JP2010078986A - Equipment controller by speech recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve a problem that when electronic equipment is controlled from a remote place, a method using a remote controller gives inconvenience that the remote controller is not found instantaneously and a method for performing audio recognition without using switches, possibly gives annoyance due to frequent wrong response to noise etc. <P>SOLUTION: Speech recognition is carried out first in a manual mode using the remote controller etc., and sound input data as an object of speech recognition and other sound input data are stored in a tutor signal database 120. Based upon the stored tutor signal data, a determination parameter database 124 is generated by performing automatic determination parameter learning 122 and after sufficiently high precision is obtained, an automatic mode is entered to perform speech/non-speech automatic determination 114 on sound input data. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention provides a function that can be operated with only one's own voice without having a device such as a remote control at hand when trying to operate a remote electrical device in a home or office. Regarding technology.

  Infrared remote controls are widely used as means for operating devices such as TVs and air conditioners from a distance. On the other hand, a voice recognition technology that recognizes the utterance content by acquiring human voice with a microphone and comparing it with a model prepared in advance has been established, and there is an attempt to control the device using it. . When controlling a device using voice recognition, for example, as disclosed in Patent Document 1, for example, a voice recognition switch is generally prepared and a user explicitly instructs the start of voice recognition. Further, as disclosed in Patent Document 2, there is a technique in which the utterance of a specific word is used instead of the voice recognition switch.

Japanese Patent Laid-Open No. 4-7998 JP 2000-322078 A

  There are problems with controlling a device using a remote control, such as when the remote control cannot be found during a dredging, and there are too many buttons on the remote control to understand how to use. These problems can be reduced by using the voice recognition function. However, since the conventional method requires the voice recognition switch to be activated, it is necessary to have a device having at least one button at hand. . In this case, the problem that the device cannot be found at the time of dredging still remains. If the voice recognition switch is always activated without using the voice recognition switch, there is a problem that the user's undesired behavior frequently occurs due to false reaction to ambient noise or utterances such as chats not intended for device operation. . In order to avoid this problem, there is a method of using the utterance of a specific word instead of the voice recognition switch, but it is still difficult to eliminate a false reaction to noise or the like. In addition, even a simple operation such as “turn on electricity” always requires at least two utterances, which impairs user convenience.

  An object of the present invention is to provide a device control apparatus based on voice recognition that is highly accurate and convenient.

  In the present invention, in a specific environment where speech recognition is used, if it is possible to sufficiently investigate the nature of the sound that may cause a false reaction, and also sufficiently investigate the nature of the speech input that is desired to react correctly, Utilizing the property of being able to accurately discriminate between voice input and other input. In order to examine these properties, it is necessary to collect a sufficient amount of data that is known to be responsive to speech recognition and data that is known not to be responsive.

  Therefore, in the present invention, device control by voice is first performed in combination with a voice recognition switch. Thereby, it is possible to sufficiently collect sound data to be reacted. Further, by continuing to capture sound data even when the voice recognition switch is not pressed, it is possible to sufficiently collect data that should not be reacted. Based on these data, the module for determining the activation of voice recognition is sufficiently learned, and when there is no fear of erroneous reaction, the user is informed that the voice recognition switch is unnecessary. If the user confirms this and designates that voice recognition is to be automatically started thereafter, the device can be controlled by voice without a voice recognition switch.

  That is, in order to achieve the above object, in the present invention, a switching unit that switches a speech recognition activation method between manual and automatic, and sound data that is a candidate for input speech for speech recognition are detected regardless of the activation method. A detection unit, an automatic determination unit that automatically determines whether or not the detected sound data is a target of speech recognition, and whether or not the detected sound data is a target of speech recognition Is provided with an accumulator that accumulates data in a database when it is self-evident and a learning unit that learns parameters used in an automatic determination unit based on the data accumulated in the database. .

