JPH09212192A - Speech recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain necessary speech recognition precision and power consumption reduction by switching the kind of a speech recognition process according to recognition precision that an application requires and then performing the speech recognition process at a clock frequency corresponding to the throughput required by the actually used speech recognition process. SOLUTION: Under specified recognition conditions, a process control part 16 is placed in operation to places only necessary parts of a speech recognition process part 15 selectively in operation at the clock frequency with the value of a source voltage needed to secure the specified speech recognition precision, thereby recognizing a speech inputted to a microphone 7. At the same time, a motor control part 10 is placed in operation according to this process result to perform operation corresponding to the contents of the speech.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition apparatus capable of significantly reducing power consumption required for voice recognition processing while ensuring necessary accuracy of voice recognition.

[0002]

2. Description of the Related Art A voice recognition apparatus for recognizing a voice includes a feature extraction unit for extracting a feature of a voice and a reference pattern (standard pattern) which is a feature pattern obtained by the feature extraction unit.
And a determination unit that determines which word is the voice that is the voice recognition target based on each similarity obtained by the similarity calculation unit. Often used as an element. In this case, the feature extraction unit uses a method of generating feature patterns and the like by analogically extracting feature information such as the size of each frequency component forming an audio signal by a bandpass filter or the like. Also, digitally, after the voice signal (analog signal) is converted into voice data (digital signal) by an A / D converter, the voice data is digitally processed by a DSP (digital signal processor), etc. Often the method of generating is used.
In addition, the similarity calculation unit and the determination unit perform a similarity calculation algorithm and a determination process for selecting the closest word by executing a similarity calculation algorithm by a microprocessor capable of performing various data processing. In many cases, either the method of performing the calculation or the method of calculating the degree of similarity by a logic circuit using dedicated hardware, the determination processing of selecting the closest word, or the like is used.

[0003]

However, among the devices using the above-mentioned conventional voice recognition device, especially the devices using voice recognition, those which emphasize portability, toys, etc. are assumed to be battery-driven. However, there is a problem in that the power consumption of the part related to the voice recognition is large, and if the part is constantly operated, the battery will be exhausted immediately. Therefore, as a method for solving such a problem, there have been heretofore known Japanese Patent Laid-Open Nos. 62-245296 and 3-20.
A technique disclosed in Japanese Patent No. 2899 and a technique disclosed in Japanese Patent Laid-Open No. 58-55991 have been proposed. Among these techniques, JP-A-62-245296 and JP-A-3-20
In the technology disclosed in Japanese Patent No. 2899, it is detected whether or not voice is input, and a predetermined period of time is exceeded,
When it is detected that no voice is input, the power supply to each part of the circuit is stopped to eliminate unnecessary power consumption. Further, in the technique disclosed in Japanese Patent Application Laid-Open No. 3-202899, when it is detected that no voice is input for a predetermined period of time or longer, the power supply to each circuit unit is stopped to waste power consumption. In addition to the above, by constantly operating the voice detection unit, when voice input is detected, power supply to the voice recognition unit such as the feature extraction unit is started, and the voice recognition process is restarted. When a voice is input, the voice recognition unit is immediately operated to recognize the voice while eliminating unnecessary power consumption.

However, these JP-A-62-245
The techniques disclosed in Japanese Patent Application Laid-Open No. 296, Heisei 3-202899 and Japanese Patent Application Laid-Open No. 58-55991 have the following problems. That is, JP-A-62-2
In the technology disclosed in Japanese Patent No. 245296 and Japanese Patent Application Laid-Open No. 3-202899, when the power is turned off without inputting voice for a certain time or longer, the power switch is operated each time the voice recognition operation is desired. However, there is a problem that the operation becomes complicated. Further, in the technique disclosed in Japanese Patent Laid-Open No. 58-55991, such a problem seems to have disappeared at first glance, but in reality,
Since the voice detector with low power consumption is configured by a very simple integrating circuit, there is a problem that when noise is input, the voice detector malfunctions due to this. Further, since it takes some time from the input of voice to the detection of voice input by the voice detection unit and the start of the voice recognition processing by the voice recognition unit, the voice recognition unit recognizes the voice. When the process started,
There is a problem in that the start end of the voice is lost and the voice recognition accuracy is significantly reduced. Therefore, an analog delay element such as a BBD (backed brigade type element) is inserted in the front side of the voice recognition unit, and after the voice detection unit detects the voice input, the voice signal bulletin is input to the voice recognition unit. Although it is possible to solve the above-mentioned problem by doing so, such a method can no longer be said to be a configuration aiming at low power consumption.

