JP2011033717A - Noise suppression device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To process acoustic signals input from two microphones to generate signals suppressed in noise. <P>SOLUTION: A noise suppression device calculates a cross spectrum from the acoustic signals obtained with the two microphones; measure a degree of time variation of a phase component of the cross spectrum; makes a frequency component with small variation a voice component; regards a part of large time variation as a noise component to calculate a correction coefficient to suppress amplitude of the noise component; and uses the calculated correction coefficient to generate the signal suppressed in noise from the acoustic signal. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

The present invention relates to a noise suppression device for clearly detecting speech in a noisy environment.

Conventionally, there has been proposed a noise suppression device for accurately detecting a speech component by suppressing a noise component of an acoustic signal input from a microphone in a noisy environment.
Among them, a typical method is a spectral subtraction method (for example, Non-Patent Document 1, hereinafter SS method).
The SS method estimates the power spectrum of the target speech by measuring the power spectrum of noise in a section where there is no utterance, and subtracting the noise spectrum from the power spectrum at the time of speech. A technique has been proposed (for example, Patent Document 1).
Non-Patent Document 2 discloses an adaptive array technique for removing noise using a 2-channel microphone input.
This technique emphasizes the voice in the main path based on the sum of the signals of the two microphones, generates a reference signal that does not include the target voice in the sub path based on the difference signal, and extracts the noise component included in the main path from the sub path. The reference signal is deformed and subtracted to suppress noise.

JP-A-8-2221092

S. F. Boll, “Suppression of acoustic noise in speculation using subtraction”, IEEE Trans. Acoustic. SpeechSignal Process. , Vol. 27, no. 27, p. 113-120, 1979 Griffiths, L.M. J. et al. And Jim, C.I. W. "An alternative approach to linearly constrained adaptive beamforming", IEEE Transactions on Antennas and Propagation, 1982, vol. 30. , P. 27

However, when using the SS method, the environmental noise constantly fluctuates, and due to the error that occurs between the noise level measured during non-speech and the noise level included during utterance, the noise is reduced during spectral subtraction. When residual or excessive pulling occurs and a human hears it, an unpleasant noise called musical noise is generated.
On the other hand, Patent Document 1 takes measures against overdrawing by changing the coefficient of spectrum subtraction for the pause section in the voice. However, if the noise level increases, it becomes difficult to determine whether or not the voice is uttered, and correct noise measurement cannot be performed. Further, the SS method has a problem that it cannot cope with sudden noise.

In the adaptive array technique of Non-Patent Document 2, there is a problem that it is difficult to create a reference signal to be obtained as a difference signal between two microphones so that no audio signal is included. This is due to the fact that a blind spot can be formed only in one direction in order to prevent an audio signal from being included in a two-channel microphone array. In a general reverberant environment, there are components that are input after the audio signal is reflected by an object such as a wall, in addition to a path that is spatially input directly to the microphone.
Even if the direct wave with high power can be canceled, the reflected wave may not be cancelled.As a result, the target signal is subtracted when noise is subtracted from the main path due to the influence of the residual audio component contained in the reference signal. There is a problem that the adaptive processing works in the direction of removing the voice signal, and the quality of the processed signal is significantly impaired.
In addition, the adaptive array technology needs to know the timing signal (whether or not the voice is emitted) for operating the adaptive processing, and there are not a few errors in a general noise environment, so the adaptive processing proceeds in an unintended direction. There is also a problem that it ends up.

Therefore, in order to solve such a problem, an object of the present invention is to realize a noise suppression device capable of accurately suppressing only a noise component by a simple method using input signals from two microphones.

The present invention is a noise suppression device that processes acoustic signals acquired by two sound collectors and suppresses noise components, and evaluates the degree of temporal variation in phase difference between two acoustic signals for each frequency component. An amplitude correction coefficient is calculated so that the amplitude component of the frequency component becomes small with the frequency component having a large degree of time variation of the phase difference as a noise, and the noise component is calculated by applying the amplitude correction coefficient to the acoustic signal. Provided is a noise suppression device having an amplitude correction unit that outputs a suppressed signal.

