JP2002287792A - Voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make it unnecessary to input an entire voice input object again at the time of correction in the case of the voice input object having a hierarchical structure and to prevent degradation of the recognition rate. SOLUTION: A voice information holding part 41 holds voice information inputted from a voice input part 27. When a voice is inputted for correction, a comparison and discrimination part 42 compares current inputted voice information with preceding voice information held in the voice information holding part 41 to discriminate resembling parts between them. A dictionary control part 34 of a voice recognition part 30 uses also a recognition result stored in a preceding result storage part 33 to perform dictionary control of taking only a part corresponding to a correction part as a comparison object pattern candidate. For example, when 'Aichi Prefecture, Kariya City, Showa town' is erroneously recognized as 'Aichi Prefecture, Kariya City, Shoei town' and 'Showa town' is inputted with voice for correction, only the town name part (concretely, Showa town, Shoei town, or the like) following Aichi Prefecture, Kariya City is taken as the comparison object pattern candidate.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a case where a destination setting or the like in a navigation system can be input by voice, and more particularly to a voice recognition technique suitable for coping with a correction input in the case of erroneous recognition. About.

[0002]

2. Description of the Related Art Conventionally, a speech recognition apparatus that compares an input speech with a plurality of pre-stored comparison target pattern candidates and obtains a speech having a high degree of coincidence as a recognition result has already been put into practical use. It is used, for example, for a user to input a destination name to be set in a navigation system by voice. In particular, when the driver uses the in-vehicle navigation system, voice input does not involve button operation or screen gaze, so that it is effective because the safety is high even when the vehicle is running.

In order to satisfy such a function, it is necessary to easily specify a sufficiently detailed point. Specifically, it is necessary to be able to input not only the level of the prefecture and the city but also the level of the name of the town under the city and the level of the capital in the town and village. Furthermore, if the user wants to set, for example, "Showa-cho, Kariya-shi, Aichi", it must be pronounced separately for each level, such as "Aichi-ken", "Kariya-shi", "Showa-cho". Then, since it is troublesome for the user, it is preferable that input (batch input) can be performed continuously.

[0004] However, in the case of such a configuration corresponding to the batch input, a situation in which the usability is degraded for the user may be assumed. That is when the user has to re-enter the destination to be set. That is,
This is because the current recognition technology cannot always say that the recognition result is completely accurate, so that a single voice input does not always result in accurate recognition. For example, when the user inputs a voice as "Showa-cho, Kariya-shi, Aichi", for example, "Shoei-cho, Kariya-shi, Aichi"
May be mistakenly recognized. Then, in the case of such an erroneous recognition, it is necessary to input a voice again as "Showa-cho, Kariya city, Aichi prefecture".

However, considering conversations in daily life, in the case of such an erroneous recognition, it is necessary to correct only the erroneously recognized town name, instead of inputting a voice as "Showa-cho, Kariya-shi, Aichi Prefecture". Is natural. That is, it is considered preferable that the second time be able to input only "Showa-cho" by voice.

To solve such a problem, there is a speech recognition apparatus disclosed in Japanese Patent Application Laid-Open No. H11-38994. According to this technology, if a word or a group of words constituting a higher layer of a comparison target pattern candidate obtained by connecting a plurality of words hierarchically is set as an abbreviation target in the recognition processing, the setting is performed. Even if the upper layer constituent words or word groups are omitted, they are temporarily regarded as comparison target pattern candidates, and then the input speech is compared to perform recognition processing. Therefore, for example, the user incorrectly recognized the voice recognition device as "Shoei-cho, Kariya-shi, Aichi," even though the user had input by voice as "Showa-cho, Kariya-shi, Aichi". In this case, when the user inputs the voice again, instead of inputting the voice of “Showa-cho, Kariya-shi, Aichi”, only the voice of “Showa-cho” needs to be input.

[0007]

As described above, the prior art is expected to improve the usability of the user, but on the other hand, there is a concern that the recognition rate may decrease. This is because, in the case of the conventional technology, in order to cope with rephrasing from all hierarchies, a method is adopted in which all intermediate hierarchies that may be rephrased are also regarded as recognition start points. It is because it increases.

[0008] Such a problem is not limited to the above-mentioned place names consisting of prefectures, municipalities, etc., but can be similarly applied as long as a plurality of words are similarly set as a hierarchical connection. . Therefore, the present invention is to eliminate the need to input all voices from the upper layer when re-inputting voices for correction due to erroneous recognition when the voice input to the voice recognition device has a hierarchical structure. Therefore, it is possible to reduce the burden on the user, improve the usability, and prevent the recognition rate from decreasing.

[0009]

Means for Solving the Problems and Effects of the Invention
According to the speech recognition device described in (1), when the user inputs a speech through the speech input unit, the recognition unit compares the inputted speech with a plurality of comparison target pattern candidates stored in the dictionary unit in advance. Then, a result having a high degree of coincidence is set as a recognition result, and the notification result is notified by the notifying means. And
If a predetermined confirmation instruction is issued after the notification of the recognition result, the post-confirmation processing means executes the predetermined post-confirmation processing assuming that the recognition result has been confirmed. Here, at least one of the plurality of comparison target pattern candidates stored in the dictionary means is set as a plurality of words connected hierarchically.

Under such a premise, the voice information holding means holds the voice information input at the time of the previous utterance, and if the voice input is made without a predetermined confirmation instruction after the notification of the recognition result. Comparing the voice information input at the time of this utterance with the held voice information at the time of the last utterance,
It is determined which part of the voice information at the time of this utterance is closest to the voice information at the time of the last utterance, and the closest part is determined as a corrected part. Further, when the portion determined to be a correction portion corresponds to a part of the hierarchy of the comparison target pattern candidate in which a plurality of words are hierarchically connected, the recognizing means determines that the constituent word or the word of the partial hierarchy is included. After temporarily considering the group as a comparison target pattern candidate, a comparison is performed on the voice input at the time of this utterance.

