JP2000089799A - Voice recognition system and method and recording medium stored with software for voice recognietion - Google Patents

Abstract

PROBLEM TO BE SOLVED: To correctly perform voice recognition with a small load even when there is noise in the surroundings of a voice recognition system. SOLUTION: A speaking state detecting circuit 45 detects the state that a user does not perform the inputting of vioces, that is, recognition operation is not performed. During this state, a filter arithmetic circuit 46 calculates filter constants based on inputs from two microphones 41, 42 to store them in a storage memory 47 together with corresponding correlation values. At the time of the recognition operation, a filter constant selecting circuit 49 selects a filter constant whose correlation value is the highest from the storage memory 47 and a noise cancelling circuit 410 eliminates the noise in a voice signal by using this filter constant. Since it is made unnecessary to perform complex operation and complex calculation for cancelling the noise simultaneously during actual recognition processing, the amount of calculations which should be performed in real time is reduced and then since it is made unnecessary to use circuits whose operating speeds are high speed especially, it is made easy to miniaturize a voice recognition system and to reduce the cost of the system.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an improvement in speech recognition technology for recognizing spoken words. An operation is performed.

[0002]

2. Description of the Related Art Voice recognition is known as a technique for inputting information to an electronic device such as a car audio system. In this speech recognition, data such as waveforms and parameters are prepared in advance as a recognition dictionary as features of each phrase to be recognized. Note that data for each individual phrase is referred to as voice reference data, and a portion storing a plurality of voice reference data is referred to as a recognition dictionary. In speech recognition, the spoken words are recognized by pattern matching of the user's voice input from the microphone with the recognition dictionary.

FIG. 9 shows a configuration example of a speech recognition system for performing such speech recognition. That is, this system includes a microphone (microphone) 11 and an input circuit 1
2, a recognition dictionary 13, and a pattern matching circuit 14
And a recognition result output circuit 15, which is used, for example, as a part for inputting a command to a car audio system or the like.

In this example, when the user speaks a phrase to be recognized to the microphone 11, the input circuit 12 converts the waveform of the audio signal input from the microphone 11 into digital data. Then, the pattern matching circuit 14
Compares the feature of the digital data with the voice reference data for each phrase stored in the recognition dictionary 13, and if there is a corresponding one, via the recognition result output circuit 15,
For example, the recognition result is output in the form of a phrase code or the like.

In an environment in which such a voice recognition system is actually used, the microphone 11 includes ambient noise in addition to the user's voice, and may be collected simultaneously with the originally intended voice. Many. In such a case, the S / N of the audio signal input to the pattern matching circuit 14 deteriorates due to the influence of ambient noise, and the pattern matching circuit 14 cannot sufficiently extract the characteristics of the audio data. May not be correctly collated with the voice reference data. As a result, there arises a problem that a correct recognition result cannot be obtained and the recognition rate is deteriorated.

For the purpose of solving such a problem, there is also known a conventional technique for improving recognition accuracy by removing noise from a speech signal to be recognized by adding adaptive noise cancellation processing to the speech recognition system. Have been.
Here, FIG. 10 is a block diagram showing a configuration example of such a conventional technique. This system uses an audio microphone 21
And a microphone 22 for noise, and an adaptive noise canceling circuit 24 in addition to the voice input circuit 23. In addition, this system
Pattern matching circuit 25 similar to that described in FIG.
, A recognition dictionary 26, a recognition result output circuit 27, and the like.

In this example, the voice microphone 21 is installed at a position where the voice spoken by the user can be easily recorded, while the noise microphone 22 captures the same noise as the noise source mixed into the voice microphone 21. In a location where Then, the input signals of these two microphones are input to the adaptive noise canceling circuit 24, and the adaptive noise canceling circuit 2
4 performs an adaptive filtering process based on these two inputs, thereby canceling (reducing) noise.
The input voice is output to the pattern matching circuit 25.

Here, as a general method of the adaptive filtering process, an LMS method or the like is used. That is,
The adaptive filtering process is performed by, for example, an adaptive noise canceller. In this adaptive noise canceller, a voice mixed with noise is input from one voice microphone, and a noise source signal that mainly causes noise is input from the other noise microphone. I do. Then, based on the noise source signal, by subtracting the noise estimated by the adaptive filter using the error signal or the like from the voice, a truly desired voice can be obtained.

