JP2006058479A - Controller with voice recognition function - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a controller with a voice recognition function which has improved recognition accuracy by making a sound model adaptive in compliance with the sounds, such as a voice and noise, inputted in real use. <P>SOLUTION: The controller A with the voice recognition function is equipped with: a recognition part 4 which compares the feature quantities of the voice signal extracted by a feature quantities extraction part 2 with the sound model stored in a sound model storage part 3 to thereby recognize an input sound; a control part 6 which outputs a control signal according to the recognition result to an illuminator B; and a sound model learning part 8 which relearns the sound model by using the voice signal stored in an input voice storage control part 7, and the corresponding sound model and updates the sound model of the sound model storage part 3. When the control signal different in the control content from the recognition result is outputted from a switch 5 during the time from the output of the recognition result before the lapse of the prescribed limit time, the input voice storage control part 7 corrects the recognition result and causes the sound model learning part 8 to relearn the sound model by using the corrected recognition result and the corresponding voice signal. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a control device with a voice recognition function that recognizes an operation command issued by a person and controls a device to be controlled.

  2. Description of the Related Art Conventionally, a control device with a voice recognition function that recognizes an operation command issued by a person and turns on or off a lighting load to be controlled has been provided (for example, see Patent Document 1).

  Further, as means for operating the control target device, an operation means for directly operating the control target device and a voice recognition means for recognizing an operation command issued by a person to operate the device are provided. A control device with a voice recognition function that operates a device to be controlled using an input and a recognition result by a voice recognition means has also been provided.

  FIG. 6 is a block diagram of such a conventional control device with a voice recognition function. The control device A with a voice recognition function includes a microphone 1, a feature amount extraction unit 2, an acoustic model storage unit 3, and a recognition unit. 4, a switch 5, a control unit 6, an input voice storage control unit 7, and an acoustic model learning unit 8 are provided as main components.

  The microphone 1 receives a sound that is either an operation command (voice) or noise issued by a person to cause the control target device B to perform a desired operation, and the input sound is an audio signal that is an analog electrical signal. Convert to and output.

  When detecting the input of the audio signal from the microphone 1, the feature amount extraction unit 2 performs A / D conversion on the input audio signal, for example, with the number of quantization bits set to 16 and the sampling frequency set to 16 kHz. After frequency conversion is performed in milliseconds and the analysis interval is 10 milliseconds, the feature amount of the audio signal is extracted. Here, for example, a mel frequency cepstrum coefficient or the like can be used as the feature amount of the audio signal, and the extracted feature amount is output to the recognition unit 4. The feature amount extraction unit 2 outputs the audio signal input from the microphone 1 to the input audio storage control unit 7.

  In the acoustic model storage unit 3, for one or a plurality of operation commands for controlling the control target device B, feature quantities of operation commands issued by a large number of speakers are modeled using, for example, an HMM (Hidden Markov Model). The acoustic model including the “command word” acoustic model and the “noise” acoustic model obtained by modeling the characteristic amount of noise assumed in the usage environment is stored. As an example of the acoustic model, for example, when the control target device B is a lighting device, the “light” corresponding to the vocabulary “light” used to instruct the lighting of the lighting device as the “command word” acoustic model. The acoustic model is stored, and the “noise” acoustic model corresponding to the sound other than the vocabulary related to the sound of the door and the “light” is stored.

  The recognition unit 4 compares the feature amount extracted by the feature amount extraction unit 2 with the acoustic model stored in the acoustic model storage unit 3, and compares the extracted feature amount with a model having a high degree of similarity. (Operation command or noise) is output to the control unit 6 and the input voice storage control unit 7 as a recognition result. That is, when two acoustic models “AKARI” and “NOISE” are stored in the acoustic model storage unit 3, if the input speech is related to “AKARI”, the recognition result “AKARI” is obtained. If the input sound is a door closing sound, a recognition result of “noise” is obtained.

  The switch 5 is provided to directly operate the control target device B to be controlled, and a control signal corresponding to the operation of the switch 5 is output to the control target device B. For example, when the control target device B is a luminaire, an ON operation signal for turning on (ON) the luminaire and a OFF operation signal for turning off (OFF) are output as control signals from the switch 5.

