JP3394998B2 - Noise removal device for voice input system - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise eliminating device for a voice input system.

[0002]

2. Description of the Related Art Conventionally, 2 for voice input and 2 for reference input
Noise is reduced by using two microphones and subtracting the feature amount of the sound (mainly noise) input through the reference input microphone from the feature amount of the sound (noise enters with the voice) input through the voice input microphone at the same time. There has been proposed a device for removing the.

Further, as shown in FIG. 10, there is known a noise eliminator including an adaptive filter 100 and a voice section detector 101 (Japanese Patent Laid-Open No. 2-278298 or Proceedings of the Acoustical Society of Japan, Heisei 2). September 1-8-5). This noise elimination device obtains power spectra X (f) and N (f) of voices respectively input from two microphones 104 and 105 for voice input and reference input, and from these, unknown input voice spectrum S ( f) is estimated, and at that time, a correction coefficient k which is a coefficient for correcting the gain difference between the two inputs by the adaptive filter 100.
(F) is generated and the value obtained by multiplying this by N (f) is X.
Noise is removed by subtracting from (f). The above correction coefficient k (f) is not fixed, and will be updated according to the level of both feature values that change sequentially.

By the way, if the voice input system is a personal computer, a workstation, etc., FIG.
As shown in FIG. 1, the voice recognition board 102 is inserted into the expansion bus of the computer, and the software is supplied from the floppy disk. In this case, the two microphones 104 and 105 have cords 106 and 107, respectively.
At, it is pulled out of the housing of the personal computer.
Then, the voice input microphone 104 is placed in front of the user, and the reference input microphone 105 is placed at a position where noise is equally input to the microphone 104.

[0005]

However, it is difficult to dispose the reference input microphone 105 at a position where noise is evenly distributed with the microphone 104, which is a very troublesome task for the user. On the other hand, the reference input microphone 105 may be fixedly attached to the back surface of the computer device to save the user from having to arrange the microphone 105 by himself. However, in this case, the noise of the heat dissipation fan motor is transmitted to the microphone 105. Excessive entry will reduce the accuracy of voice recognition. That is, when the reference microphone 105 is provided in the vicinity of the heat dissipation fan motor in the conventional noise eliminator, the sound of the heat dissipation fan motor excessively enters the microphone 105 and hardly enters the microphone 104. There is a significant imbalance in the level of the noise input, and even if the above-mentioned gain adjustment is performed, excessive data is subtracted from the power spectrum of the voice, which may make the feature pattern of the voice inaccurate.

Further, since the heat radiation fan motor is always a noise source while the personal computer is operating, the sound intensity is constant and the characteristic amount is also constant. It is a source of noise,
In contrast, for example, telephone or facsimile machine suddenly sounded, because emits the sound of a certain frequency in a constant cycle, be said sudden-routine noise source.

[0007] The present invention has been made in view of the above circumstances, in steady-routine noise source, in particular, the two input quantities a significant imbalance is such a case to the microphone, this is within the audio microphone It is an object of the present invention to provide a noise elimination device for a voice input system, which can surely eliminate such noise when the voice recognition is adversely affected when input.

[0008]

In order to solve the above-mentioned conventional problems, a noise elimination device for a voice input system according to the present invention includes a voice input means and a voice feature extraction section for extracting a feature amount of input voice. A voice recognition dictionary in which a plurality of voice feature amounts are registered in advance, a reference sound input unit different from voice input, a reference sound feature extraction unit for extracting the feature amount of the input reference sound, A correction coefficient generation unit that compares and calculates two feature amounts to generate a correction coefficient, a correction unit that corrects the voice feature amount in consideration of the correction coefficient and the reference sound feature amount, and a corrected voice feature amount. In a noise elimination device of a voice input system including a recognition unit for recognizing a voice by performing similarity calculation with the feature amount of each voice in the voice recognition dictionary, an average amount of the two feature amounts is obtained, The reference sound feature is based on the difference between the above two averages. Is characterized by having a Ku front correction unit for inputting the reference sound feature value of the corrected in the correction coefficient generation unit after performing correction.