Further, when such a conventional technique is applied to an actual device, there are the following problems. In other words, when the conventional technology is used for applications such as toys, it is often desirable to use some method and operate by voice as much as possible even when the power of the voice recognition unit is turned off. Kaisho 62-24529
The technology disclosed in Japanese Laid-Open Patent Publication No. 6-62 and Japanese Laid-Open Patent Publication No. 3-202899 cannot handle this. Further, in the technique disclosed in Japanese Patent Laid-Open No. 58-55991 described above, when the voice recognition unit is operated with low power consumption, the accuracy is the same as when the voice recognition unit is operated with normal power consumption. The recognition accuracy cannot be secured. Also, when using conventional technology for applications such as actual equipment,
The same recognition result may not always be necessary.
For example, when answering a quiz answer (maru, bad, etc.) by voice recognition, a very high recognition rate is required, so
It is necessary to extract the recognition feature information and calculate the matching algorithm to withstand such a request. At other times, it is sometimes preferable to incorporate a contingency in the recognition result, and in such a case, relatively simple feature information and algorithms are sufficient. Furthermore, if only the amount of input voice is judged and it is judged to be "energetic", "depressed", etc., a smaller processing amount is required. As described above, the required accuracy of the recognition result varies depending on the device equipped with the voice recognition device, but a voice recognition device that dynamically switches the operation according to the required recognition accuracy has not been developed yet. Absent.

In view of the above situation, according to the present invention, in claim 1, the type of voice recognition processing is switched according to the recognition accuracy required by the application, and the processing amount required for the voice recognition processing actually used is thereby changed. It is an object of the present invention to provide a voice recognition device capable of performing necessary voice recognition accuracy and low power consumption by performing voice recognition processing at a clock frequency according to the above. According to the second aspect, the type of the voice recognition process is switched according to the recognition accuracy required by the application, and the clock frequency according to the processing amount required for the voice recognition process actually used,
It is an object of the present invention to provide a voice recognition device that can achieve required voice recognition accuracy and low power consumption by performing voice recognition processing with a power supply voltage. In claim 3, the type of the voice recognition processing is switched according to the recognition accuracy required by the application, whereby the voice recognition processing is performed with the clock frequency and the power supply voltage according to the processing amount required for the voice recognition processing actually used. As well as
An object of the present invention is to provide a voice recognition device capable of achieving required voice recognition accuracy and low power consumption by stopping power supply to a portion not involved in the recognition processing operation.

[0007]

In order to achieve the above object, a voice recognition apparatus according to the present invention comprises, in claim 1, a microphone for converting an input voice into a voice signal publication.
The first to Nth feature extraction units that extract the feature information of the audio signal publication output from this microphone by mutually different extraction methods, and the clock signal publication are supplied to these first to Nth feature extraction units. Then, a feature amount is obtained according to the clock control unit that operates at a specified speed and the accuracy of voice recognition, and the frequency of the clock signal publication output from the clock control unit is changed according to the feature amount. It is characterized by having a feature amount designation unit. Claim 2
Then, in the voice recognition device according to claim 1, a voltage control unit that supplies a power supply voltage having a designated voltage value to the first to Nth feature extraction units is provided, and the feature amount designation unit is provided. It is characterized in that the voltage value of the power supply voltage output from the voltage control unit is changed according to the obtained characteristic amount. Further, in claim 3, in the voice recognition device according to claim 1 or 2, the clock control unit and the voltage control unit are obtained by the feature amount designation unit of the first to Nth feature extraction units. Supply of the clock signal bulletin to the portion that does not participate in the extraction of the characteristic information, corresponding to the specified characteristic amount,
The feature is that the supply of the power supply voltage is selectively stopped.

