JP4212947B2 - Speech recognition system and speech recognition correction / learning method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To surely and simply correct a recognition result when erroneous recognition of voice is arisen. <P>SOLUTION: When erroneous recognition of uttered voice is detected by a voice recognition engine 3, an interactive processing section 4 is constituted to read a word for which a user previously made correction for erroneous recognition, from a recognition word link DB7, the word is presented as a correct candidate and the erroneously recognized word and the correct word that is corrected by the user are made correspondent to each other and they are newly registered. Thus, when erroneous recognition is arisen, only a word which has high probability of being a correct one is presented to the user as an appropriate corrected candidate and recognition result is surely corrected by conducting only a simple operation in which a selection is made from correct candidates being presented. <P>COPYRIGHT: (C)2005,JPO&amp;NCIPI

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speech recognition system and a speech recognition correction / learning method, and in particular, a correspondence between a character string to be recognized and its speech pattern is registered as a speech dictionary, and a speech pattern having a high similarity to an input speech is recorded. This is suitable for use in a speech recognition system configured to recognize a character string having a character string of an input speech.
[0002]
[Prior art]
Most recent vehicles are equipped with various electronic devices such as an audio device, an air conditioner, and a navigation device. Recently, in order to avoid one-handed operation or the like when operating these electronic devices, a system is also provided that allows the electronic devices to be operated by voice recognition. By using this voice recognition technology, the driver can operate various electronic devices without removing his / her hands from the steering wheel (without manually operating an operation unit such as a remote controller or an operation panel).
[0003]
A speech recognition system usually recognizes a specific word, idiom or simple command sentence (which is simply expressed as “word” in this specification) spoken by a user as an utterance command, and synthesizes the recognized word by speech synthesis. Talk back. The user confirms the recognized word that has been talked back, and if it is correct, inputs that fact. In response to this, the system performs control according to the recognized word. On the other hand, when the recognition word talked back from the system is different from the word spoken by the user, the user performs voice input again.
[0004]
In such a speech recognition system, an acoustic model in which a character string of a recognition target word and its speech pattern are associated is registered in advance in the speech dictionary database. Then, the feature amount calculated from the user's input speech is compared with the feature amount of the acoustic model to search for the speech pattern with the highest similarity, and the character string having the speech pattern is recognized as the character string of the input speech To do.
[0005]
In such a speech recognition system, misrecognition of spoken speech is inevitable. Depending on the voice uttered by the user, misrecognition may occur continuously. As a countermeasure in this case, not only one word with the highest degree of similarity but also a function that presents a plurality of words with a higher degree of similarity to the user and selects one of them is also provided Has been. In addition, a technique has been proposed in which apparent recognition performance is improved by sequentially deriving recognition results after the highest recognition result that has been erroneously recognized as the highest recognition result (see, for example, Patent Document 1). ).
[0006]
[Patent Document 1]
Japanese Patent Application Laid-Open No. 10-63295
In addition, as a technique for increasing the recognition rate itself, various methods called “speaker adaptation” have been studied, in which the acoustic model is tuned for each speaker to reduce the occurrence of misrecognition. (For example, refer to Patent Document 2). As a typical speaker adaptation, there is a method called “enroll”. Enrollment is a process in which the user reads out words prepared in advance according to instructions from the system at the beginning of use of the system, and learns using the voice pattern of the specified word and the voice pattern related to speaker input. (For example, refer to Patent Documents 3 and 4).
[0008]
[Patent Document 2]
Japanese Patent Laid-Open No. 7-230295 [Patent Document 3]
JP 2002-132288 A [Patent Document 4]
JP 2000-148198 A [0009]
[Problems to be solved by the invention]
However, even if the system has the function of presenting multiple words with high similarity, the acoustic model of the recognition engine and the user's voice pattern are far apart if the system continuously misrecognizes In many cases, all the correct candidates presented according to the degree of similarity are misrecognitions. For this reason, even if a plurality of correct answer candidates are presented, the user cannot use them, and there is a problem that he / she has to repeat voice input many times or give up voice input and input a command using a remote control or the like. It was.
[0010]
Moreover, even if the apparent recognition performance is improved as in Patent Document 1, the actual recognition rate is not improved. In order to actually increase the recognition rate, it is necessary to perform speaker adaptation processing. However, in a normal use state where misrecognition may occur, it is not always “recognized word on the system side = correct word that the user wants to input”. Therefore, there is a problem that the recognition rate cannot be improved well even if speaker adaptation is performed only by relying on the result of speech input.