Further, in the noise eliminating device for a voice input system according to the present invention, there is provided a means for storing the difference amount generated by the pre-stage correction section.

Further, in the noise eliminating device of the voice input system according to the first or second aspect, the reference sound inputting means is provided on a circuit board constituting the noise eliminating device.

[0011]

[0012]

[0013]

According to the above construction, even if the noise excessively enters the reference sound inputting means and hardly enters the voice inputting means, and there is a significant difference in the level of noise input between the two means, the imbalance will not occur. The corrected reference sound feature amount is canceled by the correction unit and is input to the correction coefficient generation unit. Therefore, the generation of the correction coefficient in the correction coefficient generation unit is
The gain adjustment is for general noise imbalance, and the error is small, which can contribute to the accuracy of speech recognition in the recognition unit.

Further, for example, when the pre-stage correction unit is activated at the time of setting up a personal computer, the difference amount generated when the pre-stage correction unit is activated is stored, and the difference amount is used for voice recognition. Based on the above, the corrected reference sound feature amount can be input to the correction coefficient generation unit. That is, it is possible to memorize the stationary / typical noise in advance and perform the correction process.

Further, since the reference sound inputting means is provided on the circuit board which constitutes the noise removing device, the user can save the labor of arranging the reference sound inputting means. In this case, even if there is a heat dissipation fan motor near the circuit board, as described above, the significant imbalance in noise input is resolved, and there is no problem.

[0016]

[0017]

【Example】

(Embodiment 1) The present invention will be described below with reference to the drawings illustrating the embodiment. FIG. 1 is a block diagram showing a noise elimination device of a voice input system. In the figure, 1 is a voice input microphone, 2 is a reference sound input microphone, 3,
4 is a feature quantity extraction unit, 5 and 6 are changeover switches, 7 and 8 are accumulation registers, 9 and 10 are averaging registers, and 11 and 12.
Is a subtracter, 13 is an adaptive filter (correction coefficient generation unit), 1
Reference numeral 4 is a subtractor, 15 is a voice section detection unit, and 16 is a recognition unit. This noise eliminating device is incorporated in a voice recognition board 20 as shown in FIG.

The voice input microphone 1 is similar to that shown in FIG.
As shown in FIG. 4, the microphone 2 for inputting the reference sound is fixed to the fastener of the voice recognition board 20 while being pulled out to the outside of the housing of the personal computer 21 through the cord 1a, or at the microphone position. It is detachably installed so that it can be changed. Microphone 2
When the board 20 is mounted on the personal computer 21, it is located on the back side of the personal computer 21. A heat radiation fan motor 22 is installed on the back side of the personal computer 21.

The feature extracting units 3 and 4 are for extracting features of voice based on, for example, a well-known frequency spectrum method, and include a plurality of band pass filters or A / D converters that pass different frequencies. It is configured with.

The changeover switches 5 and 6 are used for the feature extraction section 3 and 3.
The feature amount extracted in 4 is switched between output to the terminal a side or terminal b side. The switching is performed, for example, by selecting the terminal b side during setup of the personal computer 21 and automatically switching to the terminal a side after a while, or selecting the terminal b side each time the personal computer 21 is powered on. Then, it is conceivable that after a while, the control unit (not shown) operates so as to select the terminal a side. Note that terminal b
During connection to the side, it is desirable that noise is not emitted so that only the sound of the heat radiation fan motor 22 is input.

The accumulation registers 7 and 8 are provided in the feature extraction units 3 and 4, respectively.
The level of each frequency component is accumulated with respect to the feature amount extracted in (3), and the averaging registers 9 and 10 average the accumulated level of each frequency component. The average amount of the feature amount is obtained by the accumulating registers 7 and 8 and the averaging registers 9 and 10. The averaging registers 9 and 10 can hold the average amount while power is supplied thereto.

