example, the automatic adjustment unit 4 selects the keyword whose evaluation value is the lowest from the ideal value. Based on the list of evaluation values for each threshold presented on the initial evaluation result screen 90 corresponding to the selected keyword, the automatic adjustment unit 4 determines the range in which the accuracy rate of the selected keyword does not decrease and the rejection rate does not increase. , raise the threshold corresponding to the keyword.

In addition, the automatic adjustment unit 4 adjusts the threshold for the keyword selected on the accuracy list screen 70 together with the keyword with the largest number of misrecognitions on the misrecognition analysis screen 80 corresponding to the keyword selected on the accuracy list screen 70. is selected as a keyword to be recognized (second recognition target word). Based on the list of evaluation values for each threshold presented on the initial evaluation result screen 90 corresponding to the selected keyword, the automatic adjustment unit 4 determines the range in which the accuracy rate of the selected keyword does not decrease and the rejection rate does not increase. , lower the threshold corresponding to that keyword.

The automatic adjustment unit 4 repeats the above operation a specified number of times, and registers a new threshold list reflecting the changed threshold in the threshold list table 40 as an adjusted threshold list. After that, the evaluation section 2 performs evaluation again using the adjusted threshold value list, and the evaluation result is recorded in the evaluation result table 50 in the same manner as in the above-described embodiment. Then, an accuracy list screen 70 presenting the evaluation value of each keyword in the post-adjustment threshold list together with the ideal value is displayed.

In this modified example, as described above, it is possible to automatically select a keyword for which the threshold value should be adjusted and adjust the threshold value corresponding to the keyword, thereby obtaining a unique effect of reducing the user's operation burden. .

<Modification 2>
In the above-described embodiment, it is assumed that any of the keywords to be recognized by the speech recognition model 10 is included in the evaluation speech. Noise speech containing other words may be added to the evaluation speech. When adding such a noise sound to the evaluation sound, the keyword ID of the corresponding entry in the evaluation data table 30 is recorded with "n/a" indicating no match. As for the recognition result of the speech recognition unit 1 for such noise speech, the correct result is that there is no keyword (rejection).

In this modification, in the evaluation after the initial evaluation (step S103 in FIG. 7), in addition to the evaluation speech containing the keyword, noise speech is input to the speech recognition unit 1, and speech recognition is performed for each keyword to be recognized. The unit 1 records the number of erroneous recognitions (erroneous reactions) of the keyword as the recognition result for the noise voice. Then, on the accuracy list screen 70 shown in FIG. 8, along with the evaluation value and ideal value for each keyword, the number of times noise voices are erroneously recognized is presented. By referring to the accuracy list screen 70, the user can easily grasp the keyword that is likely to be misrecognized with respect to noise speech, select the keyword as a threshold adjustment target, and perform the same operation as in the above-described embodiment. threshold adjustment can be performed appropriately.

As described above, in this modified example, it is possible to allow the user to accurately grasp keywords that are likely to be erroneously recognized with respect to noise speech, and to more effectively support threshold adjustment. be done.

<Modification 3>
In the above-described embodiment, the range in which the correct answer rate is sufficiently high and the rejection rate is sufficiently low in the initial evaluation result screen 90 corresponding to the keyword selected as the threshold adjustment target is set as the threshold adjustable range. However, this range is only a range derived based on the evaluation results of the evaluation unit 2 using the evaluation data, and if the threshold near the boundary of this range is set as the threshold after adjustment, more diverse voices can be input. In the actual speech recognition performed, it is also assumed that the recognition accuracy will decrease.