According to a preferred aspect of the present invention, the phase difference fluctuation evaluating unit includes a cross spectrum calculating unit that calculates cross spectra of two acoustic signals at predetermined intervals, and a buffering unit that stores a predetermined number of calculated cross spectra. And a fluctuation measuring unit that calculates a time fluctuation degree of the phase component of the cross spectrum stored in the buffering part as a time fluctuation degree of the phase difference between the acoustic signals for each predetermined period.

In a preferred aspect of the present invention, the amplitude correction unit calculates a cross spectrum in which the noise component is suppressed by multiplying the amplitude component of the cross spectrum by the amplitude correction coefficient.

Further, as a preferred aspect of the present invention, the amplitude correction unit includes a whitening unit that performs whitening processing on the amplitude component of the cross spectrum, and multiplies the amplitude component of the whitened cross spectrum by an amplitude correction coefficient. To suppress the noise component.

Further, as a preferred aspect of the present invention, the amplitude correction unit includes a filter coefficient calculation unit that calculates a filter coefficient obtained by performing inverse Fourier transform on the amplitude correction coefficient, and performs filtering on one of the two acoustic signals or a synthesized acoustic signal. An acoustic signal in which noise is suppressed is generated by applying a coefficient.

When the noise suppression device of the present invention is applied to the utterance detection device, it is possible to configure an utterance detection device that does not easily react to voice or noise other than the target direction.
Furthermore, according to the sound reproducing apparatus to which the present invention is applied, it is possible to provide easy-to-hear sound from which the influence of stationary environmental noise is removed.

It is a block diagram when the noise suppression device according to the present invention is applied to the utterance detection device (first embodiment). It is an arrangement view of an utterance detection device for recognizing an ATM user in a financial institution as a speaker. 1 is a block diagram of a noise suppression device according to a first embodiment of the present invention. It is a block diagram when the noise suppression device according to the present invention is applied to an audio reproduction device (second embodiment). It is the example which calculated | required the cross spectrum of the acoustic signal input from two microphones. It is an example which calculated | required the spectrum of the area containing the audio | voice in an acoustic signal. It is an operation | movement flow for performing a noise suppression process. It is a block diagram of the noise suppression apparatus which concerns on the 2nd Embodiment of this invention.

Hereinafter, an embodiment to which a noise suppression device according to the present invention is applied will be described with reference to the drawings.
(First embodiment)
Here, an example will be described in which the noise suppression device according to the present invention is used in an utterance detection device that detects that an operator has a conversation using a mobile phone in front of a CD / ATM of a financial institution.
Proposed utterance detection device that issues a warning from a speaker or the like when it detects that a CD / ATM operator of a financial institution is uttering sound, in order to prevent the damage of wire fraud that has been increasing in recent years Has been.

In transfer fraud, criminals sometimes use victims to direct CD / ATM operations on their mobile phones and transfer victims' money into the perpetrator's account. In many cases, a person performs an operation while having a conversation with a partner at the telephone port using a mobile phone.
This utterance detection device analyzes the audio signals from the two microphones installed at the upper left and right ends of the CD / ATM in order to prevent wire fraud, and the voice signal uttered by the operator in front of the CD / ATM It is something to detect.
In such an environment in which CD / ATM is set, the operating noise of CD / ATM and the surrounding noise inside and outside the installation booth are large, and it is necessary to suppress the surrounding noise in order to detect the sound with high accuracy.

FIG. 2 is a diagram showing an example of the arrangement of the utterance detection device for detecting the utterance of the user 4 of the ATM 3 in the financial institution. In the utterance detection device 1, the main body device is installed on the wall surface, and two microphones 2 are installed at both left and right ends of the upper part of the ATM 3 with a predetermined distance therebetween. In the present embodiment, two microphones 2 are used, but the present invention is not limited to this, and an appropriate number of three or more may be used in an appropriate arrangement. What is necessary is just to perform the below-mentioned process with the pair of microphones.

Next, the configuration of the utterance detection device 1 to which the noise suppression device according to the present invention is applied will be described with reference to FIG. The utterance detection device 1 includes two microphones 2 that are sound collectors, an amplifier 10, an A / D converter 11, a noise suppression unit 12 that is a noise suppression device of the present invention, a cross-correlation calculation unit 24, and an utterance detection unit 25. It is configured.

