JP3718088B2 - Speech recognition correction method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speech recognition correction method for correcting an error portion when a part of a character string obtained as a result of speech recognition processing has an error.
[0002]
[Prior art]
As a method of giving various operation instructions to recent in-vehicle devices, such as navigation devices and audio devices, in addition to the method in which the user presses various keys on the operation panel or remote control (remote control) unit, it can be used. There is a method of recognizing the contents of operation voices issued by a person. According to the method for recognizing the contents of the operation voice, the user does not need to remember the arrangement of various operation keys, and the user does not have to operate the keys while the vehicle vibrates while driving. The operation can be simplified. In addition, since the method for recognizing the contents of the operation voice is becoming possible with a relatively high accuracy voice recognition process especially with recent increase in processor speed, the operation instruction method using voice recognition is It is also widely used for in-vehicle devices.
[0003]
By the way, when operating sound that is subject to speech recognition is collected with a microphone, road noise and engine noise are simultaneously collected along with the operating sound, so compared to when collecting sound in a quiet environment. The recognition rate of voice recognition decreases. Therefore, it is usually necessary to correct the content of misrecognized speech. If there is an error in a part of the character string obtained as a recognition result, the user utters the same content again with a slightly different tone, for example, clearly pronounced. A second speech recognition process is performed. In this way, by repeating the voice recognition process several times for the same voice, a character string as a correct recognition result can be finally obtained.
[0004]
[Problems to be solved by the invention]
By the way, the conventional speech recognition result correcting method described above is repeated many times when a user utters the same speech and repeats speech recognition processing on the content, and utters a word with a poor recognition rate. There was a problem that it took time and effort to correct the same voice. For example, if you can display a character string as a recognition result, move the cursor to the correction location in the character string, and input correction data directly from the keyboard etc., such inconvenience will not occur. Is performed using voice recognition processing, it is not easy for a user to input a key such as 50 sounds. Therefore, a method for efficiently correcting a recognition result by inputting voice is desired.
[0005]
The present invention has been made in view of such a point, and an object of the present invention is to provide a speech recognition correction method that can efficiently correct misrecognized characters.
[0006]
[Means for Solving the Problems]
In order to solve the above-described problem, in the speech recognition correction method of the present invention, the first speech recognition means performs speech recognition processing on the input speech, and character string data as a recognition result including a plurality of characters. And the second speech recognition means performs speech recognition processing on the correction speech input in order to correct the erroneous recognition location in the obtained character string data. Correction data consisting of n characters is acquired. Then, a part of the character string data obtained by the first voice recognition means described above is replaced with correction data obtained by the second voice recognition means described above. When correcting a misrecognized part, correction data consisting of one or more n characters is acquired by performing the voice recognition process by the second voice recognition device described above, so correction is performed as in the past. It is not necessary to operate an operation unit such as a keyboard when inputting business data, and it is possible to simplify the operation when correcting a misrecognized portion. In addition, since only the voice necessary for correction is input corresponding to the misrecognized part and the recognition result is corrected, the voice to be recognized as before is changed many times by changing the tone of the voice. This eliminates the need for input and can correct the recognition result efficiently.
[0007]
Further, the correction candidate notification means notifies the correction portion that can be replaced with the correction data in the character string data, and the correction portion selection means notifies the user of one or more of the notified correction portions. It is desirable to select one according to the operation. For determining whether or not it is a correction part that can be replaced by the correction data, for example, checking whether or not the character string data that has been replaced exists as a commonly used word (word) Judgment may be made by Also, if only speech corresponding to a specific operation instruction or the like is targeted for speech recognition, it is determined whether or not the replaced character string is included in words corresponding to the specific operation instruction. May be. In this way, information that does not need to be notified is notified by notifying only the correction part that can replace part of the character string data with correction data, and not reporting the correction part that cannot be replaced. Can be prevented, and it is possible to prevent the notification regarding the correction part from becoming complicated. In addition, by selecting one of the notified correction locations that is suitable as the correction location according to the user's operation, the correction content according to the user's intention can be reliably reflected. it can.