  The control unit 6 outputs a control signal for performing an operation corresponding to the recognition result to the control target device B based on the recognition result of the recognition unit 4. That is, a control signal is input to the control target device B from both the switch 5 and the control unit 6, and an operation corresponding to the latest control signal from the control signal from the switch 5 or the control signal from the control unit 6 is performed. To do.

  The input voice storage control unit 7 stores the voice signal input from the feature amount extraction unit 2 and the recognition result input from the recognition unit 4 in association with each other, and newly inputs the recognition result and the recognition result. Are output to the acoustic model learning unit 8.

Then, the acoustic model learning unit 8 uses the recognition result input from the input voice storage control unit 7 and one or more audio signals corresponding to the recognition result to convert the acoustic model corresponding to the recognition result to MLLR ( It is adapted by the Maximum Liklihood Renier Regression method and the MAP estimation method (Maximum A Posteriori Probability Pstimation) to improve recognition accuracy.
JP 2002-289371 A

  In the control device with a speech recognition function having the above configuration, since speech recognition is performed using the acoustic model incorporated in the acoustic model storage unit 3, the sound used to create the acoustic model and the microphone 1 when input are used. There is a problem that the accuracy of recognition is lowered when the similarity to a sound (such as a user's voice) is low. For example, in the case of the “command word” acoustic model, human voice quality, inflection, and the like vary from person to person. Therefore, an acoustic model that matches the characteristics of the person actually used (voice quality, intonation, etc.) is incorporated in the acoustic model storage unit 3 in advance. Therefore, there is a problem that the accuracy of speech recognition by the recognition unit 4 is reduced. The same applies to the noise model, and the nature of the input noise varies greatly depending on the use environment of this apparatus. Therefore, a noise model that matches the input noise in actual use is incorporated in the acoustic model storage unit 3 in advance. As a result, the accuracy of speech recognition may be reduced.

  Therefore, in the control device with a voice recognition function described above, the acoustic model learning unit 8 uses the voice signal or noise voice signal input during actual use and the recognition result of the recognition unit 4 for the voice signal to generate an acoustic model. By re-learning, the acoustic model is sequentially updated and adapted to the actual usage environment, but if an incorrect recognition result is output for the input audio signal, this incorrect recognition result is displayed. Since the acoustic model is re-learned by using the acoustic model, the acoustic model becomes inaccurate, and as a result, the recognition performance of speech recognition may be deteriorated.

  The present invention has been made in view of the above problems, and its object is to improve recognition accuracy by adapting an acoustic model according to sounds such as voice and noise input during actual use. Another object is to provide a control device with a voice recognition function.

  In order to achieve the above object, according to the first aspect of the present invention, a sound that is either a voice or a noise uttered by a person to operate the control target device is input, and the input sound is converted into a voice signal that is an electrical signal. A sound conversion unit that converts and outputs, a feature amount extraction unit that extracts a feature amount of an input sound from an audio signal, and an acoustic model unit that stores an acoustic model in which the feature amount is modeled for each of a plurality of sounds and noises , A recognition unit that recognizes an input sound by comparing the feature amount extracted by the feature amount extraction unit with the acoustic model stored in the acoustic model unit, and a control for performing an operation according to the recognition result of the recognition unit A control unit that outputs a signal to the control target device, an operation unit that directly outputs a control signal corresponding to the operation to the control target device, an input voice storage unit that stores the voice signal and the recognition result of the recognition unit in association with each other, and In the input audio storage The acoustic model learning unit that re-learns the acoustic model of the input sound using the existing recognition result and the audio signal corresponding to the recognition result and updates the acoustic model stored in the acoustic model unit, and the control from the recognition unit And a timer unit that limits a predetermined time limit from the time when the recognition result is input to the control unit, and a control signal that is different in control content from the recognition result of the recognition unit is output from the operation unit during the time limit operation of the timer unit. In this case, the input speech storage unit corrects the stored recognition result based on the content of the control signal output from the operation unit, and the acoustic model learning unit uses the corrected recognition result and the input speech. It is characterized by re-learning the model.