With the above arrangement, in the voice recognition device according to the first aspect, the feature quantity necessary for voice recognition is obtained by the feature quantity designating section according to the accuracy of the voice recognition, and the clock is obtained according to the feature quantity. Voice data output from the microphone by changing the frequency of the clock signal publication output from the control unit and changing the operation speed of the first to Nth feature extraction units that extract feature information by mutually different extraction methods. By extracting the feature information of the voice recognition processing, the application performs the voice recognition processing at the operation speed according to the processing amount required for the voice recognition processing actually used, thereby reducing the necessary voice recognition accuracy and reducing the power consumption. And achieve. According to a second aspect, in the voice recognition device according to the first aspect, the power supplied from the voltage control section to the first to Nth feature extraction sections in accordance with the feature quantity obtained by the feature quantity designation section. In the application by changing the voltage value of the voltage,
The voice recognition process is performed at an operating speed and a power supply voltage according to the amount of processing required in the voice recognition process that is actually used, and the required voice recognition accuracy and low power consumption are achieved. Also,
According to a third aspect of the present invention, in the voice recognition device according to the first or second aspect, the clock control section and the voltage control section obtain the feature amount designation section of the first to Nth feature extraction sections. The amount of processing required for the voice recognition process actually used in the application by selectively stopping the supply of the clock signal bulletin and the supply of the power supply voltage to the portion that does not participate in the extraction of the characteristic information, corresponding to the characteristic amount. With the operation speed and the power supply voltage according to the above, only the part related to the extraction of the characteristic information is operated to perform the voice recognition process, and the required voice recognition accuracy and the low power consumption are achieved.

[0009]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will now be described in detail with reference to the embodiments shown in the drawings. FIG. 1 is a front view showing an example of a robot toy using an example of the form of the voice recognition device according to the present invention. A robot toy 1 shown in this figure has a body portion 2 formed in a cylindrical shape, two legs 3 movably attached to a lower portion of the body portion 2, and the body portion 2 described above.
Two arm parts 4 movably attached to the upper part of
A neck portion 5 fixed to the upper portion of the body portion 2, and the neck portion 5
A head 6 fixed on the top, a microphone 7 fixed to a side portion of the head 6, a battery 8 housed in the body 2, and a battery 8 housed in the body 2. A voice recognition device 9 for recognizing the voice input to the microphone 7 by the electric power obtained by the power source, and a control signal gazette (drive signal gazette) incorporated in the body 2 and output from the voice recognition device 9. And a plurality of motors 11 that are built in the body portion 2 and that generate a driving force according to the drive voltage output from the motor control portion 10, and the body. Of the inside of the part 2, it is built in at the base of each of the legs 3 and the arms 4,
A plurality of actuators 12 for moving the leg portion 3 and the arm portion 4 by the driving force obtained by the motors 11 are provided.