[0008]
The speech recognition dictionary includes a speech recognition dictionary used for the speech recognition processing by the first speech recognition unit described above, and based on the corrected character string data in which the correction data is replaced by the character string replacement unit. It is desirable to update the contents. The fact that the character string data has been modified means that the recognition rate for the voice corresponding to the original character string data is low, so the content of the voice recognition dictionary corresponding to this voice is changed to the corrected character string data. By updating based on the recognition rate, the recognition rate can be improved.
[0009]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, a voice recognition device according to an embodiment to which a voice recognition correction method of the present invention is applied and a vehicle-mounted system including the voice recognition device will be described with reference to the drawings.
[0010]
[First Embodiment]
FIG. 1 is a diagram illustrating a configuration of an in-vehicle system according to the first embodiment. The in-vehicle system shown in FIG. 1 performs various operations such as navigation processing in response to various input instructions given by voice from the user, and performs voice recognition processing on the voice uttered by the user. A voice recognition device 1 that identifies a character string corresponding to the voice uttered by the user, and displays a map around the vehicle position by detecting the vehicle position, or a route to the destination selected by the user A navigation device 2 that performs search, route guidance, and the like; a display device 3 that displays a voice recognition result output from the voice recognition device 1; a map image around the vehicle position output from the navigation device 2; and a voice recognition device 1 and an audio unit 4 for outputting various guidance voices output from the navigation device 2.
[0011]
The voice recognition device 1 described above includes a microphone 10, a voice recognition unit 12, a voice recognition dictionary 14, a correction voice recognition unit 16, a correction dictionary 18, a character string replacement unit 20, a correction candidate search unit 22, and a correction candidate storage unit 24. , A recognition result determination unit 26, an image generation unit 28, and a dictionary update unit 30.
[0012]
The microphone 10 collects sound uttered by the user and converts it into an electrical signal. The voice recognition unit 12 searches the voice recognition dictionary 14 to perform voice recognition processing on the voice signal input via the microphone 10 and specifies character string data corresponding to the voice uttered by the user. .
[0013]
The correction voice recognition unit 16 performs a voice recognition process on the correction voice by searching the correction dictionary 18 when a correction voice necessary for correcting the voice recognition result is input. The character data corresponding to the correction voice (hereinafter referred to as “correction data”) is specified. In the present embodiment, it is assumed that one sound character is considered as the correction sound.
[0014]
The character string replacement unit 20 acquires the character string data specified by the voice recognition unit 12 and the correction data specified by the correction voice recognition unit 16, and converts each character data included in the character string data to the correction data. Generate character string data replaced with. Specifically, for example, if the character string data specified by the voice recognition unit 12 is “Takizawa” and the correction data specified by the correction voice recognition unit 16 is “ka”, the character string The replacement unit 20 replaces the character data “K” included in the character string data “Takizawa” with the correction data “K”, and the characters included in the character string data “Takizawa”. Character string data “Takazawa” that replaces data “ki” and correction data “ka”, and character string that replaces character data “za” and correction data “ka” included in character string data “Takizawa” Data “Takikawa” and character string data “Takizawa” in which the character data “wa” included in the character string data “Takizawa” and the correction data “ka” are replaced are generated.
[0015]
The correction candidate search unit 22 searches the speech recognition dictionary 14, and each character string data generated by the character string replacement unit 20 (each character string data after the replacement process) is converted into a character string to be subjected to speech recognition. It is checked whether or not it is registered in the recognition dictionary 14. If each character string data is registered in the speech recognition dictionary 14, the correction candidate search unit 22 stores the character string data in the correction candidate storage unit 24 as “correction candidate data”.
[0016]
The recognition result determination unit 26 notifies the user of the voice recognition result obtained by the voice recognition unit 12, determines whether the voice recognition result is appropriate based on an instruction input given by the user, and outputs the result. If the voice recognition result is incorrect, the recognition result determination unit 26 notifies the user of the correction candidate of the voice recognition result extracted by the correction candidate search unit 22 and gives an instruction given by the user. Based on the input, the optimum correction result is determined and output. The speech recognition result or the optimum correction result determined by the recognition result determination unit 26 is output to the navigation device 2 as an output from the speech recognition device 1.