  By the way, when the recognition unit misrecognizes the input sound and the control unit causes the control target device to malfunction due to an incorrect recognition result, the user directly operates the operation unit to bring the control target device into a desired operation state. Although it is conceivable, according to the present invention, during the time limit operation of the timer unit, that is, between when the recognition result is input from the recognition unit to the control unit and until a predetermined time limit elapses, When a control signal with a different control content is output from the operation unit, the recognition result stored in the input voice storage unit is corrected, and the voice signal of the erroneously recognized input sound and the corrected recognition result are displayed. By using the acoustic model learning unit to re-learn the acoustic model, the possibility of misrecognition when the same sound is input next time is reduced, improving the accuracy of speech recognition and improving the reliability of the device be able to.

  According to a second aspect of the present invention, there is provided the operation unit according to the first aspect of the present invention, from when the recognition unit recognizes the input sound as noise in a state where the control target device is turned off until the time limit operation of the timer unit ends. When the control signal for turning on the device to be controlled is output from, the input voice storage unit corrects and stores the recognition result recognized as noise into a voice for an on operation, The same effect as the invention is achieved.

  According to a third aspect of the present invention, in the first aspect of the invention, from when the recognition unit recognizes the input sound as a sound for turning on in a state where the control target device is turned off, until the time limit operation of the timer unit ends. When a control signal for turning off the control target device is output from the operation unit during the period, the input voice storage unit corrects and recognizes the recognition result recognized as the voice for the on operation and stores the noise, The same effect as that of the invention of claim 1 is obtained.

  The invention of claim 4 is the invention of claim 1, further comprising a control state storage unit that stores the operation state of the device to be controlled, and the input sound is controlled by the current operation state stored in the control state storage unit. When the recognition unit recognizes that the voice is for operating the device, the input voice storage unit corrects and stores the recognition result recognized as the voice for operation as noise.

  In general, when the user operates the device to be controlled by voice, it is considered that a command to operate the current operation state is not issued. However, according to the invention of claim 4, the operation is performed to the current operation state. If the input sound is recognized to be a command to be input, the input sound storage unit corrects and stores the recognition result recognized as the sound for operation as noise. When the acoustic model learning unit relearns the acoustic model using the signal and the corrected recognition result, it is possible to reduce the possibility of erroneous recognition when the same sound is input next time.

  According to a fifth aspect of the present invention, in the first aspect of the present invention, a human sensor is provided to detect the presence or absence of a person in a detection area including at least the sound collection range of the sound conversion unit, and the human sensor detects the presence of a person If the recognition unit recognizes the input sound as a voice for operation in a state in which the recognition is not performed, the input voice storage unit corrects and recognizes the recognition result recognized as the voice for the operation as noise. To do.

  By the way, when there is no person in the sound collection range of the sound converter, it is obvious that the sound input to the sound converter is noise. According to the invention of claim 5, the human sensor detects the presence of a person. If the input sound is recognized as operation sound without detection, the input sound storage unit corrects and stores the recognition result as noise. When the acoustic model learning unit re-learns the acoustic model using the recognized recognition result, it is possible to reduce the possibility of erroneous recognition when the same sound is input next time.

  According to a sixth aspect of the present invention, in the first aspect of the invention, when the recognizing unit determines that the new input sound is noise until a predetermined time elapses after the recognizing unit recognizes the input sound as speech. The input voice storage unit deletes the recognition result recognized as noise and the data of the voice signal corresponding to the recognition result.

  According to the invention of claim 6, when the input sound is continuously input within a predetermined time, the input sound storage unit deletes the recognition result recognized as noise and the data of the sound signal corresponding to the recognition result. Therefore, in a device in which voice for operation is input continuously, the acoustic model learning unit will erroneously delete the voice signal of the input sound that was mistakenly recognized as noise and the recognition result data. It is possible to prevent re-learning based on the received data, and to reduce the possibility of erroneous recognition when the same sound is input next time.

  As described above, according to the present invention, the recognition unit recognizes during the time limit operation of the timer unit, that is, until the predetermined time limit elapses after the recognition result is input from the recognition unit to the control unit. When a control signal whose control content is different from the result is output from the operation unit, the recognition result stored in the input voice storage unit is corrected, and the voice signal of the misrecognized input sound and the corrected recognition result When the acoustic model learning unit re-learns the acoustic model using, the possibility of misrecognition when the same input sound is input next time is reduced, and the accuracy rate of speech recognition is improved and the reliability of the device is improved. There is an effect that can improve the nature.

  Embodiments of the present invention will be described below with reference to the drawings.