  According to the present invention, after a while after purchase and installation of the device control apparatus, it is possible to perform voice control that does not require any equivalent of a remote controller, and the problem that the remote controller cannot be found in the event of a drought does not occur.

  Hereinafter, an effective embodiment of the present invention will be described with reference to the drawings. First, the basic configuration of the present invention will be described. In the following description, function blocks may be referred to as “parts” or “means”, such as “automatic determination of voice / non-voice” as an automatic determination unit or automatic determination means.

  FIG. 1 is a conceptual diagram showing a basic configuration of the present invention. In this configuration, the input sound data from the voice input device (such as a microphone) 102 is continuously captured while the device is turned on. Speech segment candidate detection 104 is performed on the obtained sound data. As a method for extracting speech segment candidates, for example, after sound data is digitized by an A / D converter, frame data is cut out with a time width of about 10 milliseconds, and the square sum of the amplitude of the speech waveform is obtained. For example, a method may be considered in which the frame power is used and the value is compared with a preset threshold value. Such detection of a speech section is a method widely known by engineers in the field. At this time, the role of the speech section candidate detection unit 104 is to extract candidates to the last, and since it is assumed that what is to be processed as speech and what is not to be processed by subsequent processing are unnecessary. There is no problem in detecting a lot of data, but on the contrary, it is necessary to avoid losing necessary data as much as possible. In order to perform such detection, a method such as setting the threshold value to be extremely small may be used.

  Next, in the operation mode check 106, it is confirmed whether the operation mode of the target device is set to “manual” or “automatic”. The operation mode is set in the operation mode setting 108, details of which will be described later.

  FIG. 2 is an extraction of the main processing procedure when the operation mode is “manual” from FIG. First, the user input detection 112 detects whether or not there is an input from the user input device 110. Here, the user input device is a device for the user to instruct activation of voice recognition, and assumes a remote control button, for example, as will be described in detail later. Here, when the activation of voice recognition is instructed, for example, when a button on the remote control is pressed, the voice recognition 116 is executed using the input sound data. At the same time, the input voice data is additionally stored in the teacher signal database 120 as a “voice” teacher signal. In the voice recognition 116, the voice recognition unit executes voice recognition with reference to a preset list of equipment control commands, and the resulting control command is executed by the equipment control 118. . On the other hand, when activation of voice recognition is not instructed, such as when a button on the remote control is not pressed, the voice recognition 116 is not executed, and the input voice data is used as a “non-voice” teacher signal as a teacher signal database. 120 is additionally stored.

  FIG. 3 shows the main processing procedure extracted from FIG. 1 when the operation mode is “automatic” in the operation mode check 106. First, the automatic determination means for performing the voice / non-voice automatic determination 114 is activated. Here, the computer automatically determines whether the input sound data is “voice” or “non-voice”. Here, “speech” refers to all sound data that should be input for speech recognition, and “non-speech” refers to all sound data that should not be input for speech recognition. That is, the user's general speech, laughter, cough, etc. should be interpreted as being included in the category of “non-speech” as well as ambient noise. In the automatic determination, data stored in the determination parameter database 124 is used, and details thereof will be described later. If it is determined that the input sound data is “voice” as a result of the automatic voice / non-voice determination, the voice recognition 116 is executed using the sound data. The device control 118 is executed according to the recognition result as in the case of the operation mode “manual”. However, in the case of “automatic”, since it is only an estimation that the input sound data is “speech”, the data is not additionally stored in the teacher signal database 120. If the result of the voice / non-voice automatic determination is “non-voice”, the device does nothing and waits for the next input.