As described above, even if the comparison target pattern candidates other than the comparison target pattern candidates set in the dictionary means in advance, the constituent words or word groups of the partial hierarchy are temporarily compared with the comparison target pattern candidates having the hierarchical structure. Since it is regarded as a pattern candidate and used for the recognition process, only a part of the hierarchical portion can be restated for correction. For example, if the user uttered "Showa-cho, Kariya-shi, Aichi", but incorrectly recognized the voice recognition device as "Shoei-cho, Kariya-shi, Aichi," When the user inputs the voice again, instead of inputting the voice of "Showa-cho, Kariya-shi, Aichi", it is sufficient to input only the voice of "Showa-cho". If the device side misrecognizes it, it is very natural to correct only the misrecognized part (the town name "Showa-cho" in the above example) from the habit of conversation in everyday life. . Forcing special attention only when using a voice recognition device is not preferable in terms of usability. Therefore, as in the case of the speech recognition apparatus of the present invention, it is possible to cope with a very natural behavior in a daily conversation habit of correcting only the part of erroneous recognition. The user's burden of having to input a voice from the user can be reduced, and usability can be further improved.

Further, while being able to cope with such rephrasing (correction), erroneous recognition can be reduced as compared with the above-described conventional technique. Because, in the case of the prior art,
Since it is a concept that all intermediate layers that are possible in order to cope with restatement from all layers are regarded as recognition start points, the number of pattern candidates to be compared greatly increases. On the other hand, in the case of the present invention, a correction point is specified in advance, and only a part corresponding to the specified correction point is set as a comparison target pattern candidate. For example, in the case of the above example, there is no correction up to Kariya city in Aichi prefecture.
Only Showa-cho, Matsuei-cho, etc.) may be set as comparison target pattern candidates. That is, all the comparison target pattern candidates that do not have Kariya City in Aichi Prefecture as the upper layer are excluded from the target, so that only a relatively small number is required. Reducing the number of pattern candidates to be compared contributes to a reduction in the possibility of erroneous recognition and a reduction in recognition processing time.

From the viewpoint of reducing the burden on the user and improving the usability, it is not preferable to restate all of the cases of partial erroneous recognition, but from the viewpoint of erroneous recognition prevention,
Matching is easier when there are more recognition targets. Therefore, although there is a trade-off between the merits of both, if correction input is performed for all layers below the erroneously recognized layer, the rear end can be specified, and improvement in recognition accuracy can be expected. In this case, as described in claim 2, when comparing the voice information at the time of the present utterance with the voice information at the time of the previous utterance, it may be performed from the end of the voice information.

In this case, it is indispensable for the user to comply with such a correction method in order to improve its effectiveness. It is preferable to notify The notification timing may be, for example, a time when the apparatus is started.
Of course, the notification may be made periodically.

On the other hand, telephone numbers and the like are area code-city code-****. These may be regarded as three levels, but it is also possible to treat each digit as a level in numerical input. In such a case, for example, when a single digit number is erroneously recognized and corrected and input by one digit, new erroneous recognition is likely to occur. That is, the same number may be present in other digits. Therefore, it is preferable to correct and input the preceding and succeeding numbers (for example, three digits in addition to one digit before and after). In other words, it is considered that, for example, by inputting at least three digits, the corresponding portion can be easily specified.

Also in this case, the effectiveness cannot be improved unless the user observes such a correction method.
It is preferable to notify the user of such a correction method as shown in FIG. The voice information at the time of this utterance and the voice information at the time of the previous utterance to be compared by the correction portion determination means may be voice waveform information as described in claim 5 or as described in claim 6. May be a voice feature parameter. Also, when comparing these, claim 7
The method may be performed using a DP matching method as shown in FIG. 8 or may be performed using a hidden Markov model as described in claim 8.

By the way, when a portion determined to be a correction position corresponds to a part of the hierarchy of the pattern candidate to be compared, constituent words or word groups of the partial hierarchy are temporarily regarded as the pattern candidate to be compared. , Not only "normal input" but also "input for correction". In other words, when the correction operation detecting means detects a predetermined operation of the user indicating that the current utterance is a correction of the previous utterance content, the correction operation detecting means detects the "partial hierarchy temporarily regarded as a comparison target pattern candidate". Only the constituent words or word groups of "are used. In other words, if it is known that the correction is the previous correction, it is not necessary to correspond to the "normal input", so that the number of pattern candidates to be compared is reduced accordingly, and erroneous recognition can be further reduced.

If the user can select one of a plurality of voice input methods to be corrected for correction, the correction portion determination means corrects the correction according to the selected correction method. If the location is determined, more accurate recognition can be performed. For example, in the example described so far, there is a method of inputting only a portion to be corrected, a method of inputting a portion including a portion to be corrected, and a method of inputting before and after a portion to be corrected. According to the selection of, any of these corrections can be dealt with.

It has been described that, when a predetermined confirmation instruction is given after the notification of the recognition result, the recognition result is determined to be determined and the process proceeds to the predetermined post-confirmation processing. For example, when used in a car navigation system, a process itself for setting a destination as a recognition result or a process for instructing a device that executes the destination setting process to set the destination may be considered. . Regarding the “predetermined confirmation instruction” after the notification of the recognition result, it is conceivable that the instruction may be made by voice (for example, input by uttering “yes”) or by operating switches.