[0011]
It is also conceivable to use a command input by the user by operating a remote controller or the like as a correct value for speaker adaptation. However, on the system side, whether the command entered by the remote control is a command that was corrected and input after switching to remote control operation as a result of repeated voice recognition errors, or entered by a user's arbitrary operation regardless of voice recognition. I ca n’t figure out. Therefore, an input command from a remote controller or the like cannot be used as a correct value for speaker adaptation.
[0012]
From such a situation, in-vehicle electronic devices have generally used enrollers in which correct words are known in advance as speaker adaptation techniques. However, in order to increase the speech recognition rate using enrollment, the user must bother to read out some words prepared in advance on the system side at the beginning of using the system. For this reason, there is a problem that an extra burden is generated on the user because the operation is not directly related to the operation that the user actually wants to perform on the electronic device.
[0013]
The present invention has been made to solve such a problem, and it is an object of the present invention to make it possible to reliably and easily correct a recognition result when erroneous recognition of speech occurs.
Another object of the present invention is to make it possible to reliably and easily correct the result of speech recognition, and to actually improve speech recognition performance without performing enrollment that takes time and effort.
[0014]
[Means for Solving the Problems]
In order to solve the above-described problems, in the present invention, a misrecognized word and a correct word corrected by a user are registered as a record in the database, and then erroneously recognized for the same utterance Is generated, the correct answer corrected by the user previously is presented as the correct answer candidate for this time. According to the present invention configured as described above, when erroneous recognition occurs, it is possible to present to the user only those having a high probability of correct answers from past correction results.
[0015]
In another aspect of the present invention, the candidate selected by the user after presenting the correct answer candidate as described above is provided to the speaker adaptation means as a recognition result to be originally recognized. According to the present invention configured as described above, it becomes possible to accurately grasp the “correct answer that the user wants to input” on the system side even in a normal use state, and speaker adaptation is performed using the correct answer and the uttered voice. Can be performed appropriately.
[0016]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram illustrating a configuration example of the speech recognition system according to the present embodiment.
[0017]
As shown in FIG. 1, the speech recognition system of the present embodiment includes an operation unit such as a remote controller 1, a microphone 2, a speech recognition engine 3, a dialogue processing unit 4, a speech synthesis engine 5, a speaker 6, A recognition word link DB (database) 7, a screen display control unit 8, a display 9, and a speaker adaptation module 10 are provided. The remote controller 1 includes an utterance button 1a, a correction button 1b, an erroneous recognition button 1c, a joystick 1d, and an OK button 1e.
[0018]
The remote controller 1 is an operator for a user to perform various operations on an electronic device (such as an audio device or a navigation device) that uses the voice recognition system of the present embodiment. Performed by the remote control 1. The utterance button 1a is a button for instructing the start of speech recognition processing by utterance. That is, the voice input acceptance state by the utterance is set in accordance with the operation timing of the utterance button 1a. The joystick 1d is an operator that is used to correct a misrecognized word to a correct word when a voice misrecognition occurs.
[0019]
The correction button 1b is a button for instructing to redo voice input when, for example, a word different from what the user originally wanted to say is accidentally uttered. The misrecognition button 1c is used to instruct to correct a misrecognized word using the joystick 1d when misrecognition occurs, that is, when the recognized word talked back from the system is different from the word spoken by the user. It is a button.
[0020]
In this embodiment, in this way, two “return buttons”, which are conventionally only one button for instructing re-input of voice input or correction of a misrecognized word, are a correction button 1b and a misrecognition button 1c. It is divided into. As a result, the system side can clearly determine that correction of the recognized word has been instructed as a result of erroneous speech recognition.
[0021]
The OK button 1e instructs the determination of a desired menu item selected using the joystick 1d, or the input information (for example, destination) when the final content of the information input using the voice recognition process is correct. ) Is a button for instructing the system to set. The OK button 1e may be provided as a dedicated button alone as shown in FIG. 1, or may be configured to be used also as the speech button 1a or the joystick 1d.
[0022]
The voice recognition engine 3 compares the uttered voice input from the microphone 2 with a voice dictionary prepared in advance, and recognizes a word related to the uttered voice. Then, a command corresponding to the uttered voice is executed to an audio device or a navigation device (not shown) through the dialogue processing unit 4.