The subtractor 11 subtracts the average amount of the voice feature amount from the average amount of the reference sound feature amount to generate a difference amount F (f) between the two average amounts. That is, the heat radiation fan motor 22 input from the reference microphone 2
Is generated and a relatively small amount of the sound of the heat radiation fan motor 22 input from the voice microphone 1 is generated.

The subtractor 12 subtracts the difference amount generated by the subtractor 11 from the reference sound feature amount generated by the feature extraction unit 4. As a result, even if there is a significant difference in the level of noise input between the microphones 1 and 2, the difference is eliminated. That is, before the reference sound feature amount is input to the adaptive filter 13, the reference sound feature amount is corrected according to the level difference. It is possible to entrust such correction to the adaptive filter 13, but the adaptive filter 13 updates the correction coefficient for gain adjustment according to the level of both feature amounts that change sequentially, and the heat radiation fan motor 11 Since it is not necessary to update the stationary / typical noise as described above, such stationary / typical noise can be removed in advance before the adaptive filter 13 to improve the accuracy of the correction coefficient. .

As described above, the adaptive filter 13 updates the correction coefficient for gain adjustment in accordance with the levels of the two feature quantities that change sequentially. Correction coefficient k (f)
Is calculated by the formula (X (f) + c) / (N (f) + c). When the input is small, k (f) has a large error, and the noise component cannot be properly estimated. Therefore, the constant c> 0.
Has been introduced to prevent this.

The subtractor 14 generates a feature quantity S (f), and S (f) = X (f) -k (f) .N (f).
It is calculated by the formula.

The voice section detector 15 detects a voice section and outputs the discrimination signal to the adaptive filter 13. The adaptive filter 13 receives the above determination signal and generates a correction coefficient in the non-voice section.
More precisely, even when there is no voice component in a certain band and a voice component in another band, it is possible to update the estimated value of the noise component in the band without the voice component, more accurately, in the voice section. We are trying to improve the noise removal performance against the noise of.

The recognition unit 16 includes a voice recognition dictionary in which a plurality of voice feature amounts are registered in advance, and the degree of similarity between the corrected voice feature amount and the feature amount of each voice in the voice recognition dictionary. This is to perform speech recognition by performing calculation.

According to the above configuration, even if noise is excessively input to the reference sound input microphone 2 and hardly enters the voice input microphone 1, even if there is a significant difference in the noise input level between the two microphones, it is not possible. As for the equilibrium, the accumulation registers 7 and 8 and the averaging register 9 which constitute the pre-stage correction unit,
10 and the subtractors 11 and 12, and the canceled reference sound feature amount is input to the adaptive filter 13 that is the correction coefficient generation unit. Therefore, the generation of the correction coefficient in the adaptive filter 13 becomes a correction coefficient for gain adjustment for general noise imbalance, and there are few errors, and the accuracy of voice recognition in the recognition unit 16 can be increased.

Further, since the reference sound microphone 2 is provided on the voice recognition board 20, it is possible to save the user the trouble of disposing the microphone 2 himself. In this case, even if the heat radiation fan motor 22 is present near the board 20, there is no problem because the front stage correction unit eliminates the problem due to the significant imbalance of the noise input as described above.

It should be noted that it is also possible to memorize this in advance and perform a correction process for stationary / typical noise such as the heat radiation fan motor 22. For example, at the time of setting up a personal computer, the pre-stage correction unit is activated for a while, and the difference amount generated when the pre-stage correction unit is activated is stored in the non-volatile memory. The reference sound feature amount may be corrected on the basis of the predetermined difference amount, and the corrected reference sound feature amount may be input to the adaptive filter 13. If the nonvolatile memory is provided in this way, the stored contents are retained even when the power is cut off, so that the difference amount needs to be generated only once. Further, even if the pre-stage correction unit is configured by software, the difference amount is generated only once, so the load on the CPU is small.

In this embodiment, the reference sound microphone 2 is provided near the heat radiation fan motor 22, but the reference sound microphone 2 is placed near a cooler or other OA equipment. Even in the case where the noise is generated, it is possible to eliminate such stationary / typical noise.