Further, regarding the notification of the recognition result,
According to an eleventh aspect of the present invention, it is conceivable to perform the recognition by outputting the contents of the recognition result by voice from a predetermined voice generating device. When used for in-vehicle equipment such as car navigation systems, if audio is output, the driver does not need to shift his or her viewpoint to the display device, which is advantageous in terms of further ensuring safe driving. I can say. However, the present invention is not limited to audio output, and the content of the recognition result may be displayed as an image using characters or symbols on a display device capable of displaying characters or symbols on a screen, for example. Or may be notified by both voice and image, or other notification methods may be adopted. Although it was stated that audio output is advantageous when applied as in-vehicle equipment,
Of course, there are situations in which the vehicle is not running, so if both the voice and the image are used, the driver can perform both the confirmation by display and the confirmation by voice.

In the case where the recognition result when a speech input for correction is made is displayed and displayed, for example, the first recognition result or a portion corrected from the previous recognition result may be used. Is preferably displayed in a different display mode so as to be distinguishable from other portions. For example, changing colors or enlarging characters. In this way, the correction part in the whole becomes clear.

The application of the above-described speech recognition apparatus is, for example, a navigation system.
In this case, a voice recognition device and a navigation device are provided, and the voice input means of the voice recognition device is provided with at least predetermined navigation process-related data which needs to be specified when the navigation device performs the navigation process. The instruction is used by the user to input by voice, and the post-determination processing means may be configured to output the result of recognition by the recognition means to the navigation device. In this case, the “predetermined navigation processing-related data” is a typical destination, but other than that, it is necessary to specify the navigation processing, such as selecting conditions for a route search. Instructions are included. In this case, the navigation processing-related information as a recognition result is notified. As described above, when a part of the place name once input by voice is erroneously recognized, the erroneous part is recognized. The user's usability is improved, for example, the user has only to input again.

[0023]

Embodiments of the present invention will be described below with reference to the drawings. It is needless to say that the embodiments of the present invention are not limited to the following examples, and can take various forms as long as they belong to the technical scope of the present invention.

FIG. 1 shows a control device 1 having a voice recognition function.
It is a block diagram showing the structure of the system containing. In addition,
The control device 1 of the present embodiment is mounted on an automobile (vehicle), and interacts with a vehicle occupant (mainly, a driver) as a user by voice, and controls the navigation device 15 mounted on the vehicle. To control.

As shown in FIG. 1, the control device 1 of this embodiment
Is operated by a user to input various commands, data, and the like by an external operation, a display device 5 for displaying an image, a microphone 7 for inputting voice, and a voice input. It is connected to a talk switch 9, a speaker 11 for outputting a voice, and a well-known navigation device 15 for detecting a current position (current position) of the vehicle and providing route guidance.

The navigation device 15 is a well-known GPS device for detecting the current position of the vehicle, a CD-ROM storing route guidance data such as map data, place name data, facility name data, and the like. A CD-ROM drive for reading data from the ROM, operation keys for the user to input commands, and the like are provided.
For example, when a user inputs a destination and a command for instructing route guidance to the destination via an operation key from the user, the navigation device 15 is optimal for reaching the current position of the vehicle and the destination. The display device 5 displays a road map including a simple route and provides route guidance. The display device 5 displays not only a road map for route guidance by the navigation device 15 but also various images such as an information search menu.

The control device 1 includes a CPU, a ROM,
A control unit 50 mainly composed of a microcomputer including a RAM and the like, an input unit 23 for inputting commands and data from the switch device 3 to the control unit 50, and an image data output from the control unit 50 Screen output unit 25 that converts the image signal into an image signal and outputs the image signal to the display device 5 to display an image on a screen, and an audio input unit 27 that converts an audio signal input from the microphone 7 into digital data.
From the audio signal input via the audio input unit 27,
A speech recognition unit 30 for recognizing and acquiring a keyword as a word spoken by the user (hereinafter, also referred to as an utterance keyword) and a speech signal input through the speech input unit 27 similarly to the user's previous speech A correction point determination unit 40 that determines a correction point from the utterance content and the current utterance content, and a voice that converts the text data output from the control unit 50 into an analog voice signal, outputs the analog voice signal to the speaker 11, and sounds the speaker 11 An output unit 28 and an equipment control interface (equipment control I / F) 29 for connecting the navigation device 15 and the control unit 50 so as to enable data communication are provided.

The voice input unit 27 extracts frame signals of, for example, a period of about several tens of milliseconds at regular intervals in order to analyze the characteristic amount of the input voice, and outputs the input signal in a voice section containing voice. It is determined whether there is a noise section or a noise section containing no voice. Since the signal input from the microphone 7 includes not only the speech to be recognized but also noise, the speech section and the noise section are determined.
As this determination method, many methods have been proposed as compared with the related art.For example, the short-time power of the input signal is extracted at regular time intervals, and it is determined whether the short-time power of a predetermined threshold or more has continued for a certain time or more. A technique of determining whether a section is a speech section or a noise section is often adopted. When it is determined that the input signal is in the voice section, the input signal is output to the voice recognition unit 30.

Here, the configurations of the speech recognition section 30, the correction location determination section 40 and the control section 50 will be described in more detail with reference to FIG. The voice recognition unit 30 includes a collation unit 31,
A dictionary unit 32, a previous result storage unit 33, and a dictionary control unit 34 are provided. The dictionary unit 32 stores the IDs of a plurality of keywords (comparison pattern candidates) that are assumed to be spoken by the user and are to be recognized by the control device 1, and dictionary data including the structures thereof. Then, the collating unit 31
Then, collation (recognition) is performed using the voice data input from the voice input unit 27 and the dictionary data of the dictionary unit 32, and the ID of the keyword having the highest likelihood of recognition is output to the control unit 50 as a recognition result. This recognition result is stored in the previous result storage unit 33.
Is also stored. The previous result storage unit 33 updates and stores the recognition result obtained by the matching unit 31. Therefore, only the previous recognition result and the dictionary control unit 3
Reference numeral 4 controls the dictionary of the dictionary unit 32 based on the previous recognition result stored in the result storage unit 33 and the correction location determined by the correction location determination unit 40. The contents of the dictionary control will be described later.