Since it is desirable to collect sound from all directions, the microphone 2 is omnidirectional. The microphones 2 are installed with a predetermined distance (for example, 50 cm). This predetermined distance is determined to be a value that can specify that the operator in front of the ATM 3 uttered according to the sampling period, the assumed distance range with the speaker, and the like.
Note that this predetermined distance is a distance necessary for accurately detecting the voice direction, and does not impose any restrictions on the noise suppression device according to the present invention.
The microphone 2 has almost the same sensitivity and characteristics, but it is not necessary to prepare a specially high quality microphone.

The amplifier 10 is an amplifier that amplifies the acoustic signal collected by the microphone 2. This amplification factor is appropriately set according to the input voltage of the A / D converter 11.
The A / D converter 11 samples an acoustic signal, which is an amplified analog signal, at two sampling channels simultaneously at a predetermined sampling frequency and converts it into a discrete time signal (digital signal).
Since the amplifier 10 and the A / D converter 11 are both well-known components, detailed description is omitted.

The noise suppression unit 12 includes a phase difference variation evaluation unit 13 and an amplitude correction unit 14. In the noise suppression unit 12 in the first embodiment, the cross spectrum of the signals of the two channels input from the A / D converter 11 is obtained, noise suppression processing is performed on the frequency axis, and the cross spectrum in which noise is suppressed is obtained. Output.

The phase difference variation evaluation unit 13 measures whether each frequency band is dominant in the frequency axis on the frequency axis. Specifically, the analysis is performed at very fine time intervals, the cross spectrum of the two-channel signal is obtained, and the degree of time variation of the phase component (phase difference between the two acoustic signals) is measured.
In the amplitude correction unit 14, the phase difference variation evaluation unit 13 calculates a noise suppression coefficient such that the frequency component having a large variation in the phase difference between the two microphones is regarded as noise and the amplitude value of the frequency component becomes small. Then, noise suppression processing is performed by multiplying the chord suppression coefficient calculated for the original spectrum of the signal on the frequency axis. Detailed processing of the noise suppression processing will be described later.

The cross-correlation calculation unit 24 calculates a cross-correlation function by performing inverse Fourier transform on the cross spectrum of the two microphones 2 in which the noise component is suppressed every predetermined time, and outputs the cross-correlation function to the utterance detection unit 25.

The utterance detection unit 25 evaluates the peak height, peak width, and peak continuity of the cross-correlation value sequence calculated by the cross-correlation calculation unit 24, and determines whether there is a utterance from the designated direction.
In a voiceless sound frame, a disordered acoustic signal appears at the inputs of the left and right microphones 2, so that the cross-correlation value is relatively small, whereas in a voiced sound frame, for example, the operator 4 of the ATM 3 When speaking, since the sound from the front direction appears at the inputs of both microphones 2 in the same phase, the cross-correlation value becomes relatively large.
Therefore, in the utterance detection unit 25, the peak height that gives the maximum value of the cross-correlation value sequence is greater than or equal to a certain value, the width satisfies a certain value, and the peak position is close to a predetermined direction. When it is satisfied over a plurality of frames, it is determined that the voice is emitted.

The configuration when the noise suppression device according to the present invention is applied to the utterance detection device 1 has been described above.
Note that the noise suppression unit 12 which is a noise suppression device according to the present invention can be realized as a part of software constituting the utterance detection device 1. It can also be realized by a two-channel signal input function and a module that outputs a noise-suppressed signal.

Next, a specific noise suppression principle of the noise suppression unit 12 which is the noise suppression device of the present invention will be described in detail.
In general, when an audio signal is analyzed, such as when a person is talking, frequency analysis is performed using an analysis window of 20 ms to 40 ms with an analysis period (shift width) of 10 ms to 20 ms. This is based on the fact that the statistical property of the voice signal does not change for about 10 ms.
FIG. 6 is an example of a power spectrum when a voiced part of a female voice under a noisy environment is cut out and analyzed with a 30 ms Hamming window (the horizontal axis is frequency [Hz] and the vertical axis is intensity [dB]). .
The audio component of this signal has a sharp peak at 300 Hz, 600 Hz, and 900 Hz in the band of 1 kHz or less, and can be distinguished from the audio component, but 1 kHz or more is buried in noise, and which band is the audio component It is difficult to distinguish. In this way, it is difficult to determine which band has the dominant speech component or environmental noise by looking at only one frame analysis result.