[0023]
The speech synthesis engine 5 synthesizes the words recognized by the speech recognition engine 3 and talks back from the speaker 6. In response to this, the user listens to the recognition word that has been talked back to confirm whether or not a misrecognition has occurred. If there is no erroneous recognition, the user performs voice input for the next process. On the other hand, if there is a misrecognition, the user presses the misrecognition button 1c to correct the recognized word. The screen display control unit 8 controls the display 9 to display a remote control operation screen when correcting the recognized word.
[0024]
The dialogue processing unit 4 executes a series of dialogue processing with the user when performing voice recognition. That is, processing for instructing the speech recognition engine 3 to start speech recognition processing in response to the user's operation of the utterance button 1a, and supplying the words recognized by the speech recognition engine 3 to the speech synthesis engine 5 to talk to the user For example, when the user presses the erroneous recognition button 1c as a result of the talkback, the screen display control unit 8 is controlled to provide a remote control operation screen to the user.
[0025]
The dialogue processing unit 4 detects the erroneously recognized word (by the voice recognition engine 3 when the erroneous recognition button 1c is pressed by the user when the erroneous recognition of the uttered voice by the voice recognition engine 3 is detected. (Recognition result) and the correct word corrected by the user using the joystick 1d after the operation of the erroneous recognition button 1c are associated with each other and registered in the recognized word link DB7. Thus, the dialogue processing unit 4 constitutes the correct word registration means of the present invention.
[0026]
When the dialog processing unit 4 controls the screen display control unit 8 at the time of correcting the recognized word and presents the above-described remote control operation screen on the display 9, the user previously corrected the word that was erroneously recognized at that time Is read from the recognized word link DB 7 and presented to the user as a list of correct answer candidates. Thus, the dialogue processing unit 4 and the screen display control unit 8 constitute the correct candidate presenting means of the present invention.
[0027]
Further, the dialogue processing unit 4 displays the speech voice input from the microphone 2 and the correct word corresponding thereto (recognition result when there is no misrecognition or correction result when there is misrecognition) as a speaker adaptation module. 10 is also performed.
[0028]
For example, when the misrecognition of the utterance voice by the voice recognition engine 3 is not detected, that is, when the utterance button 1a is pushed without the misrecognition button 1c being pushed, the dialogue processing unit 4 The result of recognition by the speech recognition engine 3 is provided to the speaker adaptation module 10. When the recognition word is corrected by pressing the misrecognition button 1c, the utterance voice at that time and the correction result (selection result from the correct answer candidate) are provided to the speaker adaptation module 10. . Thus, the dialogue processing unit 4 also constitutes information providing means of the present invention.
[0029]
The speaker adaptation module 10 performs speaker adaptation processing using the speech pattern from the microphone 2 and the correct speech pattern provided by the dialogue processing unit 4. The correct speech pattern is prepared in advance by the speaker adaptation module 10 as an acoustic model, and the speaker adaptation is performed using the corresponding speech pattern based on the correct word notified from the dialogue processing unit 4. Note that various methods can be applied to the contents of the speaker adaptation processing, but any known method can be applied to them, and a detailed description thereof will be omitted here.
[0030]
FIG. 2 is a conceptual diagram showing the data structure of the recognized word link DB 7. In FIG. 2, “link word” is a word that the user has previously corrected for the recognition result using the remote controller 1. That is, for example, a result that the user has previously corrected to “Saga Prefecture” or “Chiba Prefecture” using the remote controller 1 for the result of erroneous recognition of “Fukushima Prefecture” by the speech recognition engine 3 is recognized. It is registered in the word link DB7.
[0031]
Next, the operation of the speech recognition system according to the present embodiment configured as described above will be described. Before explaining the operation of the voice recognition system, the state transition of the utterance command which is the premise thereof will be explained. Usually, a plurality of utterance commands prepared in the system are managed in a plurality of layers according to the operation contents for the system. For example, when a destination is set by an address in the navigation device, as shown in FIG. 3, the address is divided into three layers and input, and finally the OK button 1e is pressed to set the input address as the destination. Set.