On the other hand, the correction location judging section 40 includes a voice information holding section 41 and a comparing / judging section 42. The voice information holding unit 41 holds voice information input from the voice input unit 27, and the comparison / judgment unit 42 stores the voice information
Then, the voice information input this time is compared with the previous voice information held in the voice information holding unit 41 to determine a similar part between the two. In this comparison, the sound waveform itself may be compared, or a feature parameter of the sound may be used. The comparison is performed using the DP matching method or the hidden Markov model. Then, the similar part is determined to be a correction part, and the speech recognition unit 30
Is output to the dictionary control unit 34. Note that the information about the correction point is also output to the control unit 50.

The control section 50 includes a recognition result storage section 51, a final recognition result determination section 52, a post-processing section 53, and the like. The recognition result storage unit 51 stores the recognition result output from the speech recognition unit 30. The recognition result stored here is stored until it is deleted. Then, the final result determination unit 52 determines the final recognition result based on the one or more recognition results stored in the recognition result storage unit 51 and the correction part input from the correction part determination unit 40. One or more recognition results stored in the recognition result storage unit 51 are cleared (deleted) when a predetermined confirmation instruction is given.

In the post-processing section 32c, for example, when the above-mentioned predetermined confirmation instruction is issued, the device control I / F 29
To execute a "post-confirmation process" for instructing the navigation device 15 to perform a predetermined process by sending data to the navigation device 15 via the PC or to output the recognition result output from the voice recognition unit 30 to the voice output unit 28 as text data. Then, a process for instructing the speaker 11 to sound is executed.

The recognition result sent from the voice recognition unit 30 to the control unit 50 may be all of the upper comparison target patterns as the final recognition result, or only the uppermost pattern among them. However, in the following description, in order to facilitate understanding, it is assumed that only the highest order is sent unless otherwise specified.

Further, in this embodiment, when the user presses the talk switch 9, voice input becomes possible thereafter. Note that it is assumed that no voice is input even though the talk switch 9 is pressed. Therefore, if there is a silent section for a predetermined time or more after the talk switch 9 is pressed and voice input is enabled, voice input is disabled. Move to a possible state. Note that the voice input unit 27 monitors the timing at which the talk switch 9 is pressed, and it is sufficient to detect that the talk switch 9 is pressed, but the voice recognition unit 30 and the correction location determination unit 40
Monitors the timing at which the talk switch 9 is pressed and the time during which the pressed state is continued, so that it can be determined whether the talk switch 9 is a click operation or a double-click operation. Specifically, when the talk switch 9 is turned off within a relatively short time (for example, within 0.5 seconds) after the talk switch 9 is turned on, it is regarded as a click operation. If the click operation is performed twice consecutively within a predetermined interval (for example, within 0.5 seconds), it is regarded as a double click operation. In the system of this embodiment, a click operation is performed in the case of normal voice input, and a double click operation is performed in the case of voice input for correction. Is detected by the voice recognition unit 30 and the correction part determination unit 40, and the corresponding processing is executed. It should be noted that when a so-called long-press operation (for example, an operation of holding down for two seconds or more) is performed instead of the double-click operation, the correction input may be performed.

Next, the operation of the system of this embodiment will be described.
An example in which a destination for a route search is input by voice using the navigation device 15 will be described with reference to the flowcharts of FIGS. First, in S10, which is the first step in FIG. 3, it is determined whether or not the talk switch 9 is turned on (pressed). If the talk switch 9 is turned on (S10: YES), the voice extraction processing is performed. Is performed (S20). This voice extraction processing is performed by the voice input unit 27.
In the process, it is determined whether or not the voice section is a voice section or a noise section based on the voice data input via the microphone 7, and the data of the voice section is extracted and output to the voice recognition unit 30 and the correction part determination unit 40. It is.

Next, a voice recognition process is performed (S30). Details of the voice recognition processing will be described with reference to FIG. As described above, the voice recognition unit 30 and the correction location determination unit 40
Monitors the timing at which the talk switch 9 is pressed and the time during which the pressed state is continued, so it is first determined whether or not the talk switch 9 is clicked (S31). And if it is a click operation (S31:
YES), since it is a normal voice input, the recognition processing is executed as it is (S32). On the other hand, if it is not a click operation (S31: NO), for example, when double-clicking or long-pressing is performed and the input is for correction, the correction point determination unit 40 determines a correction point (S3).
4) The speech recognition unit 30 controls the dictionary based on the determined correction part (S35). In this case, the recognition process in S32 is performed using the controlled dictionary.

Here, the dictionary control in S35 will be described. First, the dictionary data stored in the dictionary unit 32 according to the present embodiment will be described. The dictionary data includes not only the data of the vocabulary itself as the comparison target pattern but also the data indicating the hierarchical structure when the vocabulary of the comparison target pattern is a plurality of words connected hierarchically. It is remembered. More specifically, syllable data constituting a vocabulary is allocated to each side of a tree structure (indicated by an arrow (→) in FIG. 5) as shown in FIG. In FIG. 5, a single circle (○) represents a vertex,
Double circles (◎) represent accepted vertices, that is, vertices for words. Then, the vertex indicated by the arrow A in FIG. 5 becomes the “root”, from which the word is completed by tracing syllables assigned to each side according to preorder traversal. Here, "preceding forward scan" means that a root is visited, and then a subtree having a child as a root is sequentially scanned (this scan is also a forward forward scan). Here, "parent"
Means the vertex immediately before, "child" means the next vertex, and "sibling" means vertices having the same parent.