The noise suppression unit 12 estimates the dominance of the speech component at each frequency with a time interval that is much finer than a normal analysis cycle, and will be described in detail below.
The peak at 600 Hz in FIG. 6 is a voice component. FIG. 5A shows the cross spectrum (complex number) of the left and right channels of this 600 Hz component, and the time transition for 10 ms is displayed.
Similarly, FIG. 5B shows a time transition for 10 ms of a cross spectrum of 1840 Hz which is a noise component.
FIG. 5 will be described. FIG. 5 shows the time variation of the cross spectrum at a specific frequency. The cross spectrum means a Fourier transform of the cross-correlation function of the left and right channels.
In FIG. 5, the change in the circumferential direction indicates the degree of temporal variation of the phase of the cross spectrum of the specific frequency, that is, the degree of variation of the relative phase difference between the acoustic signals input from the two channels. The change in the radial direction indicates the degree of time fluctuation of the amplitude of the cross spectrum of the specific frequency, that is, the degree of fluctuation of the product of the amplitude values of the two channels.

In FIG. 5, black circles are values in the analysis center frame, and the locus when the previous and next 5 frames are shifted by 1 ms and analyzed is indicated by a bold line.
From FIG. 5A, it can be seen that the cross spectrum component of 600 Hz, which is an audio component, hardly fluctuates in both phase and amplitude for 10 ms. On the other hand, FIG. 5B shows that the 1840 Hz component, which is a noise component, hardly changes in amplitude (product) of the left and right channels in 10 ms, but the phase difference fluctuates greatly.

Since the sound component does not change in characteristics for 10 ms, the directionality of the sound is strong, and the reverberation characteristics are constant in 10 ms, the fluctuation in the phase difference between the signals input from the two microphones is There are few ingredients. On the other hand, the noise component has low directivity, and signals from various sound sources randomly reach the left and right microphones, so that the variation in the phase difference between the signals input from the two microphones is large.
Therefore, the frequency component of interest is measured by measuring how much the phase of the cross spectrum, that is, the phase difference between the signals input from the two microphones fluctuates within a predetermined time (here, 10 ms). It is possible to determine whether or not the voice component is dominant or the noise component is dominant. The frequency band determined to be a noise component can be substantially suppressed by reducing its amplitude intensity.

The specific principle of noise suppression of the noise suppression unit 12 has been described above.
Here, an example has been described in which the variation in the phase difference between the acoustic signals input from the two microphones is measured from the cross spectrum calculation result using FFT. Thereby, there is an advantage that the amount of calculation can be reduced, but when there is no need to consider the amount of calculation, the phase component is directly calculated from the FFT calculation results of each of the acoustic signals input from the two microphones, A phase difference may be obtained from these differences, and the degree of variation thereof may be measured. In addition, the cross spectrum can be obtained by calculating a cross-correlation function in the time domain and performing Fourier transform on the cross-correlation function.
Moreover, in order to measure the fluctuation of the phase difference and discriminate between the speech component and the noise component, it is appropriate to estimate from the section of about 10 ms from the nature of the speech as described above, and the analysis cycle is also longer than the normal speech analysis. However, it is appropriate to carry out in a short time of about 1 ms. These may be appropriately determined depending on the balance between calculation amount and accuracy.

Next, a specific noise suppression processing procedure of the noise suppression unit 12 will be described. The noise suppression unit 12 includes a phase difference variation evaluation unit 13 and an amplitude correction unit 14 as shown in FIG. The phase difference fluctuation evaluation unit 13 further includes a preprocessing unit 15, a frame cutout unit 16, an FFT calculation unit 17, a cross spectrum calculation unit 18, a buffering unit 19, and a fluctuation measurement unit 20, and the amplitude correction unit 14 is white. And an amplitude correction coefficient calculation unit 22 and a suppression processing unit 23.
Hereinafter, the function of each unit and a specific processing procedure for noise suppression will be described with reference to the flowchart of FIG. 7 and the configuration diagram of FIG. 3 as appropriate.