[0017]
The image generation unit 28 generates image data for displaying various images such as a speech recognition result and a correction candidate of the speech recognition result based on an instruction given from the recognition result determination unit 26. A specific display example of the image generated and displayed by the image generation unit 28 will be described later.
[0018]
The dictionary update unit 30 checks whether or not the speech recognition result has been corrected based on the information output from the recognition result determination unit 26. If the correction has been performed, the dictionary update unit 30 The contents of the data stored in the recognition dictionary 14 are updated.
[0019]
The voice recognition unit 12 and the voice recognition dictionary 14 described above are the first voice recognition unit, the correction voice recognition unit 16 and the correction dictionary 18 are the second voice recognition unit, and the character string replacement unit 20 is the character string replacement unit. Further, the correction candidate search unit 22, the correction candidate storage unit 24, the recognition result determination unit 26, the image generation unit 28, and the display device 3 are used as the correction candidate notification unit, the recognition result determination unit 26 is used as the correction location selection unit, and the dictionary update unit. Reference numeral 30 corresponds to the dictionary updating means.
[0020]
The in-vehicle system according to the present embodiment has the above-described configuration. Next, the speech recognition process performed in the speech recognition apparatus 1 will be described in detail. 2 and 3 are flowcharts showing the operation of the speech recognition process performed in the speech recognition apparatus 1. FIG. For example, consider a case in which the name of a departure place or a destination at the time of performing a route search is input to the navigation device 2 by voice. A description will be given of an operation when a recognition error is made and the character string “Kakizawa” is corrected.
[0021]
The voice recognition unit 12 determines whether or not voice input has been performed by the user via the microphone 10 (step 100). Until a voice input is performed, a negative determination is made in step 100 and a standby state is entered. When voice input is performed, the voice recognition unit 12 performs voice recognition processing using the voice recognition dictionary 14, and specifies character string data corresponding to the voice uttered by the user (step 101). In the example described above, “Takizawa” is specified as the character string data. The voice recognition result obtained by the voice recognition unit 12 is output to the recognition result determination unit 26.
[0022]
The recognition result determination unit 26 instructs the image generation unit 28 to generate image data for displaying the voice recognition result, and also generates audio data for outputting the voice recognition result by voice to generate the audio unit 4. Output to. As a result, the voice recognition result is displayed on the screen of the display device 3, and the voice corresponding to the voice recognition result is output from the audio unit 4 to notify the user of the voice recognition result (step 102). .
[0023]
FIG. 4 is a diagram illustrating a display example of a voice recognition result. As shown in FIG. 4, on the screen of the display device 3, an indication that the speech recognition result is “Takizawa”, and if the speech recognition result is incorrect, the user is prompted to input correction speech. Display. Further, in parallel with the display as shown in FIG. 4, for example, from the audio unit 4, is “Are the voice recognition result“ Takizawa ”? Is announced.
[0024]
When the voice recognition result is notified to the user as described above, the recognition result determination unit 26 determines whether or not a certain time (for example, 30 seconds) has passed (step 103), and the certain time. If it has not elapsed, it is determined whether or not voice input has been performed (step 104). When a predetermined time has elapsed without voice input, an affirmative determination is made in step 103, and the recognition result determination unit 26 outputs the voice recognition result to the navigation device 2 (step 105). In addition, when a voice input is performed by the user before the predetermined time has elapsed, an affirmative determination is made in step 104, and the recognition result determination unit 26 determines whether or not the input voice is a single character. (Step 106). For example, whether or not the input voice is a single character is always monitored by the correction voice recognition unit 16, and the recognition result determination unit 26 inputs based on information obtained from the correction voice recognition unit 16. It is determined whether or not the voice is a single character.
[0025]
If the input voice is not a single character, a negative determination is made in step 106, and the recognition result determination unit 26 determines that the input voice is not a voice for correcting the recognition result but a voice related to the next operation instruction or the like. The voice recognition result obtained in step 101 is output to the navigation device 2 (step 107). Thereafter, the process returns to step 101, and the operation after the voice recognition process is performed on the input voice related to the next operation instruction or the like (the input voice subjected to the determination process in step 104).