[0032]
That is, in the example of FIG. 3, words such as “address”, “telephone number”,... Are managed in the hierarchy in the initial state. If, for example, “address” is spoken in this hierarchy, the process proceeds to the next lower hierarchy 1. In this level 1, the prefecture name is managed, and words such as “Fukushima Prefecture”, “Saga Prefecture”, “Chiba Prefecture”, etc. can be input as utterance commands. When a desired prefecture name is uttered at level 1, the process proceeds to level 2, which is one level lower. In level 2, it is possible to input a city name as an utterance command.
[0033]
Similarly, when a desired city name is uttered at level 2, the system proceeds to level 3 which is one level lower. In this hierarchy 3, it is possible to input the remaining part of the address as an utterance command. When the remaining part of the address is spoken, the process proceeds to the final level 4. In level 4, by pressing the OK button 1e, the address input by the utterance is set as the destination. When the correction button 1b or the misrecognition button 1c is pressed in each of the levels 1 to 4 as described above, a return process is performed and the level returns to the level one above.
[0034]
4 and 5 are flowcharts showing an operation example of the speech recognition processing according to the present embodiment. 4 is a flowchart showing the operation of the hierarchical processing performed in each hierarchy shown in FIG. 3, and FIG. 5 is a flowchart showing the operation of the erroneous recognition correction processing included in FIG.
[0035]
In FIG. 4, the dialogue processing unit 4 determines whether or not the utterance button 1a has been pressed (step S1). If it is determined that the utterance button 1a has been pressed, the dialogue processing unit 4 activates the voice recognition engine 3 to set the voice input acceptance mode, and further determines whether or not the voice input engine 1 is in the initial state of FIG. 3 (step S2).
[0036]
If it is not the initial state, the dialogue processing unit 4 determines that the correct answer has been obtained by the speech recognition in the previous hierarchy, holds the following information as learning data (step S3), and stores it in the speaker adaptation module 10 Transmit (step S4).
i) Waveform data of speech (eg, user's voice waveform when uttering “Shiga Prefecture”) ii) Recognition result (eg, “Shiga Prefecture”)
iii) Information “recognition result = correct answer”
Thereafter, the user utters a desired word and inputs it from the microphone 2 (step S5). In response to this, the speech recognition engine 3 once exits the speech input acceptance mode and performs the recognition processing of the input word. Then, the speech synthesis engine 5 synthesizes the recognition result and talks back from the speaker 6 (step S6). After the talkback, the dialogue processing unit 4 executes a process for transitioning to the next layer (step S7).
[0038]
Note that the speaker adaptation module 10 uses the speaker adaptation algorithm based on parameter update, for example, based on the information i) to iii) provided from the dialogue processing unit 4 in step S4. Execute.
[0039]
If it is determined in step S1 that the speech button 1a has not been pressed, the dialogue processing unit 4 determines whether or not the correction button 1b has been pressed (step S8). When the correction button 1b is pressed, the dialogue processing unit 4 executes a return process for transitioning to the previous hierarchy (step S9).
[0040]
On the other hand, when determining that the correction button 1b is not pressed, the dialogue processing unit 4 further determines whether or not the erroneous recognition button 1c is pressed (step S10). When the misrecognition button 1c is pressed, the dialogue processing unit 4 determines that the result obtained by the speech recognition in the previous hierarchy is an error, and holds the following information as learning data (step S11). .
I) Waveform data of speech (eg, user's voice waveform when uttering “Shiga”) II) Recognition result (eg, “Fukushima”)
III) Information “Recognition Result = Error” Then, the dialogue processing unit 4 executes the return process for transitioning to the previous hierarchy (step S12), and then executes the erroneous recognition correction process shown in FIG. 5 (step S13).
[0041]
In FIG. 5, the dialogue processing unit 4 searches the recognized word link DB 7 using the erroneously recognized word from the speech recognition engine 3 (“Fukushima Prefecture” in the above example) as a key (step S21). As a result of this search, it is determined whether or not a link word that has been corrected by the user before the erroneously recognized word is registered in the recognized word link DB 7 (step S22).
[0042]
When one or more such link words are found, a remote control operation screen as shown in FIG. 6A including the link word as a correct candidate and further including the word “other” is displayed on the display 9. Present (step S23). If there is an actual correct answer among the correct answer candidates, the user operates the joystick 1d to select it. In this case, the dialogue processing unit 4 confirms that any word has been selected from the remote control operation screen shown in FIG. 6A (step S24), and determines whether or not the selected word is “other”. If any link word is selected from among the correct answer candidates other than “others”, the process jumps to step S29.