The LMS (Least Mean Square) method uses a gradient vector for updating the filter constant of such an adaptive filter, that is, a gradient vector, to an expected value by an averaging operation that takes a long time from an error signal and an input signal. Is a method of updating the filter constant using the instantaneous estimated value for each sample instead of obtaining

In this case, the pattern matching circuit 25 compares and compares the features of the voice data after noise cancellation inputted from the adaptive noise cancellation circuit 24 with the voice reference data stored in the recognition dictionary 26. ,
If there is a corresponding one, the recognition result is output via the recognition result output circuit 27.

[0011]

However, the prior art using the above-described adaptive filtering processing has a problem that the load that must be processed simultaneously is large. In other words, in the case of operating an electronic device or the like in real time by voice recognition, in the related art, two complicated processes, an adaptive filtering process and a pattern matching process, must be performed simultaneously in parallel when voice is being uttered. did not become. For this reason, a circuit having a high operation speed is required, and there is a problem that the obstacle is large in terms of cost, size reduction, and the like.

The present invention has been proposed in order to solve the above-mentioned problems of the prior art, and has as its object to provide a voice recognition apparatus capable of performing a recognition operation correctly with a small load even in an environment where noise is present in the surroundings. To provide recognition technology.

[0013]

In order to achieve the above-mentioned object, a speech recognition system according to claim 1 comprises a first and a second speech recognition system.
Means for calculating a filter constant for filtering processing based on each audio signal from each microphone, and an audio signal from a second microphone with respect to an audio signal from the first microphone. A means for removing a noise component by performing a filtering process based on the characteristic based on the calculated filter constant, and a means for recognizing a spoken word based on a voice signal from which noise has been removed. Means for determining whether or not sound is being input from a microphone, and means for controlling so as to calculate the filter constant when the sound is not being input. The speech recognition method according to claim 3 is a method in which the invention according to claim 1 is viewed from the viewpoint of a method.
Calculating a filter constant for a filtering process based on each audio signal from each of the microphones; and an audio signal from a second microphone with respect to the audio signal from the first microphone. Removing a noise component by performing a filtering process based on a characteristic based on a filter constant, a step of recognizing a spoken word based on the voice signal from which the noise has been removed, and It is characterized by including a step of determining whether or not the input is performed, and a step of performing control so as to calculate the filter constant when the sound is not input. According to a fifth aspect of the present invention, the inventions of the first and third aspects are viewed from the viewpoint of a recording medium on which computer software is recorded. Speech recognition software for recognizing words in speech using a computer is provided. In the recorded recording medium, the software causes the computer to calculate a filter constant for a filtering process based on each audio signal from each of the first and second microphones, and to process the audio signal from the first microphone. The noise component is removed by performing a filtering process based on the audio signal from the second microphone and the characteristic based on the calculated filter constant, and based on the audio signal from which the noise has been removed, Recognize uttered words and phrases, determine whether voice is input from the microphone, When the serial audio is not input, and wherein the to calculate the filter constant. According to the first, third, and fifth aspects of the present invention, since the calculation of the filter constant for noise removal is performed while the actual speech input is not being performed, it is not necessary to perform the calculation in parallel with the actual recognition processing. Voice recognition can be performed correctly with a small load.