[0026]
If the input voice is a single character, the correction voice recognition unit 16 determines that the input voice is a voice for correcting the recognition result, and sets the correction dictionary 18 for the voice. Using this, voice recognition processing is performed, and correction data corresponding to the voice is specified (step 108). In the example described above, since “T” is input by the user in order to correct “Takizawa” that has been misrecognized to “Kakizawa”, voice recognition processing is performed on this voice “K”. Corresponding correction data is identified.
[0027]
Next, the character string replacement unit 20 acquires character string data to be corrected from the voice recognition unit 12 and also acquires correction data from the correction voice recognition unit 16 (step 109). After that, the character string replacement unit 20 replaces the first character data included in the character string data to be corrected with the correction data, and outputs it to the correction candidate search unit 22 (step 110). In the above-described example, the character string data “Kakizawa” generated by replacing the first character data “ta” of the character string data “Takizawa” to be corrected with the correction data “ka” is output.
[0028]
The correction candidate search unit 22 searches the speech recognition dictionary 14 and the character string data (character string data after replacement) output from the character string replacement unit 20 is stored in the speech recognition dictionary 14 as a character string to be subjected to speech recognition. It is checked whether or not it is registered (step 111). If the replaced character string data is registered in the speech recognition dictionary 14, the correction candidate search unit 22 stores the character string data as correction candidate data in the correction candidate storage unit 24 (step 112). If the replaced character string data is not registered in the speech recognition dictionary 14, a negative determination is made in step 111. In this case, the correction candidate search unit 22 performs the correction candidate data shown in step 112. The storage operation is not performed.
[0029]
Next, the character string replacement unit 20 checks whether or not the character string replacement processing is completed by checking whether or not the last character data included in the character string data to be corrected has been replaced with correction data. Determination is made (step 113). If the replacement process has not been completed, a negative determination is made in step 113, and the character string replacement unit 20 replaces the character data to be subjected to the replacement process next with the correction data (step 114). In the example described above, the second character data “ki” of the character string data “Takizawa” is replaced with the correction data “ka”, and the character string data “Takazawa” is output. In the second and subsequent processing, the character data “Takikawa” in which the third character data “za” of the character string data “Takizawa” is replaced with the correction data “ka”, and the character string data “ Character string data “Takizawa” in which the last character data “Wa” of “Takizawa” is replaced with correction data “ka” is output. When the replacement process is performed, the process returns to step 111, and the operations after the determination as to whether or not the character string data after the replacement is stored in the speech recognition dictionary 14 are repeated.
[0030]
When the replacement process is completed, an affirmative determination is made in step 113, and then the recognition result determination unit 26 determines whether correction candidate data is stored in the correction candidate storage unit 24 (step 115). If correction candidate data has been stored, the recognition result determination unit 26 reads the correction candidate data, sends an instruction to the image generation unit 28, and displays the correction candidates (step 116). FIG. 5 is a diagram illustrating a display example of correction candidates, and character string data corresponding to each of “Kakizawa”, “Takazawa”, “Takikawa”, and “Takizawa” is stored as correction candidate data. The example of a display when it was done is shown. As shown in FIG. 5, numbers are added to the respective correction candidates such as “1: Kakizawa”, “2: Takazawa”, “3: Takikawa”, “4: Takizawa”. In addition to the display, a display prompting the user to select a number corresponding to the optimum correction result is performed.
[0031]
When a user selects a suitable correction result from the displayed correction candidates, the recognition result determination unit 26 directs the correction candidate data corresponding to the selected correction candidate to the navigation device 2 as a voice recognition result. (Step 117). In the example described above, it is assumed that the first “Kakizawa” is selected by the user. As a method for selecting correction candidates by the user, the user may be allowed to select a number added to each correction candidate via a predetermined operation unit (not shown). It is also possible to have the person input the number by voice and perform a voice recognition process on this to select a correction candidate.
[0032]
Next, the dictionary update unit 30 acquires information related to the corrected speech recognition result from the recognition result determination unit 26, and updates the contents of the speech recognition dictionary 14 in accordance with the speech recognition result (step 118). After the content of the speech recognition dictionary 14 is updated, the process returns to step 100 and the operations after the determination as to whether or not speech input has been performed are repeated.