(Embodiment 1)
FIG. 1 is a block diagram of a control device with a voice recognition function according to the present embodiment. A control device A with a voice recognition function includes a microphone 1, a feature amount extraction unit 2, an acoustic model storage unit 3, and a recognition unit 4. The switch 5, the control unit 6, the input voice storage control unit 7, the acoustic model learning unit 8, and the timer unit 9 are provided as main components. In addition, since it is substantially the same as the control apparatus A with the voice recognition function of FIG. 6 described in the background art except that the timer unit 9 is added, the same components are denoted by the same reference numerals, and the description thereof is as follows. Omitted.

  The timer unit 9 receives the trigger signal from the control unit 6 and starts a time limit operation. That is, when the recognition result is input from the recognition unit 4, the control unit 6 outputs a control signal for performing an operation according to the recognition result to the control target device (for example, the lighting device B), and the timer unit 9. Output a trigger signal. When the timer unit 9 receives the trigger signal, the timer unit 9 starts a time limit operation for a predetermined time. When the time limit operation is completed, the timer unit 9 outputs a timer operation completion signal to the input voice storage control unit 7.

  On the other hand, the input voice storage control unit 7 stores the voice signal input from the feature amount extraction unit 2 and the recognition result input from the recognition unit 4 in association with each other and receives a timer operation completion signal. The recognition result input this time and one or more audio signals corresponding to the recognition result are output to the acoustic model learning unit 8. The input voice storage control unit 7 is also provided with an operation input of the switch 5, and the timer operation is completed by the timer unit 9 from the time when the recognition result is input from the recognition unit 4 (time point when the timer operation starts). When a control signal having a control content different from the recognition result is given from the switch 5 until the signal is input, the recognition result is rewritten so as to match the control content of the control signal. When input, the corrected recognition result and the audio signal corresponding to the recognition result are output to the acoustic model learning unit 8 to re-learn the acoustic model.

  FIG. 2 shows a construction example when the present apparatus A is applied for controlling a lighting fixture. The lighting apparatus B to be controlled is installed on the ceiling 21 of the room 20 and the wall near the door 22 leading to the outside. 23 is provided with a switch 5 for directly operating the illumination device B, and a microphone 1 is provided in the vicinity (upper side) of the switch 5. Since the switch 5 is often operated when entering and exiting the room 20, the switch 5 is installed in the vicinity of the door 22, and for the same reason, the operation command issued by the person in the vicinity of the door 22 can be reliably collected. The microphone 1 is also installed in the vicinity of the door 22 (switch 5).

  However, since the microphone 1 is installed in the vicinity of the door 22, the sound for opening and closing the door 22 is easily collected by the microphone 1, and the opening and closing sound of the door 22 is mistakenly recognized as an operation command, and lighting is performed contrary to the user's intention. There is a possibility that the device B is turned on or off. In addition, there is a possibility that an operation command issued by a person is mistakenly recognized as an opening / closing sound of the door 22 and the lighting device B does not turn on or off, contrary to the user's intention.

  Here, when the recognition unit 4 misrecognizes noise as a command word or conversely misrecognizes an operation command as noise, when the operation of the lighting device B is different from the user's intention, the user switches to the switch 5. It is considered that the operation of the lighting device B is matched with the operation intended by the user by directly operating. Therefore, the time limit of the timer unit 9 is determined from the time required for the user to directly operate the switch 5 to change the operation of the lighting device B from the time when the recognition result of the recognition unit 4 is input to the control unit 6. If the recognition result of the recognition unit 4 is different from the operation intended by the user, the operation input from the switch 5 is input voice storage control before the timer operation completion signal is input. The operation input from the switch 5 is not given until the timer operation completion signal is inputted when the recognition result of the recognition unit 4 and the operation intended by the user match with each other. I can judge.

  As described in the background art, the input voice storage control unit 7 stores the voice signal input to the feature amount extraction unit 2 and the recognition result by the recognition unit 4 in association with each other. Are output to the acoustic model learning unit 8 and the acoustic model is re-learned by the acoustic model learning unit 8, but the input speech storage control unit 7 recognizes from the recognition unit 4. If a control signal having a control content different from the recognition result is given from the switch 5 until the timer operation completion signal is inputted from the timer unit 9 from the time when the result is inputted, the recognition result is converted to the control content of the control signal. Therefore, the recognition result can be corrected to match the operation command issued by the user. Therefore, the acoustic model learning unit 8 re-learns the acoustic model using the corrected recognition result and the corresponding speech signal, so that the acoustic model is accurate and the recognition performance of speech recognition can be improved. .