According to a second aspect of the present invention, there is provided a voice recognition system comprising: a first microphone for inputting a voice to be uttered; and a second microphone for inputting the same noise as the noise mixed in the first microphone. A switch for turning on at the time of inputting the voice, a recognition unit for recognizing a phrase from a given voice signal and outputting an end signal when the recognition is completed, and uttering based on the state of the switch and the end signal. Means for determining whether or not the filter constants are calculated; calculating means for adaptively calculating a filter constant so that the correlation value between the audio signals of the microphones is high; and calculating the calculated filter constant and calculating the filter constant. A memory for storing a plurality of combinations of the correlation values between the two voice signals at the time of the speech, and a combination of the correlation values at a predetermined interval only during the period when the speech is not being performed. And control means for operating the computing means,
Means for selecting a filter constant corresponding to the largest correlation value among the correlation values stored in the memory; and a filtering process based on the selected filter constant and two audio signals from each of the microphones. Means for outputting a sound signal from which noise has been cancelled, and wherein the recognizing means is configured to recognize a phrase based on the sound signal from which the noise has been cancelled. . According to a fourth aspect of the present invention, the second aspect of the present invention is viewed from the viewpoint of a method, wherein a first microphone for inputting a voice to be spoken and the first microphone are provided.
And a second microphone for inputting the same noise as the noise mixed in the microphone of the first embodiment. A recognition step of recognizing a phrase from a given voice signal, and outputting an end signal when the recognition is completed. Determining whether or not a speech is being made based on a state of a switch for turning on at the time of voice input and the end signal, and adapting so that a correlation value between each voice signal of each microphone becomes high. Calculating a filter constant; storing a plurality of combinations of the calculated filter constant and a correlation value between the two audio signals when the filter constant is calculated; Controlling the computer to execute the calculating step at a predetermined interval only for a time interval; A step of selecting a filter constant corresponding to the maximum correlation value, based the the selected the filter constant and two audio signals from the microphones,
Outputting a voice signal from which noise has been canceled by performing a filtering process, wherein the recognizing step recognizes a phrase based on the voice signal from which the noise has been canceled. According to a sixth aspect of the present invention, the inventions of the second and fourth aspects are viewed from the viewpoint of a recording medium on which software of a computer is recorded. Software for voice recognition for recognizing words and phrases in speech using a computer. On a recording medium recording the software, the software comprises: a first microphone for inputting an uttered voice to the computer;
A second microphone for inputting the same noise as the noise mixed into the first microphone; and a second microphone for recognizing a phrase from a given voice signal, and when the recognition is completed, outputting an end signal. Is determined based on the state of the switch for turning on at the time of input and the end signal, and the filter constant is adaptively adjusted so that the correlation value between the audio signals of the microphones is increased. A plurality of combinations of the calculated filter constant and the correlation value between the two audio signals at the time when the calculated filter constant is calculated are stored at a predetermined interval only while the speech is not being performed. The calculation of the filter constant is executed, and a filter constant corresponding to a maximum correlation value is selected from among the stored correlation values. By performing a filtering process based on two audio signals from a microphone, a noise-cancelled audio signal is output, and in the recognition, a phrase is recognized based on the noise-cancelled audio signal. It is characterized by.

According to the second, fourth and sixth aspects of the present invention, two microphones (microphones), that is, a first microphone for voice and a second microphone for noise are used, and the user does not input voice. A state in which the recognition operation is not performed is detected, and during that state, a correlation between the input waveforms between the two microphones is detected, and a constant of a filter provided in the second microphone is set between the two microphones. Calculate so that this correlation is high, stop setting if it is longer than a certain time,
Let the filter constant be determined. Such operations are executed at regular time intervals, and filter constants for each operation are separately stored. And when the recognition operation is actually performed,
After detecting the start of the user's voice recognition operation, from among the set values of the filter constants stored so far, the one having the highest correlation is selected, and a filter having a characteristic determined from the filter constant is selected. The difference between the audio signal from the first microphone and the audio signal from the second microphone is input to a speech recognition circuit to perform speech recognition processing.

As a result, since the signal subjected to the noise cancellation processing is subjected to the recognition pattern matching processing, the recognition operation can be performed without lowering the recognition rate even in a noise environment. In addition, the constant of the noise canceling filter is set in consideration of the vacant time in which the recognition process is not performed, so to speak. Therefore, there is no need to simultaneously perform complicated operations for noise cancellation and calculations during the actual recognition processing. As a result, the amount of computation that must be performed in real time can be reduced, and it is not necessary to use a specially high-speed circuit, which facilitates downsizing and cost reduction of the speech recognition system.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, an embodiment of a navigation system according to the present invention (hereinafter referred to as "embodiment") will be specifically described with reference to the drawings. In each of the drawings used in the following description, the same members and members of the same type as those in the drawings described earlier are denoted by the same reference numerals, and description thereof will be omitted.

This embodiment is realized using various hardware devices and a computer controlled by software. In this case, the software is created by combining the instructions according to the description in this specification, and the common parts with the above-described prior art use the technique described in the prior art. Further, the software includes not only program codes but also data prepared in advance for use in executing the program codes. The software realizes the effects of the present invention by utilizing a physical processing device such as a CPU and various chipsets incorporated in the navigation system.

However, specific hardware and software configurations for realizing the present invention can be variously modified. For example, depending on the circuit configuration and the processing capability of the CPU, a certain function may be realized by a physical electronic circuit such as an LSI, or may be realized by software. Also, regarding the parts that use software, various forms of software, such as compilers and assemblers, can be considered. A recording medium on which software for realizing the present invention is recorded is also an aspect of the present invention by itself.