Then, when the user of the robot toy 1 utters any one of the preset words, the voice is recognized by voice while the power is supplied only to the portion necessary for the voice recognition processing. Perform the operation specified by this voice. The speech recognition apparatus 9 includes a speech recognition processing section 15 and a processing control section 16 as shown in FIG. 2, and operates the processing control section 16 under designated recognition conditions to perform designated speech recognition. Clock frequency required to ensure accuracy,
Based on the value of the power supply voltage, only the necessary portion of the voice recognition processing unit 15 is selectively operated to recognize the voice input to the microphone 7, and based on the result of this processing,
The motor control unit 10 is operated to perform an operation corresponding to the content of the voice. The voice recognition processing unit 15 includes a voice input unit 17, a voice section detection unit 18, a spectrum extraction unit 19, a volume detection unit 20, a similarity calculation unit 21, and
The voice spectrum dictionary 22 and the result output unit 23 are provided, and the voice signal bulletin output from the microphone 7 is digitized based on the clock signal bulletin supplied from the processing control unit 16 and the power supply voltage. After converting the data into data, if the voice recognition processing by the spectrum is specified while extracting the voice section information based on the voice data, the spectrum of the voice data is extracted based on the voice section information, and the voice spectrum is extracted. If the word indicated by the voice data is specified by the similarity calculation process using the dictionary 22, and if the voice recognition process by the voice segment information is designated, the voice segment based on the voice segment information of the voice segment is used. If the word indicated by the data is specified and the voice recognition processing based on the volume information is specified, the volume information of the voice data above. Based After identifying the words represented by the speech data, it generates a drive signal publication corresponding to a word that has been identified, and supplies it to the motor control unit 10.

The voice input section 17 includes an amplifier circuit for amplifying the voice signal publication output from the microphone 7, a filter circuit for filtering the voice signal publication output from the amplifier circuit to remove noises, and the like. It is equipped with an A / D conversion circuit for converting the audio signal publication output from the device into audio data, and takes in the audio signal publication output from the microphone 7, amplifies this, and filters it to remove noise. After the removal, it is A / D converted into voice data, which is supplied to the sound volume detection unit 20, the voice section detection unit 18, and the spectrum extraction unit 19. The sound volume detection unit 20 detects the sound volume of the sound data output from the sound input unit 17 and generates sound volume information indicating the sound volume of the sound data,
This is supplied to the voice section detection unit 18 and the result output unit 23.

The voice section detection unit 18 is the voice input unit 1 described above.
When a voice is input to the microphone 7 based on the output of 7 and the output of the volume detector 20, the voice detection information is generated by detecting the voice and the silent section in the voice data, for example, voice. And the boundary between voice and "stop"
Such as “t” between “st” and “p” is detected, and these detection results are used as silent section information, and the voice detection information and the silent section information are provided to the spectrum extraction unit 19 , And the result output unit 23. The spectrum extracting unit 19 divides the voice data output from the voice input unit 17 into voice segments for frequency analysis based on the voice detection information and the silence period information output from the silence period detecting unit 18, and analyzes the time spectrum. After extracting the characteristic information such as the pattern, the LPC cepstrum coefficient, the number of zero crosses, and the power value, the characteristic pattern indicating the characteristic of the audio data is generated based on the characteristic information, and the characteristic pattern is generated by the similarity calculation unit 21. And the result output unit 23.

Further, the speech spectrum dictionary 22 includes words to be recognized, for example, "Mezameyo", "Buster",
A standard pattern indicating characteristics such as “advance” and “kill” is stored. When a read instruction is issued from the similarity calculation unit 21, the standard pattern of the designated word is read and the standard pattern of the specified word is read. It is supplied to the calculation unit 21.
When the spectrum extraction unit 19 outputs a feature pattern indicating the feature of the voice data, the similarity calculation unit 21 outputs the feature pattern and the voice spectrum dictionary 22.
Pattern matching processing with each standard pattern stored in, to calculate the degree of similarity between the standard patterns stored in the If there is a standard pattern having a small distance, it is determined as a word corresponding to the voice data, and the determination result is supplied to the result output unit 23.

The result output unit 23, based on the voice recognition instruction output from the processing control unit 16, outputs the sound volume detection unit 2 described above.
0, the voice section detection unit 18, the spectrum extraction unit 19, the similarity calculation unit 21, and the voice spectrum dictionary 22 are set to the required state to perform the voice recognition process, and the voice recognition process is performed. The recognition result is taken in, the word corresponding to the voice data is determined, a drive signal bulletin is generated according to the determination result, and this is supplied to the motor control unit 10.