Reference Example The noise removal device of the voice input system of the reference example is suddenly
The configuration is suitable for fixed noise. Figure 5
[ Fig. 6] is a block diagram of a noise elimination device of a voice input system of the present reference example . In the figure, 31 is a voice input microphone, 32 is an amplifier, 33 is a feature extraction unit composed of a 15-channel band-pass filter, 34 is a feature amount correction unit for correcting voice feature amount, 35 is a recognition unit, and 36 is a recognition unit. Is a telephone, 37 is a signal source detector, and 38 is a signal source dictionary.

The signal source dictionary 38 is preliminarily registered with feature quantities of sounds emitted from the telephone 36 which is a signal sound source or a facsimile device (not shown). Specifically, the telephone 36, the frequency components f ₁ ~f ₁₅ ring tones, noise factor k _n for each frequency, ring time t _1, and the non-ringing time t ₂ is stored. Noise factor k _n, as the level of each frequency of the ringing tone that made for correction, consistent with the frequency level at which the actual ringing tone is input through the microphone 31,
It is determined in consideration of the microphone position and the phone position.

The signal source detector 37 detects whether or not a ringing tone is emitted from the telephone 36, which is a signal tone generating source, and whether or not the generation of the ringing tone is stopped. The signal source detection unit 37 can detect the generation of a ringing tone by directly detecting a ringing signal from the telephone or by including a bandpass filter and a sound detector corresponding to the frequency component of the ringing tone. On the other hand, the detection of whether or not the ringing tone has stopped is performed, for example, by taking an off-hook signal from the telephone 36, or information on the ringing time t ₁ and the non-ringing time t ₂ stored in the signal source dictionary 38, It can be performed by the timer information and the judging means for judging based on these information.

The signal source detector 37 can select a signal source dictionary suitable for the signal sound from the signal source dictionary unit 38 when the signal sound is detected. For example, the signal source can be discriminated by the frequency of the signal sound, and the signal source dictionary can be selected based on the discrimination result. Further, a ringing tone is produced based on the selected noise coefficient of the signal source dictionary and H and L signals which will be described later, and the produced ringing tone is supplied to the correction section 34.

FIG. 6 is a timing chart showing an example of the relationship between the audio signal and the detection signal. A (t) indicates a voice (including a ringing tone that becomes noise) input from the voice input microphone 31. B (t) indicates a signal sound detection section. C (t) is the ringing time t ₁ ,
High, L created by the information of the non-ringing time t ₂
The ow signal is synchronized with the ringing section and the non-ringing section of the actual ringing tone.

FIG. 7 shows the characteristic amount Tv of the above A (t).
(T) is shown in frame units. On the other hand, FIG.
Shows the feature amount Tn (t) of the produced ring back tone in units of frames. Note that "0", "12", and "3" in the figure
Each value of 5 ”,“ 20 ”, and“ 6 ”is each frequency (channel)
Is a noise coefficient k _n for. Although the respective numerical values are completely the same as the corresponding numerical values in FIG. 7, this is for convenience of explanation, and in reality, there is some deviation.
Further, C in the figure shows the ringing section of the signal sound as H and the non-ringing section as L, and corresponds to the above-mentioned C (t). That is, the feature amount read from the signal source dictionary 38 is
It is shown that the data is output to the correction unit 34 only in the H section.

FIG. 9 shows the feature value Tx (t) of the audio signal after correction by the correction section 34 in units of frames. According to the figure, “12”, “35”, “20”, and “6” of each frequency, which is the characteristic amount of the ringing tone, are erased, which indicates that the correction is accurately performed.

According to the above construction, even for sudden and steady noise such as telephone ringing noise, it can be reliably removed to improve the accuracy of voice recognition and the voice dictionary is not affected by noise. Can be created. Further, as compared with the two-microphone structure, that is, the one including the feature extracting unit for the reference sound, the feature extracting unit for the reference sound is not necessary, so that the cost can be reduced. Of course, 2 shown in FIG.
It is also possible to provide a noise elimination device having a structure in which the above configuration is added to the structure of a microphone.