First, in step S10, preprocessing is performed on the discrete signal converted by the A / D converter 11 at a predetermined sampling period (for example, 8 kHz). The preprocessing unit 15 is a frequency band unnecessary for processing the input discrete signal, for example, a low-frequency cut filter that cuts a frequency component of 70 Hz or less, and a high frequency band that compresses the dynamic range of the signal and increases numerical calculation accuracy. It consists of emphasis processing.
These are not essential processes. Moreover, although it is necessary to make it the same structure in both the left and right channels, the FIR (Finite Impulse Response) type, IIR (Infinite)
Impulse Response) There is no type restriction.

Next, in step S <b> 20, frame extraction processing is performed by the frame extraction unit 16 on the signal processed by the preprocessing unit 15. The frame cutout unit 16 cuts out a fixed-length frame (for example, 30 ms) from the acoustic signal with a predetermined shift width. Here, the shift width represents the analysis cycle, and is cut out in a very short cycle as compared with the normal speech analysis as described above. Here, the shift width is 1 ms. When a frame is cut out, it is cut out by multiplying the acoustic signal by using a Hamming window as a window function. The window function is not limited to the Hamming window, and a Hanning window or the like may be used.

Next, in step S30, the acoustic signal cut out by the frame cutout unit 16 is converted into the FFT (Fast) by the FFT calculation unit 17.
(Fourier Transform) calculations are performed and converted to frequency components. In the case of analysis conditions using 8 kHz sampling and a 30 ms analysis window, the number of signal points is 240, so 256 is adopted as the FFT size. A 240-point signal is input to the first 240 points, and 0 is input to the subsequent 16 points.

Next, in step S40, the cross spectrum calculation unit 18 calculates a cross spectrum from the FFT calculation result of the left and right channels by the following calculation formula (Formula 1).

Here, Y (k, t) is the cross spectrum at frequency number k and frame number t, X ₁ (k, t) is the left channel FFT result, and X ₂ (k, t) is the right channel FFT result. is there. * Represents a conjugate of a complex number.
Here, the cross spectrum is calculated after obtaining the Fourier transform of the left and right channel signals. However, the cross spectrum may be obtained by obtaining the cross correlation function of the left and right channels and then performing the Fourier transform.

The calculated cross spectrum is stored in the buffering unit 19. The buffering unit 19 is a ring buffer of a predetermined size, and when a new cross spectrum is input exceeding the predetermined size, the oldest one is erased sequentially.
The size of the ring buffer may be prepared as much as necessary for observing the temporal variation of the phase difference. As described above, here, the observation time of the time variation of the phase difference is 10 ms, the calculation of the cross spectrum is performed every 1 ms, and the cross spectrum for 5 ms before and after is sufficient, so the buffer size is 11.
In step S50, it is determined whether a predetermined period set in advance has elapsed. This predetermined period is a period for performing a noise suppression process described later.
The buffering unit 19 outputs the cross spectrum data accumulated in the subsequent fluctuation measurement unit 20 in order to perform noise suppression processing every time a predetermined period (here, 10 ms) elapses.

In step S60, the fluctuation measurement unit 20 uses the cross spectrum sequence temporarily stored in the buffering unit 19 for each predetermined period to determine which band is the voice component dominant and the noise component dominant. The degree of phase fluctuation is calculated.
In the utterance detection device 1, it is not necessary to shorten the utterance determination cycle so much. Here, the determination period is set to 10 ms, and the data in the buffering unit 19 is processed at a rate of once per 10 cross spectrum inputs to the buffering unit.

A specific method of calculating the degree of fluctuation of the phase component by the fluctuation measuring unit 20 will be described.
Now, dealing with the number of 2M + 1 of the cross spectrum (in this case, M = 5), the frame number of middle and _{t 0.} At this time, for example, the evaluation value of Formula 2 can be used as the degree of phase variation of the cross spectrum.