As described above, there are various possible modes for realizing the present invention using a computer. In the following, a virtual circuit block for realizing individual functions included in the present invention and the embodiments will be described. The invention and embodiments will be described.

[1. Configuration] FIG. 1 is a functional block diagram showing the configuration of this embodiment. That is, in this embodiment, the audio microphone 41 and the noise reference microphone 42
A speech input circuit 43, an utterance start switch 44, an utterance state detection circuit 45, a filter operation circuit 46, a filter constant operation result storage memory 47, an operation control circuit 48, and a filter constant selection circuit 49. , A noise canceling circuit 410, a pattern matching circuit 411, a recognition dictionary 412, and a recognition result output circuit 413.

The voice microphone 41 is a first microphone for inputting voice to be uttered.
It is installed mainly at the position where the uttered voice of the user is recorded.
The noise reference microphone 42 is a second microphone for inputting the same noise as the noise mixed into the audio microphone 41, and the same noise as the noise source mixed into the audio microphone is recorded, and It is installed at a position where the voice level of the speaker can be input as low as possible.

The utterance start switch 44 is a switch for the user to turn on when the voice is input. The pattern matching circuit 411 recognizes a word from a given voice signal, and outputs an end signal when the recognition is completed. The utterance state detection circuit 45 is a means for determining whether or not voice is being input from the voice microphone 41. Specifically, based on the state of the utterance start switch 44 and the end signal, This is a means for determining whether or not the user is speaking.

A filter calculation circuit (also referred to as a noise cancellation filter calculation circuit) 46 is means for calculating a filter constant for filtering processing based on the input from the microphones 41 and 42, that is, the audio signal. More specifically, it is a calculating means for adaptively calculating a filter constant so that the correlation value between the respective inputs of the microphones 41 and 42 becomes higher.

The arithmetic operation control circuit 48 is means for controlling the filter arithmetic circuit 46 to calculate a filter constant for the adaptive filtering process when the voice is not input. Is control means for operating the arithmetic means at predetermined intervals only while the speech is not being made.

The storage memory 47 is a memory for storing a plurality of combinations of calculated filter constants and correlation values between two inputs when the filter constants are calculated. Further, the filter constant selection circuit 49 is means for selecting a filter constant corresponding to the largest correlation value from among the correlation values stored in the storage memory 47, and specifically, is stored in the storage memory 47. The correlation values for the past k times are compared, the one with the largest correlation value is extracted, and a filter constant corresponding to the largest correlation value is sent to the noise cancellation circuit 410.

The noise canceling circuit 410 performs a filtering process on the basis of the selected filter constant and the two inputs from the microphones 41 and 42, that is, the audio signal, to convert the audio signal from which the noise has been canceled. Output means, specifically, a voice microphone 4
This is a means for removing a noise component by performing a filtering process on the audio signal from No. 1 based on the audio signal from the noise reference microphone 42 and the characteristic based on the selected filter constant.

FIG. 2 is a conceptual diagram showing that noise is canceled by the operation of the filter 111 and the subtraction circuit 112 realized by the noise canceling circuit 410. Also, the pattern matching circuit 41
1 is configured to recognize an uttered word or phrase based on an output from which noise has been removed, that is, an audio signal.

The specific format and calculation method of the filter constant, the specific method of noise cancellation, etc.
A known technique described in the related art may be appropriately selected and used.

[2. Operation] In the present embodiment configured as described above, the calculation of the filter constant for noise removal is performed while no actual speech input is being performed, and is performed concurrently with the actual recognition processing. Since there is no need, it is possible to perform speech recognition correctly with a small load.

[2-1. Operation of utterance state detection circuit] That is, first, the utterance state detection circuit 45 determines whether or not the user is currently uttering. That is, the user turns on the utterance start switch 44 when uttering a word, and the pattern matching circuit 411 outputs a recognition operation end signal when the recognition of the uttered word ends. Therefore, as shown in the flowchart of FIG. 3, the utterance state detection circuit 45 waits for a signal (SWon signal) sent from the utterance start switch 44 when it is turned on (step 1), and receives the SWon signal. (Step 2) An utterance state signal indicating whether or not the utterance is being set is set to "Speaking" (Step 3).