If the voice recognition using the spectrum extraction processing is designated by the voice recognition instruction output from the processing control unit 16, the volume detection unit 20, the voice section detection unit 18, the spectrum extraction unit 19, the similarity degree. Computing unit 21,
All of the voice spectrum dictionary 22 is put into an operating state, and it is determined whether the voice data is "awake,""buster,""advance," or "kill" as shown in FIG. 3, and the voice recognition instruction is given. If the voice recognition using the number of silent sections is specified by, the above-mentioned volume detection unit 2
0, the voice section detector 18, the spectrum extractor 19, the similarity calculator 21, and the voice spectrum dictionary 22, only the volume detector 20 and the voice section detector 18 are activated, and the voice data is “ It is determined whether it is a “ter” or a “pappappa”. Further, if the voice recognition using the sound volume is designated by the voice recognition instruction, among the sound volume detection unit 20, the voice section detection unit 18, the spectrum extraction unit 19, the similarity calculation unit 21, and the voice spectrum dictionary 22, Only the sound volume detection unit 20 is activated, and it is determined whether the sound data is “loud voice” or “loud voice”.

Further, the processing control unit 16 is provided with the robot toy 1.
Based on the voice recognition instructions specified by the user of
A feature amount designating section 24 that issues a designated voice recognition instruction and generates a clock frequency instruction signal bulletin and a voltage value instruction signal bulletin according to the voice recognition instruction, and a clock output from the feature amount designating section 24. A clock control unit 25 for generating a clock signal publication having a frequency according to the frequency indication signal publication, and a voltage control for producing a power supply voltage having a voltage value according to the voltage value indication signal publication output from the feature quantity designating unit 24. And a section 26. When voice recognition using spectrum extraction is instructed by the user of the robot toy 1, voice recognition instruction indicating that voice recognition processing is to be performed using spectrum extraction is issued to the voice recognition processing unit 15. , A clock signal bulletin having the highest frequency, a power supply voltage having the highest voltage value are generated and supplied to the voice recognition processing unit 15.

When the user of the robot toy 1 gives an instruction for voice recognition using a silent section, the voice recognition processing unit 15 is instructed to perform voice recognition processing using the silent section. While issuing a recognition instruction, a clock signal bulletin having a normal frequency and a power supply voltage having a normal voltage value are generated and supplied to the voice recognition processing unit 15. As a result, the power consumption of each element constituting the voice recognition device 9, for example, a CMOS circuit or the like, which consumes power when the clock signal publication is switched, is kept low, and the power used by each element ( (Current x voltage) is kept low, and power consumption is reduced. Further, when the user of the robot toy 1 gives an instruction for voice recognition using a volume section, the voice recognition processing unit 15 is given a voice recognition instruction indicating that the voice recognition processing is performed using the volume. With
The clock signal bulletin having the lowest frequency and the power supply voltage having the lowest voltage value are generated and supplied to the voice recognition processing unit 15. As a result, the power consumption of each element constituting the voice recognition device 9, for example, a CMOS circuit or the like, which consumes power when the clock signal publication is switched, is kept low, and the power used by each element ( (Current x voltage) is suppressed to a low level, further reducing power consumption.

As described above, in this embodiment, the processing control unit 16 is operated under the specified recognition condition, and the sound is generated at the clock frequency and the value of the power supply voltage required to secure the specified sound recognition accuracy. Only the necessary portion of the recognition processing unit 15 is selectively operated to recognize the voice input to the microphone 7, and the motor control unit 10 is operated based on the result of the processing to cause the content of the voice to be recognized. Since the operation corresponding to the above has been performed, the type of voice recognition processing is switched according to the recognition accuracy required by the application, and the clock frequency corresponding to the amount of processing required for the voice recognition processing actually used is thereby changed. , The voice recognition process is performed with the power supply voltage, and the power supply to the part that is not involved in the recognition process operation is stopped. It can be achieved with lower power consumption.