  For example, the recognition unit 4 is input from the feature amount extraction unit 2 even though the user utters an operation command (for example, “Akari”) to turn on the lighting device B in a state where the lighting device B is turned off (off). If the determined feature amount is compared with the acoustic model stored in the acoustic model storage unit 3 and is determined to be similar to the acoustic model of “noise”, the control unit 6 and the input voice storage control unit 7 are notified of “noise”. The recognition result is output. When the recognition result “noise” is given to the control unit 6, the control unit 6 outputs a trigger signal to the timer unit 9 and does not output any control signal to the lighting device B. Remains off. At this time, since the lighting device B does not light even though the user utters “light”, it is determined that the on operation command “light” is not correctly recognized, and the switch 5 is directly turned on to turn on the lighting device. Try to light B. When the switch 5 is turned on, an on operation signal of the switch 5 is given to the lighting device B, the lighting device B is turned on, and an on operation signal is given to the input sound storage control unit 7. The input voice storage control unit 7 stores the voice signal input from the feature amount extraction unit 2 and the recognition result of the recognition unit 4 in association with each other. When the timer operation completion signal is received from the timer unit 9, the current recognition is performed. The result and the feature quantities of one or more audio signals corresponding to the result are output to the acoustic model learning unit 8, but the recognition result (“noise”) of the recognition unit 4 before receiving the timer operation completion signal. When the switch 5 gives an ON operation signal different from the above, the input voice storage control unit 7 determines that a misrecognition has occurred, changes the current recognition result from “noise” to “light”, and changes the “ The recognition result “Akari” and the corresponding speech signal are output to the acoustic model learning unit 8. At this time, the acoustic model learning unit 8 re-learns the “AKARI” acoustic model by using the voice signal misrecognized as “NOISE”, so that when the user next utters “AKARI”, this utterance Can be recognized as “Akari”, and the recognition accuracy can be improved.

  For example, when the noise is input to the microphone 1 in a state where the lighting device B is turned off (off), and the feature amount extraction unit 2 extracts the feature amount of the noise and outputs it to the recognition unit 4, the recognition unit 4 When the feature amount input from the feature amount extraction unit 2 is compared with the acoustic model stored in the acoustic model storage unit 3 and it is determined that the feature is similar to the acoustic model of “AKARI”, the control unit 6 and the input voice A recognition result “Akari” is output to the storage control unit 7. When the recognition result “Akari” is given to the control unit 6, the control unit 6 outputs a trigger signal to the timer unit 9, and outputs a lighting control signal to the lighting device B. Turn on (ON). At this time, although the user does not speak “Akari”, the lighting device B is turned on. Therefore, it is determined that the noise is erroneously recognized as “Akari”, and the switch 5 is directly turned off to perform illumination. Device B is going to be turned off. When the switch 5 is turned off, an off operation signal of the switch 5 is given to the lighting device B, the lighting device B is turned off, and an off operation signal is given to the input sound storage control unit 7. In the input voice storage control unit 7, since an off operation signal different from the recognition result (“light”) of the recognition unit 4 is given from the switch 5 before receiving the timer operation completion signal from the timer unit 9, erroneous recognition is prevented. It is determined that it has occurred, the stored recognition result is changed from “light” to “noise”, and the changed recognition result “noise” and the feature amount of the speech signal are output to the acoustic model learning unit 8. . At this time, since the acoustic model learning unit 8 re-learns the “noise” acoustic model by using the “noise” voice signal erroneously recognized as “light”, the next time “noise” is input to the microphone 1. In this case, there is a high possibility that the voice signal can be correctly recognized as “noise”, and the recognition accuracy can be improved.

  As described above, when the lighting device B performs an operation different from the user's intention due to the misrecognition of the recognition unit 4, the user directly operates the switch 5 to correct the operation state of the lighting device B to a desired operation. Therefore, by detecting an operation to be corrected by the user using the switch 5, it is possible to correct the correspondence between the speech signal used for re-learning the acoustic model and the recognition result. Therefore, by re-learning the acoustic model using the correct recognition result, the probability that the recognition unit 4 correctly recognizes the next time the same sound is input is improved.