The utterance state signal is supplied to the arithmetic operation control circuit 4
8, output to the filter constant selection circuit 49 and the pattern matching circuit 411, the utterance state detection circuit 45 continues,
The end of the recognition operation is awaited from the pattern matching circuit 411 (step 4). When the end of the recognition operation is recognized by receiving the end signal (step 5), the output utterance state signal is set to "non-speech". (Step 6) to notify the arithmetic operation control circuit 48 and the like that the speech is no longer being made.

[2-2. Operation of Arithmetic Operation Control Circuit] On the other hand, the arithmetic operation control circuit 48 sends a command to the filter operation circuit 46 based on the speech state signal input from the speech state detection circuit 45, so that every m time intervals n
During the time, the filter operation circuit 46 is controlled so as to execute the operation of the filter constant, that is, the filter operation process.

FIG. 4 is a flow chart showing the operation of the arithmetic operation control circuit 48. The processing of step 44 is performed for each of the counters for measuring two types of time, m time and n time, from non-speech to non-speech. At the beginning after the change during the utterance (step 42) and at the beginning after m or n hours have elapsed, the counter is reset and initialized. Otherwise, both counters are incremented. It represents processing.

That is, in this procedure, the arithmetic operation control circuit 48 first refers to the speech state signal (step 4).
1), the utterance state signal changes from "active" to "non-active"
(Step 42), an operation start command is sent to the filter constant calculation circuit 46 (step 43). And
While monitoring whether or not the utterance state signal changes from "non-speech" to "speech" (step 45), when n hours have passed (step 46), an operation end command is sent to the filter operation circuit 46 (step 47).

If the speech state signal has not yet changed from "speaking" to "non-speaking" (step 4).
8) Until m hours have passed (step 49), step 4
Repeat steps 4 to 48. However, in this repetition while waiting for the elapse of the m-time, since the operation start command (step 43) has not been issued to the filter operation circuit 46, the operation of the filter constant is not performed. After the elapse of m hours, an operation start command is sent to the filter operation circuit 46 again (step 43), and the operation of the filter constant is started.

FIG. 5 is a timing chart showing a state in which the filter operation circuit 46 operates during the non-speech state and for n hours at m time intervals as a result of the above control.

[2-3. Operation of Filter Operation Circuit] Next, FIG. 6 shows an operation procedure of the filter operation circuit 46 controlled by the operation operation control circuit 48 as described above.
That is, the filter operation circuit 46 outputs the speech sound for the recognition operation from when the operation control circuit 48 issues the operation start command (step 61) to when the operation end command is issued (step 63). While not present, two audio signals recorded by the audio microphone 41 and the noise reference microphone 42 are input, and the filter constant is adaptively adjusted so that the correlation value between the two audio signals becomes high. Calculation is performed (step 61).

When an operation end command issued from the filter operation control circuit 48 is received (step 6).
3) The operation of calculating the filter constant (filter operation) is stopped (step 64), and the correlation value between the filter constant determined by the calculation up to that point and the two audio signals based on the constant is calculated as follows: The data is output to and stored in the storage memory 47 (step 65).

While the utterance state signal indicates that the utterance state signal is not in the utterance state, the storage memory 47 stores a set of the filter constant passed from the filter operation circuit 46 and the corresponding correlation value as shown in FIG. Are stored a plurality of times up to k times in the past.

[2-4. Noise removal at the time of speech] The filter constant calculated and stored in this manner is used to cancel (remove) noise from a speech signal to be subjected to speech recognition. Here, FIG. 8 is a flowchart showing the operation procedure of the filter constant selection circuit 49 and the noise cancellation circuit 410 when removing noise from an audio signal. That is, first, the filter constant selecting circuit 49 refers to the past k correlation values stored in the storage memory 47 (step 81), compares the correlation values with each other (step 82), and sets Are detected (step 83).

The filter constant selecting circuit 49 stores the filter constant corresponding to the maximum correlation value in the storage memory 4.
7 (step 84), and sends this filter constant to the noise cancellation circuit 410 (step 85).

The noise canceling circuit 410 which has received the filter constant performs the following processing by inputting two signals recorded from the voice microphone 41 and the noise reference microphone 42. That is, as shown in FIG. 2, a signal obtained by passing the audio signal from the audio microphone 41 and the audio signal from the noise reference microphone through a filter having the characteristic of the filter constant transmitted from the filter constant selection circuit 49. And a subtraction process is performed using, and a signal obtained by the subtraction is output to the pattern matching circuit 411 as an audio signal after noise cancellation.