  FIG. 4 shows a series of processing procedures relating to learning of determination parameters for voice / non-voice discrimination extracted from FIG. The teacher signal stored in the teacher signal database 120 is used in the determination parameter learning 122. The determination parameter learning 122 is activated when a certain amount of teacher signals are accumulated. However, while it is desirable that the processing from the voice segment candidate detection 104 to the device control 118 described above be performed without a large delay, the determination parameter learning 122 is not required to have such responsiveness. In some cases, the activation of the determination parameter learning 122 may be delayed. In the determination parameter learning 122, various feature amounts are extracted from the sound data included in the teacher signal database 120, and the relationship between the set of feature amounts and a label of speech or non-speech is learned. Such a classification problem into two classes has been well studied in the field of machine learning, and various techniques can be easily applied to engineers in the field. After describing a method using linear discriminant analysis, which is the most typical of them, the remaining part of FIG. 4 will be described again.

  FIG. 5 is a diagram showing a state of determination parameter learning by linear discriminant analysis. Here, an example is shown in which two types of feature quantities are used, with the feature quantity 1 on the horizontal axis and the feature quantity 2 on the vertical axis. The feature value here refers to features that are often used in speech analysis such as cepstrum and fundamental frequency from simple values such as average volume and observed interval length, mixed Gaussian distributions and hidden Markovs created from sound databases. Likelihood using a model or the like may be used, but the following discussion is not detailed here because it does not depend on what is used as a feature quantity. In FIG. 5, it is assumed that “teaching” is 13 and “non-speech” is 26, for a total of 39 teacher data. Linear discriminant analysis draws a straight line on this plane so that black circles (voice) are gathered on one side and white circles (non-voice) are gathered on the opposite side. As a quantitative standard, the distance between each point and the straight line (one side of the straight line is defined as positive and the other side is defined as negative) is calculated, and the difference between the sum of the distances for the black circles and the sum of the distances for the white circles is maximized. What should I do? Further, as shown in the figure, when the total number of black circles is different from the total number of white circles, one black circle is counted as two to balance the overall weight. Such a criterion provides a straight line that best separates the black and white circles. In general, when the feature quantity is multidimensional, a hyperplane that divides the multidimensional feature space into two is obtained. The diagonal lines shown in the figure are the results obtained by linear discriminant analysis. In this example, if it is determined that the upper right corner of the straight line is a black circle and the lower left corner is a white circle, only a total of two, one black circle and one white circle, are erroneously classified. The judgment performance for teacher data is about 92% for black circles and about 96% for white circles, and averages about 94%.

  In FIG. 4, when the determination parameter learning 122 is completed, the obtained result is stored in the determination parameter database 124. In the example of FIG. 5, the slope and intercept values representing diagonal lines are stored as parameters. With this parameter, when new sound data comes in, it is converted into a feature value and expressed as a point on the plane, and by looking at which side of the line the point is the sound and It can be determined whether it is non-voice. Further, in the example of FIG. 5, the result that the determination performance with respect to the teacher data is about 94% is obtained, and this corresponds to the role of the determination accuracy estimation 126. In the determination accuracy estimation 126, the accuracy of the result obtained when new data is determined is estimated from the determination accuracy for the teacher data. This is because, when it is expected that sufficient accuracy cannot be obtained by the voice / non-voice automatic determination, it is desirable to use manual activation by a button on the remote controller. The result obtained by the determination accuracy estimation 126 is presented to the user on the determination accuracy display 128. However, since it is not always easy for the user to show a value such as 94% at this time, for example, a red LED is turned on when the prediction accuracy is lower than a preset threshold, and a blue LED is turned on when the prediction accuracy is high. Also good.

  Hereinafter, returning to FIG. 1, the operation mode setting will be described. If the user looks at the determination accuracy display 128 and wishes to change the operation mode, the user changes the operation mode by operating a button or the like (operation mode setting 108). For example, when the operation mode is changed from manual to automatic, in the subsequent processing, the “automatic” processing is advanced in the operation mode check 106. When the user is dissatisfied with the performance of the apparatus, the operation mode can be returned to “manual” to operate again using the remote controller or the like. It should be noted that the change from “manual” to “automatic” can usually be performed within a few days after the voice recognition is performed several hundred times.