Here, D (k, t) is the phase error evaluation value at frequency number k and frame number t, θ (k, t) is the phase of cross spectrum Y (k, t) at frequency number k and frame number t. Information. As is apparent from Equation 2, this evaluation value represents an average value of the degree of phase variation from the immediately preceding frame of the cross spectrum within the range of 10 ms. The smaller the variation of the phase component of the cross spectrum, the smaller the D ( The value of k, t) is close to 0, and the value of D increases as the fluctuation of the phase component increases.

Up to the above, the phase difference fluctuation evaluating unit 13 performs the process of evaluating the fluctuation of the phase difference between the signals of the two channels.
In this example, the value of Equation 2 is used as the evaluation value of the degree of fluctuation of the phase difference. However, in addition to this evaluation value, the dispersion of phase information, the dispersion of the cross spectrum, etc. It can be used as an evaluation value.

Hereinafter, the noise suppression process will be described based on the evaluation value D (k, t) calculated in step S60 as the process of the amplitude correction unit 14.
First, in step 70, the amplitude correction coefficient calculation unit 22 calculates an amplitude correction coefficient using the evaluation value D (k, t) calculated by the variation measurement unit 20. Here, the function of Expression 3 is used as an example of the amplitude correction coefficient.

The function of Equation 3 outputs a value close to 1 with an input close to 0, and the output becomes smaller as the input has a larger absolute value. Here, γ is a parameter for controlling the inclination of correction, and the larger the value, the higher the suppression rate.
The amplitude correction coefficient calculation unit 22 calculates the amplitude correction coefficient at frame t ₀ and frequency k as f (D (k, t ₀ )).
That is, if the frame t ₀ and the frequency k are noise components, the value of D (k, t ₀ ) increases, so the value of f (D (k, t ₀ )) decreases, and if it is a speech component, D Since the value of (k, t ₀ ) decreases, the value of f (D (k, t ₀ )) increases, resulting in an amplitude correction coefficient that suppresses noise components.

Further, as an option, processing for smoothing the amplitude correction coefficient between frames may be added. For example, the following update is performed.

Here, A (k, t) is an amplitude correction coefficient at frequency number k and frame number t, and A '(k, t) is a smoothed amplitude correction coefficient at frequency number k and frame number t. β is a parameter that controls the smoothing window length and is a positive number close to 1 and not exceeding 1.

Next, in step S80, the suppression processing unit 23 performs noise suppression processing.
First, before performing noise suppression processing, whitening of the cross spectrum stored in the buffering unit 19 is performed. One cross spectrum is extracted from the buffering unit 19 at the above-described determination period (for example, 10 ms), and whitened (flattened) by the whitening unit 21. The extracted cross spectrum is the center cross spectrum among the 11 cross spectra stored in the buffering unit 19.
This whitening has an effect of making a cross-correlation function described later into a pulse shape. There are several possible variations of this process, but an example is shown below.
A power spectrum is calculated from the cross spectrum, and an autocorrelation function of the cross-correlation function is obtained by IFFT. LPC (Linear) from the appropriate low-order terms of these numeric sequences
Predictive Coding) coefficients are obtained and further converted into LPC cepstrum coefficients. Based on the spectrum envelope by this LPC cepstrum coefficient, flattening processing is performed on the frequency axis. This whitening is not an essential process.

Next, the suppression processing unit 23 multiplies the amplitude component of the cross spectrum that has been whitened by the whitening unit 21 on the frequency axis using the amplitude correction coefficient calculated by the amplitude correction coefficient calculation unit 22. And output a cross spectrum sequence in which noise is suppressed. As described above, the output of the cross spectrum series is output once in a predetermined determination period (here, 10 ms).

In the above example, the noise correction processing is performed by calculating the amplitude correction coefficient corresponding to the degree of fluctuation of the phase component of the cross spectrum, but the scope of the present invention is not limited to this. As another aspect, a frequency component having a phase component variation degree equal to or greater than a predetermined value may be determined as a noise component, and a correction coefficient that suppresses the frequency component determined to be the noise may be introduced.

The specific noise suppression processing by the noise suppression unit 12 which is the noise suppression device of the present invention has been described above. As described above, the cross spectrum subjected to the noise suppression process is subjected to inverse Fourier transform in the cross-correlation calculation unit 24 and returned to the cross-correlation function. The cross-correlation function subjected to the noise suppression process has a property that the correlation value at the time of non-voicing becomes significantly lower than that without the process. Therefore, in the utterance detection process by the utterance detection unit 25, it is easy to set a threshold value for determining the presence or absence of utterance, and the utterance detection is accurately performed even in an environment where the environmental noise is large.