[0033]
Further, in the above-described step 115, if no correction candidate is stored, the correction candidate cannot be extracted because the character string data to be corrected includes two or more erroneous recognition locations. The recognition result determination unit 26 sends an instruction to the image generation unit 28 to inform the user that the speech recognition result has not been corrected, and to prompt the user to perform voice input again. Is displayed (step 119). When the error notification is performed, the process returns to step 100, and the operations after the determination as to whether or not the voice input has been performed are repeated.
[0034]
As described above, in the speech recognition apparatus 1 according to the present embodiment, when correcting the erroneous recognition portion included in the character string data obtained by the speech recognition unit 12, the speech recognition unit 16 performs speech recognition processing. Thus, the correction data is acquired, and the correction data is replaced with a part of the character string data obtained by the voice recognition unit 12 to correct the erroneous recognition portion. Accordingly, there is no need to input the voice corresponding to the character string data to be recognized as many times as before by changing the tone of the voice, and the recognition result can be corrected efficiently. In addition, since the correction data is input by voice input, it is not necessary to use an operation unit such as a keyboard, and the operation can be simplified. Further, when a part of the character string data obtained by the speech recognition unit 12 is replaced with the correction data, by checking whether or not the replaced character string data is registered in the speech recognition dictionary 14, Only registered character string data is extracted and notified to the user. In other words, the user is notified of only a correction location where a part of the character string data can be replaced with the correction data, and it is possible to prevent the content of notification regarding the correction location from becoming complicated. In addition, since a suitable correction location is selected in response to the user's operation, the correction according to the user's intention can be reliably performed. In addition, when the recognition result is corrected, the content of the speech recognition dictionary 14 is updated in accordance with the correction result, so that the recognition rate can be improved by repeatedly performing the speech recognition processing. It also has the advantage of being able to.
[0035]
[Second Embodiment]
By the way, in the first embodiment described above, the case where only one character is erroneously recognized included in the speech recognition result has been described. However, even when a plurality of characters are erroneously recognized by the same processing procedure, correction is also performed by speech. It can be performed.
[0036]
6 and 7 are flowcharts showing a modification of the operation of the speech recognition process performed in the speech recognition apparatus 1 shown in FIG. 1, and n or more of one character or more in the character string data obtained by the speech recognition process. An operation procedure for correcting characters is shown. The operation procedure shown in FIGS. 6 and 7 is the same as the operation procedure shown in FIGS. 2 and 3 except that the operation in step 106 for determining whether or not the input voice is one character is performed by a correction instruction by the user. The operation in step 106A for determining whether or not there is a replacement, and the operation in step 110 in which the first character data included in the character string data to be corrected is replaced with the correction data is used to correct the first n characters of the character string data. The difference is that the operation is replaced with the operation of step 110A.
[0037]
That is, in the first embodiment described above, by determining whether or not the input voice is a single character, it is distinguished whether the input voice is a correction voice or a next normal operation voice. Therefore, such a distinction cannot be made when a plurality of characters are allowed as a correction sound. For this reason, the voice input after some correction instruction is given by the user is handled as a correction voice.
[0038]
If the user does not give a correction instruction prior to the input of the correction voice, a negative determination is made in the determination operation of step 106A. Next, the recognition result determination unit 26 determines that the input voice is a recognition result. It is determined that the sound is not a sound for correction but a sound related to the next operation instruction or the like, and the operation after Step 107 for outputting the recognition result to the navigation device 2 is performed.
[0039]
On the other hand, if the user gives a correction instruction prior to the input of the correction voice, an affirmative determination is made in the determination operation of step 106A, and then the correction voice recognition unit 16 receives the input voice. Is the speech for correcting the recognition result, and the operations after step 108 for performing speech recognition processing on this speech are performed.
[0040]
In addition, there are several possible methods for giving a correction instruction by the user. For example, it is determined that a correction instruction is given when a specific key of an operation unit (not shown) is pressed, or a correction instruction is given when a user speaks a specific word toward the microphone 10. The case where it determines with a thing etc. can be considered.