Then, the pattern matching circuit 411
Compares and compares the features of the thus-input speech signal after noise cancellation with the speech reference data stored in the recognition dictionary 412, and if there is a corresponding one, recognizes it via the recognition result output circuit 413. Output the result.

[3. Effect of Embodiment] As described above, in this embodiment, the two microphones, the voice microphone 41 and the noise reference microphone 42, are used, and the user does not perform voice input, that is, the recognition operation is not performed. A state is detected, and during that state, a correlation between the input waveforms between the two microphones is detected, and a constant of a filter provided in the noise reference microphone 42 is set so that the correlation becomes high between the two microphones. The setting is stopped after a certain period of time, and the filter constant is determined. Such operations are executed at regular time intervals, and filter constants for each operation are separately stored. Then, when the recognition operation is actually performed, after detecting the start of the user's voice recognition operation, the one having the highest correlation is selected from among the set values of the filter constants stored so far. Then, the difference between the audio signal from the noise microphone and the audio signal from the audio microphone through the filter having the characteristic determined by the filter constant is input to the audio recognition circuit, and the audio recognition processing is performed.

As a result, since the signal subjected to the noise cancellation processing is subjected to the recognition pattern matching processing, the recognition operation can be performed without reducing the recognition rate even in a noise environment. In addition, the constant of the noise canceling filter is set in consideration of the vacant time in which the recognition process is not performed, so to speak. Therefore, there is no need to simultaneously perform complicated operations for noise cancellation and calculations during the actual recognition processing. As a result, the amount of computation that must be performed in real time can be reduced, and it is not necessary to use a specially high-speed circuit, which facilitates downsizing and cost reduction of the speech recognition system.

[4. Other Embodiments] The present invention is not limited to the above-described embodiments, but includes other embodiments as exemplified below. For example, the speech recognition of the present invention can be used not only for inputting a command to a so-called car navigation system mounted on a car or the like, but also for other purposes such as inputting a command to another electronic device such as a computer. .

Further, the specific method of the filtering process, the specific format of the filter constant, the specific method of calculating the filter constant, and the like are free. The method of voice recognition using can be freely selected. Further, specific values such as the specific values of the time m and the time n shown in the above-described embodiment and the number of filter constants to be stored in the storage memory can be freely determined.

[0049]

As described above, according to the present invention, even in an environment where noise is present in the surroundings, it is possible to correctly perform speech recognition with a small load, so that the apparatus can be reduced in size and cost. It is easy to plan.

[Brief description of the drawings]

FIG. 1 is a functional block diagram showing a configuration of an embodiment of the present invention.

FIG. 2 is a conceptual diagram illustrating the principle of noise removal according to the embodiment of the present invention.

FIG. 3 is a flowchart showing an operation procedure of an utterance state detection circuit according to the embodiment of the present invention.

FIG. 4 is a flowchart showing an operation procedure of an arithmetic operation control circuit according to the embodiment of the present invention.

FIG. 5 is a timing chart showing operation timings of the filter operation circuit according to the embodiment of the present invention.

FIG. 6 is a flowchart illustrating an operation procedure of the filter operation circuit according to the embodiment of the present invention.

FIG. 7 is a conceptual diagram showing a combination of a filter constant and a correlation value according to the embodiment of the present invention.

FIG. 8 is a flowchart illustrating an operation procedure of a filter constant selection circuit and a noise cancellation circuit according to the embodiment of the present invention.

FIG. 9 is a functional block diagram showing a configuration example of a conventional technique.

FIG. 10 is a functional block diagram showing another configuration example of the related art.

[Explanation of symbols]

 41 Voice Microphone 42 Noise Reference Microphone 43 Voice Input Circuit 44 Utterance Start Switch 45 Voice Status Detection Circuit 46 Filter Operation Circuit 47 Storage Memory for Filter Result Calculation Results 48 Operation Control Circuit 49 Filter constant selection circuit 410 Noise cancellation circuit 411 Pattern matching circuit 412 Recognition dictionary 413 Recognition result output circuit 111 Filter 112 Subtraction circuit

Continuation of the front page (72) Inventor Koyo Shibasaki 5-35-2 Hakusan, Bunkyo-ku, Tokyo Clarion Co., Ltd. (72) Inventor Makoto Kisani 5-35-2 Hakusan, Bunkyo-ku, Tokyo Clarion Co., Ltd. F term (reference) 5D015 CC02 DD02 EE05 9A001 BB02 EE05 HH17 JJ77 KK54