[0043]
On the other hand, when there is no actual correct answer in the correct answer candidates shown in the screen of FIG. 6A (when the user selects “other” by operating the joystick 1d), or in the recognized word link DB 7 in step S22. If it is determined that no link word is registered, all the selectable words in the scene are taken out and displayed as a list as shown in FIG. 6B (step S26). The user selects a correct word from the list by operating the joystick 1d (steps S27 and S28).
[0044]
In addition, the word which can be selected in the scene means the word of the applicable hierarchy. In the example of FIG. 6B, all the words in the hierarchy 1 in FIG. 3 managing the prefecture names such as “Fukushima Prefecture”, “Saga Prefecture”, and “Chiba Prefecture” are displayed as a list.
[0045]
When any word is selected on the remote control operation screen shown in FIG. 6A or 6B, the dialogue processing unit 4 registers the selected word in the recognized word link DB 7 (step S29).
[0046]
FIG. 7 is a diagram illustrating a registration example of the selected word with respect to the recognized word link DB 7. For example, when “Shiga Prefecture” is selected as the correct word from the remote control operation screen of FIG. 6B, the selected word is registered at the top of the link word (link word 1). When the new word “Shiga Prefecture” is registered in the link word 1, the words “Saga Prefecture” and “Chiba Prefecture” that have been registered so far move to the link word 2 and later.
[0047]
After the link word update processing, the dialogue processing unit 4 holds the following information as learning data (step S30), and when the information of I) to V) is prepared, the dialogue processing unit 4 stores them in the speaker adaptation module 10. Transmit (step S31).
IV) Selected words (eg “Shiga” selected with joystick 1d)
V) Information “selection result = correct answer” Then, the dialogue processing unit 4 executes a process of transitioning to the next layer (step S32), and ends the erroneous recognition correction process. The speaker adaptation module 10 executes speaker adaptation processing based on the information I) to V) received from the dialogue processing unit 4.
[0048]
As described above in detail, according to the present embodiment, when a misrecognition occurs, the correction result that the user has performed using the remote controller 1 in the past is presented as a correct answer candidate. A high appropriate correction candidate can be presented to the user. As a result, the user can surely correct the recognition result of the speech recognition engine 3 only by a simple operation of selecting any one of the correct answer candidates with a small number of cases.
[0049]
In addition, according to the present embodiment, the word that has been correctly obtained by recognition by the speech recognition engine 3 and the word that is selected on the remote control operation screen after the operation of the erroneous recognition button 1c are provided to the speaker adaptation module 10. Therefore, these words can be used as correct words that the user originally wanted to input. As a result, speaker adaptation learning can be performed in a normal use state of the system, and the user does not have to perform enrollment that takes time and effort. In addition, the acoustic model can be tuned to suit individual users in the background of the speech recognition process, so the speech recognition performance is also better than when using only “non-specific speaker speech recognition”. Get better.
[0050]
In the above embodiment, the remote controller 1 is used as the operation unit, but a touch panel may be used.
In the above-described embodiment, the example of displaying a list of words of the corresponding hierarchy when “others” is selected on the screen of FIG. 6A has been described. However, a software keyboard for individually inputting 50 sounds is provided. It may be displayed.
[0051]
In addition, each of the above-described embodiments is merely an example of the embodiment for carrying out the present invention, and the technical scope of the present invention should not be construed in a limited manner. In other words, the present invention can be implemented in various forms without departing from the spirit or main features thereof.
[0052]
【The invention's effect】
As described above, in the present invention, when erroneous recognition occurs, the results corrected by the user in the past are presented as correct answer candidates, so that only words with a high probability of correct answers are displayed as appropriate correction candidates. Can be presented. As a result, the user can reliably correct the recognition result only by a simple operation of selecting any one of the presented correct answer candidates when an erroneous voice recognition occurs.
[0053]
According to another feature of the present invention, a word that has been correctly obtained by speech recognition and a word that has been selected from correct candidates after the occurrence of misrecognition are provided to the speaker adaptation means as a recognition result that should be recognized. Since it is provided, the correct word can be accurately grasped by the speaker adapting means even in the normal use state, and the speaker adapting process can be appropriately performed. Thereby, even if the user does not bother to perform enrollment that takes time and effort, the speech recognition performance can be improved with certainty.