  FIG. 6 shows the flow of processing from acquisition of sound data to device control in the apparatus having the functional configuration of FIG. 1 in the form of a flowchart. Device control is actually executed when a voice segment candidate is found, the operation mode is manual, the activation button is pressed, and the voice recognition result is a valid command (206-208-210-212-218). -220-222), when the voice section candidate is found, the operation mode is automatic, the voice / non-voice judgment result is voice, and the voice recognition result is a valid command (206-208-214-216) -218-220-222). If none of these applies, and after device control is actually executed, the state returns to the sound data acquisition 202 again, and thereafter the same operation is repeated until the apparatus stops. In addition to these operations, the determination parameter learning flow described with reference to FIG. Here, the term “anytime” means when a certain amount of necessary data is accumulated.

  FIG. 7 is a diagram showing a configuration of a device control apparatus based on voice recognition when implemented as a part of a television apparatus as the first embodiment. Most of this configuration is realized as a module inside the television main body 332. The internal modules include a central processing unit (CPU) and a storage unit (memory) that constitute the processing unit of the computer described above. The CPU has functions of a speech segment candidate detection unit 304, an operation mode check unit 306, a teacher signal detection unit 312, a speech / non-speech automatic determination unit 314, a speech recognition unit 316, a determination parameter learning unit 322, and a determination accuracy estimation unit 326. Is executed as a program process. The teacher signal database 320 and the determination parameter database 324 are accumulated in the storage unit. The operation mode display unit 328 can be configured to be displayed on an independent display element or a television display. Of course, among the functions executed by the CPU, the function of the voice recognition unit and the like may be realized by a dedicated circuit.

  In addition, a remote control device 310 and a human detection sensor 330 as a human detection unit are used as necessary. As the remote control device 310, a device attached to a general television can be used, or a device having only one button specialized for the purpose of starting speech recognition may be used. In general, in the former case, signals are often transmitted by infrared rays, and in the latter case, it seems that it is common to use weak radio used in keyless entry for automobiles. The human detection sensor 330 detects whether or not there is a person in the room by the reaction of an infrared sensor, for example. Alternatively, it is an apparatus that provides information on whether there is a person in the room in cooperation with an office entrance / exit management system or the like. In this embodiment, if information about whether or not there is a person in the room can be obtained, no particular difference will occur regardless of which existing technology is used. I will omit the explanation.

  The infrared / wireless receiving unit 311 receives infrared and wireless signals sent from the remote control device 310. This is a module that plays the role of the user input detection 112 shown in FIG. 1. In addition, when a TV control signal is sent from a normal remote control device 310, the corresponding control signal is sent to the channel / volume. The data is sent to the control unit 318. The channel / volume control unit 318 is a module included in a normal television device, and changes a television reception channel, volume, and the like according to a control signal. When an operation mode change signal is sent from remote control device 310, a corresponding signal is sent to operation mode setting unit 308. In addition to this, the operation mode setting unit 308 can be directly operated with a button or the like not shown.

  The roles of the microphone 302, the speech segment candidate detection unit 304, and the operation mode check unit 306 are the same as the corresponding blocks shown in FIG. When the operation mode is “manual”, the teacher signal detection unit 312 determines whether to activate the voice recognition unit 316 according to the signal from the infrared / wireless reception unit 311. On the other hand, when the operation mode is “automatic”, the operation of the voice recognition unit 316 should be suppressed or left to the determination result of the voice / non-voice automatic determination unit 314 according to the signal from the human detection sensor 330. Output what to do. This is because if it is known that there are no people in the room, it may be desirable to regard the input sound data as “non-speech” regardless of the result of analyzing the sound. is there. Also, if the function of the human detection sensor 330 cannot be completely trusted, the operation standard of the voice / non-voice automatic determination unit 314 is temporarily corrected instead of immediately determining “non-voice”, It is also possible to take a method of increasing the possibility of the determination of “non-voice”.