Claims (6)

[Claims] 1. A first and second microphone, means for calculating a filter constant for filtering processing based on each audio signal from each microphone, and an audio signal from the first microphone Means for removing a noise component by performing a filtering process based on an audio signal from the second microphone and a characteristic based on the calculated filter constant, based on the audio signal from which the noise has been removed. Means for recognizing an uttered phrase, means for determining whether or not sound is input from a microphone, and means for controlling the filter constant to be calculated when the sound is not input. A speech recognition system, comprising: 2. A first microphone for inputting a voice to be spoken, a second microphone for inputting the same noise as noise mixed in the first microphone, and turned on when the voice is input. A recognition unit for recognizing a word from a given voice signal and outputting an end signal when the recognition is completed, and determining whether or not the speech is being made based on the state of the switch and the end signal. Means for calculating filter constants adaptively so that the correlation value between the audio signals of the microphones becomes higher; calculated filter constants; and two sounds when the filter constants are calculated. A memory for storing a plurality of combinations of correlation values between signals, and a control for operating the arithmetic means at predetermined intervals only while the speech is not being performed. Means for selecting a filter constant corresponding to the largest correlation value among the correlation values stored in the memory; and, based on the selected filter constant and two audio signals from each of the microphones, Means for outputting a noise-cancelled audio signal by performing a filtering process, wherein the recognition means is configured to recognize a phrase based on the noise-cancelled audio signal. Characteristic speech recognition system. 3. A step of calculating a filter constant for a filtering process based on each audio signal from each of the first and second microphones, and a second microphone for the audio signal from the first microphone. Removing a noise component by performing a filtering process based on the voice signal from the voice signal and a characteristic based on the calculated filter constant; and a phrase spoken based on the voice signal from which the noise has been removed. Recognizing, and determining whether or not voice is input from a microphone; and controlling the filter constant to be calculated when the voice is not input. Voice recognition method to be used. 4. A voice recognition method using a first microphone for inputting a voice to be uttered and a second microphone for inputting the same noise as noise mixed in the first microphone. A recognition step of recognizing a phrase from a given voice signal and outputting a termination signal when the recognition is completed; and a state of a switch for turning on when a voice is input and whether or not the speech is being made based on the termination signal. Determining; and adaptively calculating a filter constant so as to increase the correlation value between the audio signals of the microphones; and 2) calculating the calculated filter constant and calculating the filter constant. Storing a plurality of combinations of correlation values between two voice signals; and performing the calculation at predetermined intervals only while the speech is not being performed. Controlling to perform the following steps: selecting a filter constant corresponding to a maximum correlation value among the stored correlation values; and selecting two filter constants from the selected filter constant and each of the microphones. Outputting a noise-cancelled audio signal by performing a filtering process based on the audio signal, wherein the recognizing step recognizes a phrase based on the noise-cancelled audio signal. A speech recognition method characterized in that: 5. A recording medium recording voice recognition software for recognizing a word in a voice using a computer, wherein the software transmits the voice signal from the first and second microphones to the computer. Calculating a filter constant for the filtering process based on the audio signal from the first microphone and the audio signal from the second microphone and the characteristic based on the calculated filter constant. By performing a filtering process, a noise component is removed, based on the voice signal from which the noise has been removed, a spoken word is recognized, it is determined whether or not voice is input from a microphone, and the voice is input. When there is no voice recognition software, the filter constant is calculated. Recording medium on which 6. A recording medium storing voice recognition software for recognizing a phrase in a voice using a computer, the software comprising a first microphone for inputting voice to be spoken to the computer. And a second microphone for inputting the same noise as the noise mixed into the first microphone, and a word is recognized from a given voice signal, and when the recognition is completed, an end signal is output. Determining whether or not speech is being made based on the state of a switch for turning on at the time of voice input and the end signal, and adaptively filtering so as to increase the correlation value between the voice signals of the microphones. A set of a calculated filter constant and a correlation value between two audio signals when the calculated filter constant is calculated. A plurality of matchings are stored, the calculation of the filter constant is performed at predetermined intervals only while the speech is not being performed, and a filter constant corresponding to a maximum correlation value among the stored correlation values is selected. By performing a filtering process based on the selected filter constant and the two audio signals from each of the microphones, an audio signal in which noise is canceled is output. In the recognition, the noise is canceled. A recording medium storing voice recognition software for recognizing words and phrases based on a voice signal.