(Embodiment 2)
A second embodiment of the present invention will be described with reference to FIG. FIG. 3 is a block diagram of the control device with a speech recognition function according to the present embodiment. The control device A with a speech recognition function includes a microphone 1, a feature amount extraction unit 2, an acoustic model storage unit 3, and a recognition unit 4. The switch 5, the control unit 6, the input voice storage control unit 7, the acoustic model learning unit 8, the timer unit 9, and the control state storage unit 10 are provided as main components. In addition, since it is substantially the same as the control apparatus A with a speech recognition function demonstrated in Embodiment 1 except the point which added the control status memory | storage part 10, the same code | symbol is attached | subjected to a common component and the description is given. Omitted.

  The control state storage unit 10 receives a control signal supplied from the control unit 6 to the lighting device B and a control signal supplied from the switch 5 to the lighting device B. Among these control signals, the latest control signal is input. Based on the above, the current control state of the lighting device B is determined, and the determination result is stored. When the control state storage unit 10 receives an inquiry about the control state from the input voice storage control unit 7, the control state storage unit 10 stores the currently stored control state data ("lit" or "not lit" in the case of the lighting device B). Output to the input voice storage control unit 7.

  Here, the input voice storage control unit 7 stores the voice signal input to the feature amount extraction unit 2 and the recognition result by the recognition unit 4 in association with each other, and a new recognition result is input from the recognition unit 4. Then, the control state storage unit 10 is inquired about the current control state of the lighting device B. Then, when the current control state and the state indicated by the recognition result of the recognition unit 4 are the same state, an instruction to switch to the same state as the current control state is generally not issued. 7 determines that the voice signal is erroneously recognized, and corrects the stored recognition result. For example, when the recognition result “AKARI” is input to the input voice storage control unit 7 and the current control state captured from the control state storage unit 10 is “lit”, the input voice storage control unit 7 Since it can be determined that the command “light” is not input even though the lighting device B is already lit, the recognition result is corrected from “light” to “noise”, and the recognition result after correction And the corresponding audio signal is output to the acoustic model learning unit 8. Thus, the acoustic model learning unit 8 can re-learn the “noise” acoustic model using the noise signal misrecognized as “light”, and the accuracy of the “noise” acoustic model can be increased next time. The possibility of correctly recognizing “noise” when noise is input is improved. Therefore, even if the same noise signal is input while the lighting device B is turned off, there is a high possibility that it can be correctly recognized as “noise”. Can be prevented.

(Embodiment 3)
A third embodiment of the present invention will be described with reference to FIG. The control device with a voice recognition function of the present embodiment is provided with a human sensor 11 in the control device A with a voice recognition function of the first embodiment. Since the configuration other than the human sensor 11 is the same as that of the control device A with the voice recognition function described in the first embodiment, common components are denoted by the same reference numerals, and description thereof is omitted.

  The human sensor 11 is, for example, an ultrasonic sensor that detects an object in the detection area using ultrasonic waves, or a pyroelectric type that detects the presence or absence of a person in the detection area by detecting heat rays emitted from the human body. It consists of an infrared detection element, detects the presence / absence of a person within a detection area (including the sound collection range of the microphone 1) set within a predetermined distance range from the installation position of the microphone 1, and detects the detection result as an input voice storage control unit 7 is output.

  On the other hand, the input voice storage control unit 7 stores the voice signal input from the feature amount extraction unit 2 and the recognition result input from the recognition unit 4 in association with each other, and the recognition result is input from the recognition unit 4. Then, the recognition result and one or a plurality of audio signals corresponding to the recognition result are output to the acoustic model learning unit 8 to cause the acoustic model learning unit 8 to relearn the acoustic model. However, in the input voice storage control unit 7, when the human sensor 11 does not detect the presence of a person when the recognition result of the operation command is input from the recognition unit 4, the noise is regarded as the user's voice (command word). The input recognition result is corrected to “noise”, and the corrected recognition result and the corresponding speech signal are output to the acoustic model learning unit 8 by determining that the recognition has been mistaken. Thus, the acoustic model learning unit 8 can re-learn the “noise” acoustic model using a noise voice signal that is misrecognized as a command word (for example, “light” or “light-off”). Increasing the accuracy of the “noise” acoustic model improves the possibility of correctly recognizing “noise” the next time the same noise signal is input.