If the feature extracting section 4, the accumulating registers 7 and 8, the averaging registers 9 and 10 and the subtractor 11 shown in FIG. 1 are provided, the noise coefficient stored in the signal source dictionary 38 can be changed. Will be possible.

[0041]

As described above, according to the present invention, the accuracy of speech recognition can be improved even in the case where there is a significant imbalance in the input level to the two microphones, especially in the case of a stationary and fixed noise source. Can be increased. Thus, it is possible to arrange the reference microphones fixedly in the vicinity of the cooling fan motor and will, Ru can be eliminated user effort to determine the reference microphone location.

[Brief description of drawings]

FIG. 1 is a block diagram showing a noise removing device of a voice input system of the present invention.

FIG. 2 is a plan view showing a board on which the noise reduction device of the voice input system of the present invention is mounted.

FIG. 3 is a rear view of a personal computer that is a voice input system of the present invention.

FIG. 4 is a perspective view showing how the boat of FIG. 2 is loaded into a personal computer.

FIG. 5 is a block diagram showing a noise elimination device of a voice input system according to a reference example of the present invention.

FIG. 6 is a timing chart showing an example of a relationship between a voice signal and a detection signal in a reference example of the present invention.

FIG. 7 is an explanatory diagram showing, for each frame, the characteristic amount of a sound (both voice and noise are mixed) input from the voice input microphone in the reference example of the present invention.

FIG. 8 is an explanatory diagram showing, for each frame, a feature amount of a produced ring back tone in a reference example of the present invention.

FIG. 9 is an explanatory diagram showing the corrected feature amount in frame units in the reference example of the present invention.

FIG. 10 is a block diagram showing a noise reduction device of a conventional voice input system.

FIG. 11 is a plan view showing a board on which a noise reduction device of a conventional voice input system is mounted.

[Explanation of symbols]

1 Microphone for voice input 2 Microphone for reference sound input 3, 4 feature extraction unit 5, 6 selector switch 7,8 cumulative register 9, 10 averaging register 11,12 subtractor 13 Adaptive filter 14 Subtractor 15 Voice section detector 16 Recognition part 31 Microphone for voice input 33 Feature Extraction Unit 34 Feature correction unit 35 recognition unit 36 telephones 37 Signal Source Detector 38 Source dictionary

Continuation of the front page (56) References JP-A-58-62700 (JP, A) JP-A-3-269498 (JP, A) Patent 2836271 (JP, B2) Patent 3251980 (JP, B2) K.K. Krosghel, K .; Linh ard, Combined methods for adapt noise cancellation, SIGNAL PROCESSING IV: Theories and A applications, 1988, Vol. 1, p. 411-414 (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 15/20 G10L 21/02 JISST file (JOIS)

Claims (3)

(57) [Claims] 1. A voice input means, a voice feature extraction unit for extracting a feature amount of an input voice, a voice recognition dictionary in which a plurality of voice feature amounts are registered in advance, and a reference different from voice input. Sound input means, a reference sound feature extraction section for extracting the feature quantity of the input reference sound, a correction coefficient generation section for performing a comparison operation on the two feature quantities to generate a correction coefficient, the correction coefficient and the reference sound A correction unit that corrects the voice feature amount in consideration of the feature amount, and a voice recognition unit that performs similarity calculation between the corrected voice feature amount and the feature amount of each voice in the voice recognition dictionary. In a noise removal device of a voice input system comprising, the average amount of the two feature amounts is obtained, the reference sound feature amount is corrected based on the difference between the two average amounts, and then the corrected reference tone A pre-stage correction unit for inputting the feature amount to the correction coefficient generation unit is provided. Noise removal device of the speech input system, wherein the door. 2. The noise elimination device for a voice input system according to claim 1, further comprising means for storing the difference amount generated by the pre-stage correction unit. 3. The noise elimination device for a voice input system according to claim 1, wherein the reference sound input means is provided on a circuit board which constitutes the noise elimination device.