Therefore, in the initial evaluation result screen 90, the threshold after adjustment should be set such that the accuracy rate at the adjacent threshold is rapidly decreased or the rejection rate at the adjacent threshold is rapidly increased. In this case, a warning may be displayed on the initial evaluation result screen 90 or on the accuracy list screen 70 transitioned from the screen to indicate that there is a possibility that the accuracy may be lowered by the threshold adjustment. Judgment that the accuracy rate at the adjacent threshold is rapidly decreasing or the rejection rate at the adjacent threshold is rapidly increasing is, for example, the accuracy rate at the adjacent threshold is decreased by 10% or more. Alternatively, a condition such that the rejection rate at the adjacent threshold increases by 10% or more may be determined in advance.

As described in Modification 1 above, even in the case where the automatic adjustment unit 4 automatically adjusts the threshold value, the same warning display may be performed. That is, when the automatic adjustment unit 4 selects, as the adjusted threshold value, a threshold value at which the accuracy rate at the adjacent threshold value is sharply decreased, or at which the rejection rate at the adjacent threshold value is sharply increased. Alternatively, a warning may be displayed on the initial evaluation result screen 90 or the accuracy list screen 70, and the threshold may be adjusted when the user permits adjustment of the threshold.

As described above, in this modified example, a warning is displayed when there is a risk of the adjustment of the threshold value that would rather cause a decrease in recognition accuracy. Therefore, it is possible to more effectively support the adjustment of the threshold value. You can get the unique effect of being able to

<Modification 4>
In the above-described embodiment, an example of application to voice-triggered speech recognition that reacts only to a specific keyword was assumed, but the present invention is not limited to voice-triggered speech recognition, but also applies to continuous speech recognition that converts continuous speech into characters. can also be applied. In continuous speech recognition, it is possible to add a user word dictionary to enable recognition of technical terms in addition to general terms. Then, the ease of recognition and the difficulty of recognition of each word registered in the user word dictionary can be controlled by a threshold value.

Therefore, when the present invention is applied to continuous speech recognition, each word registered in the user word dictionary is treated in the same manner as the keyword in the above-described voice-triggered speech recognition, and the user word dictionary is processed in the same manner as in the above-described embodiment. can adequately support adjustment of the threshold set for each word in the .

The evaluation value displayed on the threshold adjustment screen in the threshold adjustment device of the above-described embodiment and modifications is based on the recognition result output by the speech recognition unit 1, and the accuracy rate for each keyword registered in the keyword table 20.・It is the evaluation value of each keyword for which the rejection rate (evaluation value) is calculated. That is, since the evaluation value for each keyword is displayed for the recognition results of voices uttered by a plurality of speakers, the evaluation target is a mixture of voice recognition results of a plurality of speakers.
However, the display method of the evaluation value is not limited to this. An evaluation value may be displayed. Similarly, on the misrecognition analysis screen, the misrecognition keyword for each speaker may be calculated, and the misrecognition keyword for each speaker and the number of appearances of the misrecognition keyword may be displayed.

<Supplementary explanation>
The threshold value adjustment devices of the above-described embodiments and modifications can be realized by using, for example, a general-purpose computer as basic hardware. That is, the function of each part of the above-described threshold adjustment device can be realized by causing one or more processors installed in a general-purpose computer to execute a program. At this time, the threshold adjustment device may be realized by pre-installing the above program in a computer, storing the above program in a computer-readable storage medium, or executing the above program via a network. It may be realized by distributing and installing this program on a computer as appropriate.

FIG. 13 is a block diagram showing a hardware configuration example of the above threshold adjustment device. For example, as shown in FIG. 13, the threshold adjustment device includes a processor 101 such as a CPU (Central Processing Unit), a memory 102 such as a RAM (Random Access Memory) or a ROM (Read Only Memory), and a HDD (Hard Disk Drive). , a storage device 103 such as an SSD (Solid State Drive), a display device 106 such as a liquid crystal panel, and an input device 107 such as a keyboard or pointing device. It has a hardware configuration as a general computer including a communication I/F 105 and a bus 108 connecting these units.