  In addition, although this embodiment is provided with the operation means by voice recognition as the operation means of the lighting device B and the means for direct operation using the switch 5, the detection result of the human sensor 11 is output to the lighting device B, The lighting load may be turned on only when a signal indicating that the human body is detected from the human sensor 11 in the lighting device B is input.

(Embodiment 4)
A fourth embodiment of the present invention will be described with reference to FIG. In addition, since the basic structure of the control apparatus A with a speech recognition function is the same as that of Embodiments 1-3, the same code | symbol is attached | subjected to a common component and the description is abbreviate | omitted.

  5 (a) and 5 (b) show, for example, a control device with a voice recognition function for voice input of a cooking recipe search device C that is installed in a kitchen and searches a recipe for cooking from the names of ingredients used for cooking and presents it to the user. The example of construction at the time of applying A is shown. The body 30 of the search device C is installed on a wall 41 around the sink of the kitchen 40. On the front surface of the body 30, a display panel 31 with a touch switch for displaying recipe search conditions and search results and a microphone 1 are provided. And are arranged.

  This search device C includes a touch panel type switch 5 for directly operating the search device C as an operation means, and performs a desired operation in accordance with a control signal input from the switch 5, for example, a user during cooking Assuming that the user's hand is dirty and the touch panel cannot be operated, a recognition unit 4 that recognizes a command word issued by a person to operate the search device C is provided, and control is performed according to the recognition result of the recognition unit 4 A desired operation is performed in accordance with a control signal output from the unit 6. In addition, when used for searching for cooking recipes, the acoustic model stored in the acoustic model storage unit 3 corresponds to the vocabulary “material search” which is a command word for transitioning the screen of the display panel 31 to the search screen. The "material search" acoustic model, the material name acoustic model corresponding to the material name generated when inputting the material name, for example, the "apple" acoustic model corresponding to the vocabulary "apple", and the command word for executing the search process Corresponding to "command" acoustic model consisting of acoustic models of many command words such as "search" acoustic model corresponding to the vocabulary "search" and sounds such as sounds other than voice for operating the device and sounds "Noise" acoustic model to be saved.

  By the way, in Embodiment 1 described above, when a recognition result is input from the recognition unit 4 to the control unit 6, the control unit 6 outputs a trigger signal to the timer unit 9 to start a time limit operation. Then, when the recognition result that the input sound is a voice is input from the recognition unit 4 to the control unit 6, the control unit 6 outputs a control signal to the control target device (search device C) according to the recognition result. Then, a trigger signal is output to a second timer (not shown) to start a time limit operation for a fixed time. The second timer outputs a timer completion signal to the input sound storage control unit 7 when a certain time is reached. When the trigger signal is input again from the control unit 6 during the time limit operation, the second timer performs the timed operation for a certain time. It is performed again from the beginning and has a so-called retriggerable function.

  On the other hand, the input voice storage control unit 7 stores the voice signal input from the feature amount extraction unit 2 and the recognition result input from the recognition unit 4 in association with each other, and the recognition result is input from the recognition unit 4. Acoustic model learning of all the recognition results except the recognition result input this time, the recognition result input before that, and one or a plurality of speech data corresponding to them among the stored recognition results The data is output to the unit 8, and the acoustic model learning unit 8 is relearned with the acoustic model. However, the recognition result input two times before is a vocabulary representing a voice operation such as “material search”, and then the input before the timer completion signal of the second timer is input. If the recognition result is “noise” and a vocabulary representing a voice operation such as “apple” is received as the current recognition result before the timer completion signal of the second timer is input after that, The voice storage control unit 7 deletes the previous recognition result that resulted in “noise” and the data of the voice signal corresponding to this recognition result.