  As an additional function, if a voice recognition activation signal is received from the infrared / wireless receiving unit 311, when the operation mode is “automatic” and “manual” for a certain period of time after that, It is also possible to take a method of performing the same operation as. This is because if it is known that the remote control device 310 is present at the user's hand, it is estimated that there is no need to deal with the problem that the remote control device 310 cannot be found at the time of a drought. This is because the need to violate the risk of reaction is considered low. In this case, not only the operation mode is changed to “manual” but also a method of temporarily correcting the operation reference so that “sound” is difficult to be detected although it is “automatic” is possible. . When the voice recognition unit 316 operates, a signal corresponding to the result is sent to the channel / volume control unit 318. In this case, the fixed time and temporary are usually several minutes.

  When the operation mode is “manual”, the teacher signal of “speech” or “non-speech” is additionally stored in the teacher signal database 320 in the same manner as in the case of FIG. Even if “automatic”, if the determination of “non-speech” is made forcibly under the condition that the signal from the human detection sensor 330 or the remote control device 310 is used within a certain period of time, the teacher signal detection unit 312 The “non-speech” teacher signal is additionally stored in the teacher signal database 320. Here, the fixed time after use is usually about several minutes. Subsequent roles of the determination parameter learning unit 322, the determination parameter database 324, the determination accuracy estimation unit 326, and the operation mode display unit 328 are substantially the same as the corresponding portions in FIG.

  FIG. 8 is a diagram showing a configuration in the case of realizing a device control apparatus based on voice recognition as a general-purpose device control apparatus for controlling various devices in a room or the like according to the second embodiment.

  The major difference from the embodiment of FIG. 7 is that all the modules included in the TV main body 332 except the channel / volume control unit 318 are included in the general-purpose control device 432 newly provided in this configuration. This is the point that has been moved. In the configuration of FIG. 7, the processing by these modules is performed inside the television apparatus, and the channel and volume are controlled according to the result. In this configuration, the processing of these modules is performed. As a result, when it is necessary to perform device control, the infrared / radio transmission unit 432 transmits the signal as an infrared or wireless signal, or the voice recognition unit 416 transmits a signal directly to an external device through a signal line. The external devices include various electric devices existing in the room such as the air conditioner 436 and the lighting 438 in addition to the television 434, and the device control signal indicates what operation is performed on which device. Sent as a combination. Each device performs an operation corresponding to the command only when it receives a control command for itself.

  FIG. 9 shows the flow of processing from acquisition of sound data to device control in the form of a flowchart in the apparatus having the configuration of the embodiment of FIGS. The difference from the flowchart shown in FIG. 6 is that when the operation mode is automatic, two processes of time determination 514 after using the remote control device and human detection determination 516 by the human detection sensor are added. is there. When it is determined by these determinations 514 and 516 that it is less than T seconds after using the remote control (the value of T is set in advance according to the use environment, but is about several minutes as described above). If the sensor does not detect a person, it is obvious that the input sound data is non-speech, no sound / non-sound determination is performed, and the sound data is non-speech teacher data. Accumulate as. When T seconds have passed since the remote control was used and the sensor detects a person, voice / non-voice determination 518 is executed as in FIG.

  FIG. 10 is a diagram showing a detailed configuration in the case of realizing a device control apparatus based on voice recognition as a part of the car navigation apparatus according to the third embodiment. In general, in a car navigation apparatus having a voice recognition function, a voice recognition start button 630 is installed at a position that is easy to operate while driving, apart from a touch panel and operation buttons 610 for operating the navigation function in general. Often there are. The signal of the voice recognition activation button is directly sent to the teacher signal detection unit 612, whereas the signal from the touch panel / operation buttons 610 is for directly operating the car navigation system or for setting the operation mode. Or something. Unlike the embodiments of the television and the general-purpose control device shown in FIG. 7 and FIG. 8, all the operation signals including the operation signal of the touch panel are transmitted by wire, but the operation after receiving these signals is the television or the like. Is almost the same as Further, when installing the car navigation device, it is also possible to connect to an operation status transmission unit 632 that provides information related to the operation status from the automobile body, for example, the engine speed. In this case, for example, when the traveling speed is larger than the reference value, the touch panel operation is invalidated in many car navigation devices.