When the above-described threshold adjustment device is realized by the hardware configuration shown in FIG. 13, for example, the processor 101 uses the memory 102 to read and execute a program stored in the storage device 103 or the like, thereby obtaining the above-described voice Functions of each unit such as the recognition unit 1, the evaluation unit 2, the display control unit 3, and the automatic adjustment unit 4 can be realized. Further, the speech recognition model 10, the keyword table 20, the evaluation data table 30, the threshold list table 40, the evaluation result table 50, and the ideal value table 60 are stored in the memory 102, the storage device 103, or the like, and can be stored as needed. It can be read out and used for processing.

It should be noted that some or all of the functions of each part of the threshold adjustment device described above are implemented by dedicated hardware such as ASIC (Application Specific Integrated Circuit) or FPGA (Field-Programmable Gate Array) (not a general-purpose processor but a dedicated hardware). processor). Further, the configuration may be such that a plurality of processors are used to implement the functions of the respective units described above. Further, the above-described threshold adjustment device is not limited to being implemented by a single computer, but can be implemented by distributing the functions to a plurality of computers.

Although the embodiment of the present invention has been described above, this embodiment is presented as an example and is not intended to limit the scope of the invention. This novel embodiment can be embodied in various other forms, and various omissions, replacements, and modifications can be made without departing from the scope of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are included in the scope of the invention described in the claims and equivalents thereof.

1 speech recognition unit 2 evaluation unit 3 display control unit 4 automatic adjustment unit 10 speech recognition model 20 keyword table 30 evaluation data table 40 threshold list table 50 evaluation result table 60 ideal value table

Claims (13)