That is, in the specific example shown in FIG. 5, when the vertices (indicated by the arrow A) which are the roots are sequentially traced, the vertices become "Aichiken" and become the accepted vertices indicated by the arrow B. Therefore, "Aichiken (Aichi Prefecture)" is one recognition target word. Then, through the reception vertex indicated by the arrow B, “Kariyashi” is formed and the reception vertex indicated by the arrow C is obtained. Therefore, "Aichi Ken Kariyashi (Kariya City, Aichi Prefecture)" is also one recognition target word. Further, there is a vertex of “child” via the receiving vertex indicated by the arrow C. FIG.
Although it is not shown, since it can be traced as "showacho" and has a reception vertex, for example, "aichikenkariya shishowacho (Showa-cho, Kariya city, Aichi prefecture)" is one recognition target word.

In this case, for example, one recognition target word such as "Aichiken Kariyashi Showacho (Kawatani City, Aichi Prefecture)" includes "Aichiken (Aichi Prefecture)" and "Kariyashi (Kariya City )) And Showacho (Showa-cho) are hierarchically connected. Therefore, it can be seen from the existence of the reception vertices indicated by arrows B and C in FIG. In other words, the accepted vertex indicates that a word is composed of the syllable data that has been traced to it. Conversely, if there is more syllable data downstream from the accepted vertex, the upstream side of the accepted vertex is in the upper hierarchy. And the downstream side is the lower hierarchy. For example, consider the receiving vertex indicated by arrow B in FIG.
"Aichiken (Aichi Prefecture)" on the upstream side is the upper hierarchy,
"Kalyashi (Kariya City) ..." on the downstream side is the lower hierarchy. That is, in this case, the word indicating the prefecture is the upper hierarchy, and the word indicating the city level or lower is the lower hierarchy. Also, considering the reception vertex indicated by arrow C in FIG. 5, “Aichiken Kariyashi (Kariya City, Aichi Prefecture)” on the upstream side is the upper layer, and although not shown in FIG. "Wachi (Showa-cho)" is the lower hierarchy.

In the above, the description of the dictionary data stored in the dictionary section 32 has been given with a specific example of Showa-cho, Kariya city, Aichi prefecture. However, basically, the prefecture is the highest level and the city level is the second level. And the town level is the third level, and data is also set for other place names. In addition,
"Basically" means "town" at the next city level in the prefecture
And "villages" come in some places.

For such dictionary data, S in FIG.
At 34, the following dictionary control is performed. That is, since the correction location determination unit 40 knows where the correction location is in the previously input speech information, the dictionary control unit 34 of the voice recognition unit 30 also recognizes the recognition result stored in the previous result storage unit 33. Only the portion corresponding to the specified corrected portion is used as a comparison target pattern candidate. For example, despite the fact that the user has input by voice as "Showa town in Kariya city, Aichi prefecture", the voice recognition unit 30 mistakenly recognizes "Shoei town in Kariya city in Aichi prefecture" and Suppose. When the user voice-inputs only "Showa-cho" for correction, the correction location determination unit 40 determines that the previous voice information (the user has entered "Aichikenkarishashiwacho") held in the voice information holding unit 41 ) Is compared with the current voice information (the user uttered "showacho") to determine which part of the previous voice information corresponds to which input. . The dictionary control unit 34 of the voice recognition device 30 that has obtained the information about the correction location, “Matsuei-cho, Kariya city, Aichi prefecture” stored in the previous result storage unit 33.
Based on the recognition result, it is understood that there is no correction to the previous speech recognition result up to Kariya city in Aichi prefecture, and that it is a correction of the town name part that is the lower hierarchy, so it follows from Kariya city in Aichi prefecture Only the street name portion (specifically, Showa Town, Matsusakae Town, etc.) is set as the comparison target pattern candidate. This is the content of dictionary control.

In other words, conversely, a positive determination is made in S31,
That is, in the case of a normal input, the recognition process (S32) is executed without performing such dictionary control. After the recognition process, the speech recognition unit 30 stores the recognition result in the previous result storage unit 33, outputs the result to the control unit 50 (S33), and proceeds to S40 in FIG.

Returning to the description of the flowchart of FIG.
At 0, the recognition result is talked back and displayed. In this talkback, the control unit 50 controls the audio output unit 28,
The recognition result is output from the speaker 11 by voice, and the screen output unit 25 is controlled to display characters or the like indicating the recognition result on the display device 5. Note that the talkback in this case is performed only for the part immediately recognized by the voice recognition unit 30. That is, in the above example, for the first voice input, talkback is made for "Matsuei-cho, Kariya-shi, Aichi" and for corrected input, talkback is made for "Showa-cho". On the other hand, in the case of display,
In the same way, only the correction part may be displayed for the correction input, but for example, a method of displaying "It is Showa-cho in Kariya city, Aichi prefecture" and changing the color of only the part of "Showa-cho" and displaying it Can be adopted.

Thereafter, it is determined whether or not the recognition is correct based on an instruction from the user (S50). Specifically, it may be based on an operation on the switch device 3 by the user, or may be based on a voice input from the microphone 7. For example, if there is a voice input indicating a positive content of "Yes", it can be determined that the recognition was correct,
If there is a voice input indicating a negative content such as "No" or "No", it can be determined that the recognition is incorrect.

If the recognition is erroneous (S
50: NO), it is determined whether or not the recognition result belongs to a predetermined category (S90). In the present embodiment, the processing for setting a destination for route guidance is premised, and the predetermined category is a category relating to a place name. If it is a predetermined category (S90: YES), S1
The process proceeds to 00, and the recognition result is temporarily stored in the recognition result storage unit 51 of the control unit 50. The recognition result temporarily stored in the recognition result storage unit 51 in this manner is stored unless the deletion processing in S80 is executed. In other words, when the correction is input many times, S100
May be executed a plurality of times, in which case, it means that all of them are stored.