(Second Embodiment)
Next, as a second embodiment, an example in which the noise suppression device according to the present invention is applied to an audio reproduction device will be described.
The sound reproducing device described in the present embodiment suppresses a noise component included in the acoustic signals input from the two microphones 2 by the noise suppressing device according to the present invention, and the clear acoustic signal in which the noise is suppressed is obtained from the speaker. It is a device that reproduces from.

Description of parts common to the utterance detection apparatus 1 according to the first embodiment will be omitted as appropriate, and the operation of the sound reproduction apparatus will be described below. Note that the flowchart of FIG. 7 noise suppression processing used in the description of the first embodiment is basically the same as that of the second embodiment.
FIG. 4 shows a block diagram of the audio reproduction device 5 according to the present embodiment.
Since the microphone 2, the amplifier 10, and the A / D converter 11 are the same as those in the first embodiment, description thereof is omitted.
The noise suppression unit 52 is the noise suppression device of the present application, and outputs a discrete signal in which noise is suppressed when an acoustic signal converted into a discrete signal by the A / D converter 11 is input. In the first embodiment, a cross spectrum in which noise is suppressed is output. However, the noise suppression unit 52 according to the second embodiment is different in that an acoustic signal in which noise is suppressed is output.

The D / A converter 58 converts the discrete signal whose noise is suppressed by the noise suppression unit 52 into an analog signal.
The analog signal converted by the D / A converter 58 is amplified by the amplifier 59 and reproduced by the speaker 60. Since the D / A converter 58, the amplifier 59, and the speaker 60 are well known, detailed description thereof will be omitted.

Next, a detailed block diagram of the noise suppression unit 52, which is the noise suppression device of the present invention, will be described with reference to FIG.
The phase difference fluctuation evaluation unit 13 calculates the degree of time fluctuation for each frequency component according to the fluctuation of the phase of the cross spectrum, and has the same function as that of the first embodiment.

The amplitude correction unit 54 includes a waveform synthesis unit 55, an amplitude correction coefficient calculation unit 56, and a suppression processing unit 57.

The waveform synthesizer 55 synthesizes the left and right channel audio signals into a single channel signal, and outputs a necessary waveform in conjunction with the processing of the amplitude correction coefficient calculator 56.
In the case of audio from a direction perpendicular to a line connecting two microphones, a simple sum may be used. In other cases, it is desirable to perform addition after performing phase shift based on the difference between arrival times of the left and right sound waves. Alternatively, either the left or right signal may be output.

In the amplitude correction coefficient calculation unit 56, the amplitude correction coefficient is calculated using the fluctuation evaluation value calculated by the fluctuation measurement unit 20, as in the first embodiment.
The amplitude correction coefficient calculation unit 56 further includes a filter coefficient calculation unit 561. The filter coefficient calculation unit 561 calculates FIR (Finite) from the calculated amplitude correction coefficient.
(Impulse Response) The coefficient of the filter is calculated. In the second embodiment, unlike the first embodiment, noise suppression processing is performed on the time axis using the calculated filter coefficient, and a time waveform subjected to noise suppression processing is output.
Here, in order to output high-quality sound, the update period of the amplitude correction coefficient calculation unit 56 and the filter coefficient calculation unit 561 is set to 1 ms. This corresponds to setting the predetermined cycle of step S50 to 1 ms in the flowchart of FIG.
That is, since the analysis period for calculating the cross spectrum is also 1 ms, the noise suppression processing in steps S60 to S80 is performed using the cross spectrum stored in the buffering 19 every time the cross spectrum is calculated. Means.
By shortening the update cycle, high-quality audio can be reproduced. Note that this update cycle may be extended if some quality degradation is allowed.