a speech recognition unit that performs speech recognition;
inputting a threshold value list whose elements are a plurality of threshold values individually set for a plurality of predetermined recognition target words and evaluation speech to the speech recognition unit, and outputting the evaluation from the speech recognition unit; an evaluation unit that calculates an evaluation value representing the recognition accuracy of each of the plurality of recognition target words by the speech recognition unit using the threshold list based on the recognition result of the speech for the target speech;
a display control unit that causes a display device to display a threshold adjustment screen for adjusting a threshold corresponding to an arbitrary recognition target word among the plurality of recognition target words,
The threshold adjustment screen presents the evaluation value calculated for each of the plurality of recognition target words by the evaluation unit together with the ideal value calculated in advance for each of the plurality of recognition target words. including an accuracy list screen to
The evaluation unit calculates, for each of the plurality of recognition target words, the number of times the speech recognition unit incorrectly recognizes another recognition target word from the evaluation speech containing the recognition target word. count for each target word,
The threshold adjustment screen is an erroneous recognition analysis screen for presenting the number of erroneous recognitions counted by the evaluation unit for each erroneously recognized recognition target word for the recognition target word selected from the plurality of recognition target words. further including,
Threshold adjuster. The threshold adjustment screen adjusts an initial evaluation threshold list for setting a threshold common to the plurality of recognition target words and the evaluation speech for the recognition target word selected from the plurality of recognition target words. further comprising an initial evaluation result screen for presenting a list of evaluation values repeatedly calculated by the evaluation unit when input is repeatedly made to the speech recognition unit while changing a common threshold in the threshold list for initial evaluation at predetermined intervals; Item 2. The threshold adjustment device according to item 1. The ideal value is determined by the evaluation unit when the initial evaluation threshold list and the evaluation speech are repeatedly input to the speech recognition unit while changing the common threshold in the initial evaluation threshold list at predetermined intervals. 3. The threshold adjustment device according to claim 2 , wherein the evaluation value represents the highest recognition accuracy by the speech recognition unit among evaluation values that are repeatedly calculated. The threshold list uses a common threshold in the initial evaluation threshold list when the ideal values of the respective recognition target words are obtained as the plurality of thresholds individually set for the plurality of recognition target words. 4. The threshold adjuster of claim 3 , comprising an initial threshold list. 2. The threshold value adjustment apparatus according to claim 1, wherein said threshold value list includes an adjusted threshold value list in which threshold values corresponding to arbitrary recognition target words are adjusted using said threshold value adjustment screen. an automatic adjustment unit that selects a first recognition target word whose threshold value is to be adjusted from among the plurality of recognition target words based on the difference between the evaluation value and the ideal value presented on the accuracy list screen; 3. The threshold adjuster of claim 2 , comprising: 3. The automatic adjustment unit adjusts the threshold corresponding to the first recognition target word based on a list of evaluation values presented on the initial evaluation result screen corresponding to the first recognition target word. 7. The threshold adjustment device according to 6 . The automatic adjustment unit further selects a second recognition target word whose threshold is to be adjusted based on the number of misrecognitions presented on the misrecognition analysis screen corresponding to the first recognition target word. 7. A threshold adjuster according to claim 6 . 3. The automatic adjustment unit adjusts the threshold corresponding to the second recognition target word based on a list of evaluation values presented on the initial evaluation result screen corresponding to the second recognition target word. 9. The threshold adjustment device according to 8 . 2. The threshold adjustment device according to claim 1, wherein said evaluation speech includes noise speech that does not include any of said plurality of recognition target words. The display control unit adjusts the threshold specified as the threshold after adjustment based on the list of evaluation values presented on the initial evaluation result screen corresponding to the recognition target word whose threshold is to be adjusted to meet a predetermined condition. 3. The threshold adjustment device according to claim 2 , wherein a warning is displayed on the threshold adjustment screen when it is determined that it applies. A threshold value list including a plurality of threshold values individually set for a plurality of predetermined recognition target words as elements and evaluation speech are input to a speech recognition unit that performs speech recognition, and the speech recognition unit outputs the threshold list. an evaluation step of calculating an evaluation value representing the recognition accuracy of each of the plurality of recognition target words by the speech recognition unit using the threshold list, based on the recognition result of the evaluation speech;
a display control step of causing a display device to display a threshold adjustment screen for adjusting a threshold corresponding to an arbitrary recognition target word among the plurality of recognition target words,
The threshold adjustment screen presents the evaluation value calculated for each of the plurality of recognition target words in the evaluation step together with the ideal value calculated in advance for each of the plurality of recognition target words. including an accuracy list screen to
In the evaluation step, for each of the plurality of recognition target words, the number of times the speech recognition unit incorrectly recognizes another recognition target word from the evaluation speech containing the recognition target word is counted. count for each target word,
The threshold adjustment screen is an erroneous recognition analysis screen for presenting the number of erroneous recognitions counted in the evaluation step for each of the erroneously recognized target words selected from among the plurality of target words to be recognized. further including,
Threshold adjustment method. A computer-readable recording medium storing a computer program, wherein the computer program causes the computer to:
a speech recognition unit that performs speech recognition;
inputting a threshold value list whose elements are a plurality of threshold values individually set for a plurality of predetermined recognition target words and evaluation speech to the speech recognition unit, and outputting the evaluation from the speech recognition unit; an evaluation unit that calculates an evaluation value representing the recognition accuracy of each of the plurality of recognition target words by the speech recognition unit using the threshold list based on the recognition result of the speech for the target speech;
a display control unit that causes a display device to display a threshold adjustment screen for adjusting a threshold corresponding to an arbitrary recognition target word among the plurality of recognition target words,
The threshold adjustment screen presents the evaluation value calculated for each of the plurality of recognition target words by the evaluation unit together with the ideal value calculated in advance for each of the plurality of recognition target words. including an accuracy list screen to
The evaluation unit calculates, for each of the plurality of recognition target words, the number of times the speech recognition unit incorrectly recognizes another recognition target word from the evaluation speech containing the recognition target word. count for each target word,
The threshold adjustment screen is an erroneous recognition analysis screen for presenting the number of erroneous recognitions counted by the evaluation unit for each erroneously recognized recognition target word for the recognition target word selected from the plurality of recognition target words. further including,
recording medium.

Images