Next, the correction method is notified (S11).
0). This is performed to acknowledge the correction method in order to cope with a situation in which the user is erroneously recognized and the user is expected to input again for correction. As a correction method, for example, a method of “correcting and inputting data in all the layers below the erroneously recognized layer” can be considered. In other words, if you uttered "Showa Town, Kariya City, Aichi Prefecture" and misrecognized "Showa Town, Okazaki City, Aichi Prefecture", instead of just uttering "Kariya City" as a correction for only the misrecognized part, This is a method of inputting voice as "Showa-cho in Kariya city". This is based on the finding that, from the viewpoint of preventing erroneous recognition, matching is easier when there are more recognition targets.
In addition, if all of the layers below the erroneously recognized layer are corrected and input, the rear end can be specified, so that improvement in recognition accuracy can be expected. When based on such a correction input rule, the correction part determination unit 40 can uniformly perform the voice information at the time of the present utterance and the voice information at the time of the previous utterance from the end of the voice information uniformly. Therefore, it is possible to determine a correction position (a position input again) with higher accuracy. In this case, since it is indispensable for the user to observe such a correction method in order to improve the effectiveness, in S110 of FIG. 3, such a correction method is notified to the user.

If the category is not the predetermined category (S90:
NO), and proceeds to S120 to execute other processing.
After the processing of S110 or S120, the process returns to S10, and the processing is repeated. On the other hand, if the determination is affirmative in S50, that is, it is determined that the recognition is correct, the final recognition result determination unit 52 of the control unit 50 determines the recognition result (S6).
0). If there is no correction input and only one recognition result is stored in the recognition result storage unit 51, it is determined as the final recognition result. If a plurality of recognition results are stored, the results are comprehensively determined to determine the final recognition result. Speaking of the above-mentioned specific example, voice input "Showa-cho, Kariya-shi, Aichi" is erroneously recognized as "Matsuei-cho, Kariya-shi, Aichi". If "Showa-cho" is correctly recognized, the recognition result storage unit 51 stores two recognition results, "Matsuei-cho, Kariya-shi, Aichi" and "Showa-cho". The information for specifying the correction location from the correction location determination unit 40 is transmitted to the control unit 50.
It is also output to ”Kawatani, Aichi Prefecture” in the first memory “Kariya City, Aichi Prefecture” and “Showa Town” in the second time are combined, "
Is determined as the final recognition result.

Even if there are two or more correction inputs and three or more recognition results are stored in the recognition result storage unit 51, the final recognition result can be determined in the same way. In other words, the correction is not limited to one time, but may be performed any number of times until it is finally correctly recognized. Therefore, if the recognition result of the talkback or the like in S40 is not correct, the user may input the correction many times.

When the recognition result is determined, a predetermined post-determination process is executed (S90). In this case, the post-determination processing means that data relating to “destination for route guidance” as a recognition result is output to the navigation device 15 via the device control I / F 29, and dictionary control in the voice recognition unit 30 is performed. This is an undoing process. The dictionary control described above
This is only a corrective measure, and if correct recognition is finally obtained, it is better not to restrict the dictionary.

Then, the recognition result temporarily stored in the previous result storage unit 33 of the voice recognition unit 30 and the recognition result storage unit 51 of the control unit 50 is deleted (cleared) (S8).
0). Such a temporary storage of the recognition result is also a countermeasure for correction, and is unnecessary when the correct recognition is finally obtained. On the other hand, if such a recognition result remains, it is inconvenient when recognizing a voice input of another content. After the process in S80, the process returns to S10, and the process is repeated.

The above is an explanation of the operation in the case where the destination for route guidance is input by voice as an example. In order to more clearly understand the features and effects related to voice recognition of the present invention, As mentioned in the processing explanation of the above-mentioned flowchart, the destination is "Showa-cho, Kariya-shi, Aichi"
The description will be continued with a specific example of specifying.

It is assumed that the user voice-inputs “Showa-cho, Kariya-shi, Aichi” through the microphone 7. Accuracy of voice recognition is 10
If it is not 0%, it may be erroneously recognized. For example, if the user mistakenly recognizes “Shoei town in Kariya city, Aichi prefecture”, the sound is output via the speaker 11.

As a result, the user is found to be erroneously recognized, and it is necessary to re-input the voice for correction. In this case, the user again hears "Showa-cho, Kariya-shi, Aichi". You do not need to input it, just input the voice of "Showacho". It is very natural to correct only the misrecognized portion (the town name “Showa-cho” in the above example) in view of customs such as conversation in daily life. Forcing special attention only when using a voice recognition device is not preferable in terms of usability. Therefore, as in the present embodiment, it is possible to cope with a very natural behavior in a daily conversation habit of correcting only the misrecognition part, so that even if it is more natural to omit the upper layer, voice input from the upper layer The user's burden of having to do so can be reduced, and usability can be further improved.

Further, while being able to cope with such rephrasing (correction), erroneous recognition can be reduced. In the case of the above-described conventional technique, all possible intermediate layers are considered to be recognition start points in order to cope with rephrasing from all layers. On the other hand, in the case of the present embodiment, the correction location determination unit 40 specifies a correction location in advance, and only a portion corresponding to the specified correction location is set as a comparison target pattern candidate. Only needs to be done. In the above example, since there is no correction up to Kariya city in Aichi prefecture, only the lower-level town name portion (specifically, Showa town, Matsusakae town,..., Etc.) may be set as comparison target pattern candidates. That is, all the comparison target pattern candidates that do not have Kariya City in Aichi Prefecture as the upper layer are excluded from the target, so that only a relatively small number is required. Reducing the number of pattern candidates to be compared contributes to a reduction in the possibility of erroneous recognition and a reduction in recognition processing time.