In the above-described embodiment, the clock frequency, the power supply voltage, and the circuit to which power is supplied are all switched depending on the type of voice recognition processing. Even if, for example, only the clock frequency is switched or only the clock frequency and the power supply voltage are switched, it is possible to obtain substantially the same effect as that of the above-described embodiment. Further, in the above-described embodiment, the voice recognition device 9 is attached to the robot toy 1. However, the voice recognition device 9 is attached to an electronic device that emphasizes portability, such as an electronic notebook and a handy terminal. Is also good. If the voice recognition device 9 is installed in such an electronic device, the voice recognition can be performed with the lowest power consumption required to ensure the required accuracy of the voice recognition, thereby minimizing the battery consumption. However, it can be operated for a long time.

[0020]

As described above, according to the present invention, in claim 1, the type of voice recognition processing is switched according to the recognition accuracy required by the application, and the voice recognition processing actually used is thereby changed. By performing the voice recognition processing at the clock frequency according to the required processing amount, it is possible to achieve the required voice recognition accuracy and low power consumption.
In claim 2, the type of the voice recognition processing is switched according to the recognition accuracy required by the application, whereby the voice recognition processing is performed with the clock frequency and the power supply voltage according to the processing amount required for the voice recognition processing actually used. The required voice recognition accuracy and low power consumption can be achieved. In claim 3, the type of the voice recognition processing is switched according to the recognition accuracy required by the application, whereby the voice recognition processing is performed with the clock frequency and the power supply voltage according to the processing amount required for the voice recognition processing actually used. It is possible to achieve the required voice recognition accuracy and low power consumption by stopping the power supply to the portion not involved in the recognition processing operation.

[Brief description of drawings]

FIG. 1 is a front view showing an example of a robot toy using an example of the form of a voice recognition device according to the present invention.

FIG. 2 is a block diagram showing a detailed circuit configuration example of the voice recognition device shown in FIG.

FIG. 3 is a table showing a processing example of the voice recognition device shown in FIG.

[Explanation of symbols]

1 ... Robot toy, 2 ... Body part, 3 ... Leg part, 4 ... Arm part,
5 ... neck, 6 ... head, 7 ... microphone, 8 ... battery, 9
... voice recognition device, 10 ... motor control unit, 11 ... motor,
12 ... Actuator, 15 ... Voice recognition processing section, 16 ...
Processing control unit, 17 ... Voice input unit, 18 ... Voice section detection unit (first to Nth feature extraction unit), 19 ... Spectrum extraction unit (first to Nth feature extraction unit), 20 ... Volume detection unit (first 1 to
Nth feature extraction unit), 21 ... Similarity calculation unit (first to Nth feature extraction unit), 22 ... Speech spectrum dictionary (first to Nth feature extraction unit), 23 ... Result output unit, 24 ... Feature amount Designated part,
25 ... Clock control unit, 26 ... Voltage control unit

Claims (3)

[Claims] 1. A microphone for converting input voice into a voice signal, first to Nth feature extraction units for extracting feature information of a voice signal output from the microphone by mutually different extraction methods, and A clock control unit that supplies a clock signal to the 1st to Nth feature extraction units and operates at a specified speed, and a feature amount is obtained according to the accuracy of voice recognition, and according to this feature amount, A voice recognition device, comprising: a feature amount designation unit that changes a frequency of a clock signal output from the clock control unit. 2. The voice recognition device according to claim 1, further comprising a voltage control unit that supplies a power supply voltage having a specified voltage value to the first to Nth feature extraction units, A voice recognition device characterized in that a voltage value of a power supply voltage output from the voltage control unit is changed in accordance with a characteristic amount obtained by a designation unit. 3. The voice recognition device according to claim 1, wherein the clock control unit and the voltage control unit are the first
To selectively stop the supply of the clock signal and the supply of the power supply voltage to the portion of the Nth feature extraction unit that corresponds to the feature amount obtained by the feature amount designation unit and is not involved in the extraction of the feature information. Characteristic voice recognition device.