  When the control device with voice recognition function A is applied to a device that continuously inputs voice, such as the cooking recipe search device C as in the present embodiment, when the user uses the device A, for example, “ Speak “material search” to switch to the material search screen, utter the material name such as “apple”, and then continuously utter “search” to execute the search operation. Therefore, there is a high possibility that the sound input before the second timer finishes the time limit operation is not noise but speech. That is, there is a high possibility that the sound input to the microphone 1 is a voice until the time limit operation of the second timer ends (that is, until a predetermined time elapses after the recognition result of voice is obtained). Therefore, if a recognition result “noise” is input from the recognition unit 4 during this period, it can be determined that the sound recognized as “noise” is actually a voice, and the input voice storage control unit 7 Then, by recognizing the recognition result “noise” and the audio signal data recognized as “noise”, the acoustic model learning unit 8 is prevented from re-learning the acoustic model using the erroneous recognition result. be able to.

1 is a block diagram of Embodiment 1. FIG. It is explanatory drawing of the usage example which applied the same to the illuminating device. 6 is a block diagram of Embodiment 2. FIG. It is a block diagram of Embodiment 3. (A) (b) is explanatory drawing of the usage example which applied Embodiment 4 to the cooking recipe search device. It is a block diagram of a prior art example.

Explanation of symbols

A control device with voice recognition function B lighting device 2 feature extraction unit 3 acoustic model storage unit 4 recognition unit 5 switch 6 control unit 7 input voice storage control unit 8 acoustic model learning unit

Claims (6)

A sound conversion unit that receives a sound that is either a voice or noise generated by a person to operate the control target device, converts the input sound into a sound signal that is an electrical signal, and outputs the sound signal;
A feature amount extraction unit that extracts a feature amount of an input sound from the audio signal;
An acoustic model unit that stores an acoustic model in which a feature amount is modeled for each of a plurality of voices and noises;
A recognition unit that recognizes an input sound by comparing the feature amount extracted by the feature amount extraction unit with an acoustic model stored in the acoustic model unit;
A control unit that outputs a control signal for performing an operation according to a recognition result of the recognition unit to a control target device;
An operation unit that directly outputs a control signal corresponding to the operation to the control target device; and
An input voice storage unit that stores the voice signal and the recognition result of the recognition unit in association with each other;
Acoustic model learning for re-learning the acoustic model of the input sound using the recognition result stored in the input speech storage unit and the speech signal corresponding to the recognition result, and updating the acoustic model stored in the acoustic model unit And
A timer unit for timing a predetermined time limit from the time when a recognition result is input from the recognition unit to the control unit;
When a control signal having a control content different from the recognition result of the recognition unit is output from the operation unit during the time limit operation of the timer unit, the input voice storage unit displays the stored recognition result as the operation unit. A control apparatus with a speech recognition function, wherein the acoustic model learning unit re-learns the acoustic model using the corrected recognition result and the input speech, based on the content of the control signal output from the control signal.   A control signal for turning on the control target device is output from the operation unit until the time limit operation of the timer unit ends after the recognition unit recognizes the input sound as noise while the control target device is off. Then, the input voice storage unit corrects and stores the recognition result recognized as noise into a voice for an on operation, and stores the voice recognition function-equipped control apparatus according to claim 1.   While the control target device is turned off, the control unit turns off the control target device from the operation unit until the time limit operation of the timer unit ends after the recognition unit recognizes the input sound as a sound for turning on the operation. 2. The control with a voice recognition function according to claim 1, wherein when the control signal to be output is output, the input voice storage unit corrects and stores the recognition result recognized as the voice for the on-operation to noise. apparatus.   The recognition unit includes a control state storage unit that stores an operation state of the control target device, and the input sound is a voice for operating the control target device to a current operation state stored in the control state storage unit 2. The control device with a voice recognition function according to claim 1, wherein the input voice storage unit corrects the recognition result recognized as a voice for operation to noise and stores the recognition result.   A human sensor for detecting the presence or absence of a person in a detection area including at least the sound collection range of the sound conversion unit is provided, and the recognition unit receives an input sound in a state where the human sensor does not detect the presence of a person. 2. The control with a voice recognition function according to claim 1, wherein when the voice for operation is recognized, the input voice storage unit corrects and stores the recognition result recognized as the voice for operation as noise. apparatus.   When the recognition unit determines that the new input sound is noise until a certain time has elapsed after the recognition unit recognizes the input sound as speech, the input speech storage unit recognizes that the input sound is recognized as noise. 2. The control apparatus with a voice recognition function according to claim 1, wherein the result and the voice signal data corresponding to the recognition result are deleted.

Images