[Brief description of the drawings]
FIG. 1 is a block diagram illustrating a configuration example of a speech recognition system according to an embodiment.
FIG. 2 is a conceptual diagram showing a structure of a recognized word link DB according to the present embodiment.
FIG. 3 is a diagram showing a hierarchy transition state related to a plurality of utterance commands prepared in the voice recognition system of the present embodiment.
FIG. 4 is a flowchart showing the operation of hierarchical processing in the speech recognition processing according to the present embodiment.
FIG. 5 is a flowchart showing an operation of an erroneous recognition correction process in the voice recognition process according to the present embodiment.
FIG. 6 is a diagram showing a remote control operation screen presented when operating a misrecognition button according to the present embodiment.
FIG. 7 is a diagram illustrating an operation example of registering a selected word with respect to a recognized word link DB according to the present embodiment.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 Remote control 1a Speech button 1b Correction button 1c Error recognition button 1d Joystick 1e OK button 2 Microphone 3 Speech recognition engine 4 Dialogue processing part 5 Speech synthesis engine 6 Speaker 7 Recognition word link DB
8 Screen display controller 9 Display 10 Speaker adaptation module

Claims (4)

A speech recognition means for recognizing a word related to the uttered speech by comparing the input uttered speech with a speech dictionary prepared in advance;
When misrecognition of the uttered speech by the speech recognition means is detected, a word previously corrected by the user for the misrecognized word is read from the recognized word link database and presented as a correct candidate, and a correct word is selected. Correct answer candidate presentation means for prompting the user,
Correct word registration means for associating a word erroneously recognized by the speech recognition means with a correct word selected by the user through processing by the correct answer candidate presenting means and registering it in the recognized word link database;
Speech synthesis means for speech-synthesizing the words recognized by the speech recognition means;
A misrecognition button for operation when the user who has input the uttered speech confirms the recognized speech to be talked back by the speech synthesizer and determines that it is misrecognized;
A correction button to be operated when the user who has input the utterance voice instructs to input the utterance voice again,
According to the presence or absence of the operation of the misrecognition button, to detect the presence or absence of misrecognition of the speech voice by the speech recognition means,
A speech recognition system, wherein the user makes a transition to a state in which re-input processing of the uttered speech by the user is possible in response to an operation of the correction button . Speaker adaptation means for performing speaker adaptation processing using the input speech pattern and correct speech pattern;
When misrecognition of the uttered voice is not detected, when the input uttered voice and the recognition result by the voice recognition means are provided to the speaker adaptation means, and the erroneous recognition of the uttered voice is detected 2. The information providing means for providing the speaker adaptation means with the input speech voice and the correction result by the user made through the processing of the correct answer candidate presentation means. The speech recognition system described in 1. A first step of recognizing a word related to the uttered voice by comparing the input uttered voice with a prepared voice dictionary;
A second step in which the words recognized in the first step are speech synthesized and talkbacked;
A third step of detecting the presence / absence of an operation of a correction button to be operated when the user who has input the utterance voice instructs to re-input the utterance voice;
A fourth step of transitioning to a state in which the user can re-input the spoken voice when the operation of the correction button is detected in the third step;
The user who has input the uttered voice confirms the recognized voice to be talked back in the second step and determines that it is misrecognized. A fifth step of detecting the presence or absence of misrecognition;
When misrecognition of the uttered speech is detected in the fifth step, a word corrected by the user before the misrecognized word is read from the recognized word link database and presented as a correct candidate, and the correct candidate A sixth step for prompting the user to select a correct word from
A seventh step of registering the word detected as erroneous recognition in the fifth step and the correct word selected by the user through the process of the sixth step in association with each other in the recognized word link database;
An eighth step of providing the speaker adaptation unit with the uttered voice input in the first step and a correction result by the user made through the process of the sixth step;
The speaker adaptation unit performs speaker adaptation processing using the speech pattern input in the first step and the correct speech pattern based on the correction result provided in the eighth step. 9. A speech recognition correction / learning method comprising: 9 steps. If no erroneous recognition of the utterance voice is detected, the processing of the sixth step and the seventh step is not performed, and the utterance voice input in the first step in the eighth step. The speech recognition correction / learning method according to claim 3 , further comprising: providing the speaker adaptation unit with the recognition result obtained in the first step.

Images