In this embodiment, the microphone 7 and the voice input unit 27 correspond to "voice input means", and the voice output unit 28, the speaker 11, the screen output unit 25, and the display device 5 correspond to "notification means." And "correction method informing means". The voice recognition unit 30 corresponds to “recognition means”, and the dictionary unit 32 in the voice recognition unit 30 corresponds to “dictionary means”.
Further, the control unit 50 corresponds to “post-confirmation processing means”. In addition, the audio information holding unit 41 in the correction location determination unit 40 corresponds to “audio information storage unit”, and the comparison / determination unit 42 corresponds to “correction location determination unit”. Also, the talk switch 9,
Voice input unit 27, voice recognition unit 30, correction location determination unit 40
Correspond to “correction operation detecting means” and “correction method detecting means”.

[Another Embodiment] Referring to FIGS. 1 to 5,
Having described one embodiment, several other embodiments will now be described. (1) In the above embodiment, with emphasis on the point of prevention of erroneous recognition, in the case of a correction input, a rule for correcting and inputting all the layers below the erroneously recognized hierarchy is used. However, from the viewpoint of reducing the burden on the user and improving the usability, it is preferable to correct and input only the erroneously recognized portion. Which of these methods is adopted is a trade-off between the merits of the two methods. For example, the user may be able to select one of the methods. In that case, it is conceivable to execute the process shown in FIG. 6 instead of the process shown in FIG. 4 as the voice recognition process in S30 of FIG.

Here, two types of correction methods are recognized, and these are distinguished by the operation method of the talk switch 9. That is, when the talk switch 9 is clicked (S231: YES), the normal recognition processing is performed, but when the talk switch 9 is not clicked (S231: YES).
NO), and the method of correction input is distinguished depending on whether or not the talk switch 9 is double-clicked. In the case of a double click (S234: YES), a normal correction point determination is performed (S235), and in the case of no double click (for example, triple click or long press) (S234:
NO), and determine the correction location by comparing "from the rear end of the input audio information" as described in the above embodiment (S2).
37).

Although the recognition accuracy is relatively improved by comparing from the rear end as in S237, it is necessary for the user to observe the correction method. On the other hand, in the case of S236, correction input of only the correction portion may be performed, so that the burden on the user is reduced and the usability is relatively improved. Of course, S
In the case of 236, since a free correction method can be performed, a portion other than the corrected portion can be input again.

The other steps are the same as those in FIG. That is, S232 and S233 of FIG. 6 are processes having the same contents as S32 and S33 of FIG.
6 is a process having the same contents as S35 of FIG. In this case, the notification of the correction method in S110 of FIG. 3 indicates that two correction methods can be selected and the designation method.

(2) In the above embodiment, an address is taken as an example. However, other than that, when a plurality of words can be handled as being connected in a hierarchical manner, the same can be applied. One of them can be a telephone number. Telephone numbers are generally area code-city code-****, so they may be considered as three layers, but for numeric input, each digit may be treated as a layer. . However. If a single digit is mis-recognized and only one digit is corrected and input, there is a possibility that the same digit may be present in another digit, and new misrecognition is likely to occur. Therefore, it is preferable to correct and input the preceding and succeeding numbers (for example, three digits in addition to one digit before and after). That is, for example, if the input is made with at least three digits, it is considered that the corresponding portion can be easily specified. In such a case, it is also preferable to notify the correction method.

(3) In the above embodiment, in the process of S90 in FIG. 3, it is determined whether or not the recognition result belongs to a predetermined category, and the predetermined category is based on the setting of a destination. It is explained that the category is related to place names. However, the gist of the present invention is not limited to such place names and the like, but in abstract terms, information that needs to be formally output after outputting a recognition result and obtaining confirmation from a user. Category. More specifically, in the case of the above-described car navigation system, this means an instruction of predetermined navigation processing related information that needs to be specified in performing the navigation processing. A representative one of the "predetermined navigation processing related information" is a destination, but also includes an instruction which needs to be specified in the navigation processing, such as selection of a condition for a route search.

(4) In the above embodiment, the example in which the voice recognition device is applied to the car navigation system has been described. However, the application destination is not limited to the above-described car navigation system 2. For example, when the voice recognition device is used for an air conditioning system, it is conceivable to adjust the set temperature, select an air conditioning mode (cooling / heating / dry), or select a wind direction mode by voice input. For example, as for the set temperature, there are instructions related to the set temperature, such as "set the set temperature to 25 degrees" or "decrease the set temperature by 5 degrees", but there are a plurality of instructions. By applying the present invention,
If the user inputs "set the temperature to 25 degrees" but misrecognizes "set the temperature to 22 degrees", the user may input only "set the temperature to 25 degrees" in the voice input again. The usability of the user is improved. The same applies to the air conditioning mode and the wind direction mode.

Further, the car navigation system and the air conditioning system are not limited to the case where they are used as on-vehicle equipment, but may be, for example, a portable navigation device or an indoor air conditioner. However, as described above, when used for in-vehicle equipment, the user may be a driver. In that case, driving itself is the most important. For other in-vehicle equipment, drive as much as possible. Preferably, there is no hindrance. Therefore, in the case of a speech recognition device on the premise of a car navigation system or an air conditioning system as an in-vehicle device, there is a further advantage. Of course, from such a viewpoint, the present invention can be similarly applied to in-vehicle devices other than the navigation system and the air conditioning system. For example, a car audio device is effective. Also, considering a configuration in which opening and closing of the power window and adjustment of the mirror angle are instructed by voice, it is effective even in such a situation.