  In the present embodiment, in addition to this, the adverse effects caused by the erroneous reaction of voice recognition are quantified based on the traveling speed, the turning angle of the steering wheel, the depression amount of the brake pedal, and the like. In other words, when the driving speed is high, turning a curve, or stepping on a brake, the driver needs more concentration than usual, so there is no false reaction in voice recognition. This is very important. Therefore, the teacher signal detection unit 612 calculates the weighted sum of the traveling speed, the turning angle of the steering wheel, the amount of depression of the brake pedal, and the like as an adverse reaction adverse effect degree, and when this value is larger than a predetermined value, Control is performed so that voice recognition is not activated regardless of the result of voice / non-voice discrimination. Alternatively, the teacher signal detection unit 612 controls to add a value that corrects the threshold used for speech / non-speech discrimination so that “speech recognition is unlikely to be activated” in proportion to the adverse reaction adverse effect level. Thereby, when there are special circumstances that should be concentrated on driving, it is possible to further reduce the risk that a dangerous situation may occur due to a malfunction of the voice recognition unit 616.

  FIG. 11 is a flowchart showing an example of the operation of the voice / non-voice automatic determination unit (314, 414, 614 in FIGS. 7, 8, and 10) in each embodiment. Here, an example of using linear discriminant analysis is shown. First, after obtaining sound data (702), a feature value is calculated based on the sound data (704). The feature quantities used here include average power, cepstrum, fundamental frequency, similarity to previously learned speech models and non-speech models, recognition result reliability when speech recognition is applied, and the like. However, the configuration of the present invention can be used as it is even if other feature quantities capable of expressing the voice quality are used. Next, after normalizing these feature quantities (706), an inner product with a linear discriminant coefficient learned in advance is obtained (708). If the value obtained in this way is larger than the threshold learned in advance, it is determined that the sound is speech, and if it is smaller, it is determined that the speech is not speech (710, 712, 714). Although an example using linear discriminant analysis has been described here, the configuration of the present invention can be used as it is even if other classification methods (for example, a support vector machine or the like) are used. The application of these classification methods can be easily realized by a developer engaged in the relevant field, and therefore will not be described in detail here.

  FIG. 12 is a flowchart showing an example of the operation of the determination parameter learning unit (322, 422, 622 in FIGS. 7, 8 and 10) in each embodiment. Again, an example of using linear discriminant analysis is shown. First, after all the learning data has been acquired (802), the feature values for them are obtained and normalized (804). After the feature quantity data thus obtained is vectorized, a covariance matrix and an intra-group average of speech data and an intra-group average of non-speech data are obtained for the whole (806). Next, after obtaining an inverse matrix of the covariance matrix (808), the obtained inverse matrix is multiplied by the difference of the intragroup average (810). As a result, a linear discrimination coefficient is obtained and stored (812). Note that the performance does not change even if a constant is applied to the entire linear discriminant coefficient, so the former is smaller when the linear discriminant coefficient is applied to speech data than when the linear discriminant coefficient is applied to non-speech data. It is convenient to reverse the sign by multiplying the whole linear discrimination coefficient by -1. Next, it is necessary to obtain an optimum threshold value (814), but since this will be described in detail in the determination accuracy estimation unit, description thereof is omitted here.