[0041]
As described above, when correcting the character string data as the recognition result, it is possible to correct a plurality of characters (may be one character) by causing the user to display some intention.
[0042]
Further, the fixed number n of voices for correction input after a correction instruction is given by the user may be a preset fixed value, but may be set freely each time. For example, the number n of correction characters is not particularly set, and the number of characters of correction data obtained by the speech recognition process for the correction speech may be adopted as the value of n. In this case, since the user determines the error location of the character string as the recognition result and utters the correction voice having the optimum number of characters to the microphone 10 each time, the recognition result can be corrected more efficiently. It can be carried out.
[0043]
In addition, this invention is not limited to the said embodiment, A various deformation | transformation implementation is possible within the range of the summary of this invention. For example, in the above-described embodiment, the case where a predetermined instruction is input to the navigation device 2 using the voice recognition device 1 to which the present invention is applied in the in-vehicle system has been described. You may make it perform a predetermined instruction | indication input with respect to other vehicle equipment, such as an audio apparatus. The present invention can also be applied to various systems other than the in-vehicle system, for example, when various instructions are input to various computers such as personal computers and workstations.
[0044]
【The invention's effect】
As described above, according to the present invention, when a part of the character string obtained as a result of the speech recognition process has a misrecognized portion, a correction voice having a predetermined number of characters is input, and the correction voice is Since the correction result is obtained by performing voice recognition processing to acquire correction data and replacing the erroneous recognition location included in the character string data with this correction data, the character string that you want to be recognized as before The recognition result can be corrected efficiently without the trouble of inputting the voice corresponding to the data many times by changing the tone of the voice. In addition, since the correction data is input by voice input, it is not necessary to use an operation unit such as a keyboard, and the operation can be simplified.
[Brief description of the drawings]
FIG. 1 is a diagram illustrating a configuration of an in-vehicle system according to a first embodiment.
FIG. 2 is a flowchart showing the operation of speech recognition processing performed in the speech recognition apparatus.
FIG. 3 is a flowchart showing an operation of speech recognition processing performed in the speech recognition apparatus.
FIG. 4 is a diagram illustrating a display example of a voice recognition result.
FIG. 5 is a diagram illustrating a display example of correction candidates.
FIG. 6 is a flowchart showing the operation of a modified example of the speech recognition process performed in the speech recognition apparatus.
FIG. 7 is a flowchart showing the operation of a modified example of the speech recognition process performed in the speech recognition apparatus.
[Explanation of symbols]
1 Voice recognition device
2 Navigation device
3 Display device
4 Audio section
10 Microphone
12 Voice recognition unit
14 Speech recognition dictionary
16 Voice recognition unit for correction
18 Correction Dictionary
20 Character string replacement part
22 Correction candidate search section
24 Correction candidate storage
26 Recognition result determination unit
28 Image generator
30 Dictionary Update Department

Claims (3)

First speech recognition means for performing speech recognition processing on input speech to obtain character string data as a recognition result composed of a plurality of characters;
Second speech recognition means for performing speech recognition processing on the input correction speech to obtain correction data consisting of one or more n characters;
A part of the character string data obtained by the first voice recognition means and different portions are replaced with the correction data obtained by the second voice recognition means to obtain a plurality of correction candidate data. and string replacement means you generate,
Correction candidate notification means for notifying the plurality of correction candidate data generated by the character string replacement means;
A correction location selection unit that selects one of the plurality of correction candidate data notified by the correction candidate notification unit according to a user operation;
A speech recognition correction method comprising: correcting the recognition result by the first speech recognition means using the correction candidate data selected by the correction location selection means . In claim 1,
A speech recognition dictionary used for speech recognition processing by the first speech recognition means;
Dictionary updating means for updating the contents of the speech recognition dictionary based on the correction candidate data selected by the correction location selecting means ;
A speech recognition correction method characterized by comprising: In claim 2,
  The correction candidate notification unit determines whether each of the plurality of correction candidate data generated by the character string replacement unit is registered in advance as a character string to be subjected to voice recognition in the voice recognition dictionary. A speech recognition correction method, wherein only the correction candidate data that has been registered is to be notified.

Images