(5) The speech recognition function (including the correction part determination function and the like) described in the above embodiment can be realized by executing a program provided in the control device 1.
Such a program can be recorded on a recording medium and distributed, or can be provided via a network, and can be loaded into a computer from a recording medium or a network and executed.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a schematic configuration of an embodiment system.

FIG. 2 is a block diagram illustrating a configuration of a speech recognition unit, a correction location determination unit, and a control unit of the system according to the embodiment.

FIG. 3 is a flowchart illustrating processing related to voice recognition and dialog control in the system of the embodiment.

FIG. 4 is a flowchart showing a speech recognition process executed in FIG. 4;

FIG. 5 is an explanatory diagram showing dictionary data stored in a dictionary unit in the speech recognition unit.

FIG. 6 is a flowchart showing another embodiment of the voice recognition process.

[Description of Signs] 1 ... Control device, 3 ... Switch device, 5 ... Display device, 7 ...
Microphone, 9 talk switch, 11 speaker, 15 navigation device, 23 input unit, 25
Screen output unit, 27: voice input unit, 28: voice output unit, 2
9: device control I / F, 30: voice recognition unit, 31: collation unit, 32: dictionary unit, 33: previous result storage unit, 34: dictionary control unit, 40: correction location determination unit, 41: voice information holding unit .., 42... Comparison / determination unit, 50... Control unit, 51... Recognition result storage unit, 52.

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G10L 15/00 G10L 3/00 551Q 15/28 561C 561D

Claims (13)

[Claims] 1. A voice input means for inputting voice, and a voice input through the voice input means is compared with a plurality of comparison target pattern candidates stored in a dictionary means in advance to determine a matching degree. A recognizing means for recognizing a higher result, a notifying means for notifying the recognition result by the recognizing means, and a predetermined confirmation instruction after the recognizing result is notified by the notifying means, the recognizing result is displayed. Post-determination processing means for executing predetermined post-determination processing as being determined, at least one of the plurality of comparison target pattern candidates stored in the dictionary means, a plurality of words hierarchically A voice recognition device set as a spliced voice recognition device, comprising: voice information holding means for holding voice information input at the time of previous speech via the voice input means; If no predetermined confirmation instruction is given after the notification of the recognition result by the voice input device and there is a voice input through the voice input device, the voice information and the voice information input at the time of the present utterance through the voice input device. The voice information of the previous utterance stored in the holding unit is compared with the voice information of the previous utterance to determine which part of the voice information of the previous utterance is closest to the voice information of the previous utterance, and the closest part is corrected. Correction part determining means, wherein the part determined as the correction part by the correction part determining means corresponds to a partial hierarchy of a comparison target pattern candidate in which the plurality of words are hierarchically connected. The recognition means is configured to temporarily consider constituent words or word groups of the partial hierarchy as the comparison target pattern candidates, and then perform a comparison with the input voice. Recognition device. 2. The speech recognition apparatus according to claim 1, wherein the correction portion determining means compares the speech information at the time of the current utterance with the speech information at the time of the previous utterance from the end of the speech information. A speech recognition device characterized by performing. 3. The speech recognition device according to claim 2, further comprising: a correction method notifying unit for notifying a voice input method to be performed by a user for correction, wherein said correction method notifying unit includes a correction method notifying unit. A voice recognition device that informs the user of input of a part including a desired part. 4. The speech recognition device according to claim 1, further comprising a correction method notifying unit for notifying a user of a voice input method to be corrected, wherein said correction method notifying unit includes a correction method notifying unit. A voice recognition device that informs the user to input a part including before and after a desired part. 5. The speech recognition apparatus according to claim 1, wherein the speech information at the time of the present utterance and the speech information at the time of the previous utterance, which are compared by the correction portion determination means, are speech waveform information. A speech recognition device, characterized in that: 6. The speech recognition device according to claim 1, wherein the speech information at the time of the present utterance and the speech information at the time of the previous utterance, which are compared by the correction portion determination means, are characteristic parameters of the speech. A speech recognition device, characterized in that: 7. The speech recognition device according to claim 1, wherein said correction portion determination means compares said speech information at the time of said current speech with speech information at the time of said previous speech by a DP matching method. A speech recognition device characterized by performing using a voice recognition device. 8. The speech recognition apparatus according to claim 1, wherein said correction portion determining means compares said speech information at the time of said present utterance with speech information at the time of said previous utterance by using a hidden Markov model. A speech recognition device characterized by performing using a voice recognition device. 9. The voice recognition device according to claim 1, further comprising a correction operation detection unit for detecting a predetermined operation of the user indicating that the current utterance is a correction of the previous utterance content. When a predetermined operation is detected by the correcting operation detecting unit, the recognizing unit uses only the constituent words or word groups of the partial hierarchy temporarily regarded as the comparison target pattern candidates, A speech recognition device configured to perform a comparison on the input speech. 10. A speech recognition apparatus according to claim 1, further comprising a step of indicating which one of a plurality of speech input methods that a user can perform for correction. A voice, comprising: a correction method detecting means for detecting a predetermined operation of a user; wherein the correction location determining means determines the correction location according to the correction method detected by the correction method detecting means. Recognition device. 11. A speech recognition apparatus according to claim 1, wherein said notifying means performs the output by outputting the contents of the recognition result as speech. 12. A speech recognition apparatus according to claim 1, wherein said notifying means performs the recognition by displaying the contents of the recognition result by characters, symbols, or the like. apparatus. 13. The speech recognition apparatus according to claim 12, wherein the notifying means displays the recognition result when the speech input for the correction is made, from the first recognition result or the previous recognition result. A speech recognition apparatus characterized in that a corrected part is displayed in a different display mode so as to be distinguishable from other parts.