  FIG. 13 is a flowchart showing an example of the operation of the determination accuracy estimation unit (324, 424, 624 in FIGS. 7, 8, and 10) in each embodiment. Similar to the case of FIG. 12, after obtaining normalized feature vectors for all learning data (902), a linear discrimination coefficient is applied to obtain a linear discrimination score (904). Next, the minimum value Cmin of the threshold C candidate values is set to C (906). Here, for example, an average value of linear discrimination scores for the entire non-voice data may be used. Next, an average discrimination rate R when the threshold candidate C is used is obtained. Here, the average discrimination rate is an average value of a rate at which audio data is correctly determined as speech and a rate at which non-audio data is correctly determined as non-speech. In the case where one of the performance for audio data and the performance for non-audio data is more important, the performance can be adjusted by taking a weighted average of both. Such processing is repeated for various candidate values C, and C when R is maximum is set as the optimum threshold Cbest. In addition, the maximum value Rmax is set as the estimation determination accuracy (908-918).

  The present invention described above in detail can be used as a home appliance such as a television having a function of facilitating operation from a remote place. It can also be used as a general-purpose control device for controlling several devices collectively. Furthermore, it can be used as a car navigation device for automobiles.

It is a block diagram of the typical system for demonstrating the basic composition of this invention. FIG. 2 is a diagram showing details of a process flow when the operation mode is manual in the configuration of FIG. 1. It is a figure showing the detail of the flow of a process in the structure of FIG. 1 when an operation mode is automatic. FIG. 2 is a diagram showing details of a flow of processing related to learning of determination parameters in the configuration of FIG. 1. It is explanatory drawing of the example of the determination parameter learning using the linear discriminant analysis with the structure of FIG. It is a figure which shows the flowchart at the time of controlling an apparatus with an audio | voice with the structure of FIG. It is a block diagram at the time of implement | achieving an apparatus control apparatus as a part of television apparatus concerning a 1st Example. It is a block diagram at the time of implement | achieving an apparatus control apparatus as a general purpose apparatus control apparatus concerning a 2nd Example. It is a figure which shows the flowchart explaining operation | movement of the apparatus control apparatus concerning the 1st, 2nd Example. It is a block diagram at the time of implement | achieving an apparatus control apparatus as a part of car navigation apparatus concerning a 3rd Example. It is a figure which shows the flowchart explaining operation | movement of the principal part of an apparatus control apparatus concerning each Example. It is a figure which shows the flowchart explaining operation | movement of the principal part of an apparatus control apparatus concerning each Example. It is a figure which shows the flowchart explaining operation | movement of the principal part of an apparatus control apparatus concerning each Example.

Explanation of symbols

102 ... Voice input device 104 such as a microphone ... Processing block 106 for detecting a voice segment candidate ... Processing block 108 for determining whether the operation mode is manual or automatic ... Processing block 110 for the user to set the operation mode to manual or automatic ... Processing block 112 in which user inputs activation of voice recognition with switch or the like ... Block 114 for detecting input from user ... Block 116 for automatically determining whether input sound data is voice or non-voice ... Block 118: Control of device based on the result of speech recognition 120: Database 122 for storing teacher signals 122: Learning determination parameters using teacher signals 124: Database for storing determination parameters obtained by learning 126 ... The accuracy of the determination based on the current determination parameter Block showing the determination accuracy of block 128 ... estimated that constant to the user.

Claims (20)

A device control apparatus for controlling a device by recognizing a user's voice by a voice recognition unit,
A switching section for switching the voice recognition activation method between manual and automatic;
A detection unit for detecting sound data that is a candidate for input speech for speech recognition;
An automatic determination unit that automatically determines whether the detected sound data is data to be subjected to speech recognition;
When it is obvious whether the detected sound data is data to be subjected to speech recognition, a teacher signal detection unit that accumulates it in a database;
A learning unit that learns determination parameters used in the automatic determination unit based on data stored in the database;
A device control apparatus based on voice recognition. When the teacher signal detection unit manually activates speech recognition by the switching unit, the input sound data is stored in the database as data to be subjected to speech recognition.
The apparatus control apparatus by voice recognition according to claim 1. The teacher signal detection unit stops the automatic speech recognition activation for a certain time after the user manually activates the speech recognition.
The apparatus control apparatus by voice recognition according to claim 1. The teacher signal detection unit accumulates the detected sound data in the database as data that should not be subject to speech recognition for a certain period of time after the user manually activates speech recognition.
The apparatus control apparatus by voice recognition according to claim 1. The teacher signal detection unit makes it difficult for automatic speech recognition to start during a certain time after the user manually starts speech recognition.
The apparatus control apparatus by voice recognition according to claim 1. It further has a human detection unit that detects whether there is a person around the device using information other than sound,
When the teacher signal detection unit detects that there is no person by the human detection unit, the automatic detection of voice recognition is stopped.
The apparatus control apparatus by voice recognition according to claim 1. It further has a human detection unit that detects whether there is a person around the device using information other than sound,
The teacher signal detection unit, when it is detected by the human detection unit that there is no person, accumulates the detected sound data in the database as data that should not be subject to speech recognition.
The apparatus control apparatus by voice recognition according to claim 1. The teacher signal detection unit changes the ease of activation of automatic speech recognition according to the degree of adverse reaction adverse effect that quantifies the adverse effect due to erroneous reaction of speech recognition according to the status of the device to be controlled,
The apparatus control apparatus by voice recognition according to claim 1. A determination accuracy estimation unit that estimates determination accuracy based on the accumulated determination parameter;
The apparatus control apparatus by voice recognition according to claim 1. An operation mode display unit that displays the determination accuracy calculated by the determination accuracy estimation unit;
10. The device control apparatus by voice recognition according to claim 9. A device control device that recognizes a user's voice and controls the device,
A switching unit for switching the voice recognition activation method to manual or automatic, and
A processing unit for processing input speech for speech recognition;
A storage unit for storing the processing result of the processing unit,
When the processing unit detects sound data that is a candidate for the input speech, the activation method is manual, and the activation method is automatic, and the detected sound data should be subject to speech recognition. As a result of determining whether or not it is data, if it is determined that the data should be the target, control to perform voice recognition of the sound data,
When it is clear whether or not the sound data is data to be recognized, the sound data is stored in the storage unit and used in the determination based on the sound data stored in the storage unit. Learn judgment parameters,
A device control apparatus based on voice recognition. When the processing unit manually activates speech recognition by the switching unit, the sound data is stored in the storage unit as data to be subjected to speech recognition.
The apparatus control apparatus by voice recognition according to claim 11. The processing unit accumulates the sound data in the storage unit as data that should not be subject to speech recognition for a certain period of time after the user manually activates speech recognition.
The apparatus control apparatus by voice recognition according to claim 12. The processing unit stops the automatic speech recognition activation for a certain time after the user manually activates the speech recognition.
The apparatus control apparatus by voice recognition according to claim 11. The processing unit performs control so that activation of automatic speech recognition is less likely to occur for a certain time after the user manually activates speech recognition.
The apparatus control apparatus by voice recognition according to claim 11. It further has a human detection unit that detects whether there is a person around the device using information other than sound,
When the processing unit detects that there is no person by the human detection unit, it stops the automatic speech recognition activation,
The apparatus control apparatus by voice recognition according to claim 11. It further has a human detection unit that detects whether there is a person around the device using information other than sound,
The processing unit accumulates the detected sound data in the storage unit as data that should not be subject to speech recognition when the human detection unit detects that no person is present,
The apparatus control apparatus by voice recognition according to claim 11. The processing unit changes the ease of activation of automatic speech recognition according to the degree of adverse reaction adverse effect that quantifies the adverse effect due to erroneous reaction of speech recognition according to the status of the device to be controlled,
The apparatus control apparatus by voice recognition according to claim 11. The processing unit accumulates the determination parameter in the storage unit, and estimates determination accuracy based on the accumulated determination parameter.
The apparatus control apparatus by voice recognition according to claim 11. An operation mode display for displaying the estimated determination accuracy;
20. The device control apparatus by voice recognition according to claim 19.

Images