The processing in the filter coefficient calculation unit 561 will be further described.
In general, the calculated spectral correction characteristic has a complicated shape, so that the FIR filter coefficient length tends to be long. For this reason, in this embodiment, to obtain the filter coefficient, the filter coefficient is calculated using an FFT length that is longer than the FFT length for which the amplitude correction coefficient is calculated. Specifically, for example, upsampling is performed by quadrupling the 256-point amplitude characteristics, and 1024-point filter coefficients are obtained by performing inverse FFT on the up-sampling, and the work is performed using a window function having an appropriate length. I am doing so.
By convolving the calculated filter coefficient with the original acoustic signal on the time axis, even if the spectrum shape is complex and changes from moment to moment, it is possible to output high-quality sound corresponding to the change.

The suppression processing unit 57 performs a filtering process on the output of the waveform synthesis unit 55 by performing a convolution operation on the time axis using the filter coefficient calculated by the amplitude correction coefficient calculation unit 56.
The waveform to be processed is a signal from the waveform synthesizer corresponding to the central portion 1 ms of the time window (30 ms) used for calculating the central cross spectrum among the 11 cross spectra stored in the buffering unit 19. Then, a filtering process is performed on this signal to output a 1 ms (= 8 points) waveform.
In addition, although the process which performs a filter process on a time axis was demonstrated here, when high quality audio | voice reproduction | regeneration is not requested | required as mentioned above, the calculated amplitude correction coefficient is not returned on a time axis, but on a frequency axis. In this case, the FFT calculation result of the original acoustic signal may be multiplied to suppress the noise component.

The discrete signal subjected to the noise suppression processing is converted into an analog signal by the D / A converter 58 as described above, amplified by the amplifier 59, and reproduced from the speaker 60.

In the foregoing, the second embodiment in which the noise suppression device of the present invention is applied to an audio reproduction device has been described.
The noise suppression device of the present invention can be applied to other than the above embodiment. For example, if the noise suppression device according to the present invention is used in the previous stage of speech recognition processing, speech other than noise and the target direction can be accurately removed from the speech to be recognized, so that errors can be greatly reduced. become. In addition, since the noise component included in the recognition target speech is suppressed, the speech recognition rate can be greatly improved.

DESCRIPTION OF SYMBOLS 1 ... Main body 10 of an utterance detection apparatus ... Amplifier 11 ... A / D converter 12 ... Noise suppression part (noise suppression apparatus of this invention)
13 ... Phase difference variation evaluation unit 14 ... Amplitude correction unit 24 ... Cross correlation calculation unit 25 ... Speech estimation unit 58 ... D / A converter 59 ... Amplifier 60 ... Speaker 2 ... Microphone 3 ... ATM
4 ... Speaker 5 ... Audio playback device

Claims (5)

A noise suppression device that processes acoustic signals acquired by two sound collectors and suppresses noise components, wherein the phase difference variation evaluates the degree of temporal variation of the phase difference between the two acoustic signals for each frequency component. An evaluation unit;
A signal obtained by calculating an amplitude correction coefficient such that an amplitude component of the frequency component becomes small using a frequency component having a large degree of temporal variation of the phase difference as a noise, and suppressing the noise component by applying the amplitude correction coefficient to the acoustic signal. An amplitude correction unit that outputs
A noise suppression device comprising: The phase difference variation evaluation unit is
A cross spectrum calculation unit for calculating a cross spectrum of the two acoustic signals for each predetermined period;
A buffering unit for storing a predetermined number of the calculated cross spectrum;
A fluctuation measuring unit that calculates a time fluctuation degree of the phase component of the cross spectrum stored in the buffering part for each predetermined period as a time fluctuation degree of the phase difference;
The noise suppression apparatus according to claim 1, comprising: The noise suppression device according to claim 2, wherein the amplitude correction unit calculates a cross spectrum in which a noise component is suppressed by multiplying the amplitude component of the cross spectrum by the amplitude correction coefficient. The amplitude correction unit includes a whitening unit that performs a whitening process on the amplitude component of the cross spectrum,
The noise suppression apparatus according to claim 3, wherein a noise component is suppressed by multiplying the amplitude component of the whitened cross spectrum by the amplitude correction coefficient. The amplitude correction unit includes a filter coefficient calculation unit that calculates a filter coefficient obtained by performing inverse Fourier transform on the amplitude correction coefficient,
The noise suppression device according to claim 2, wherein the filter coefficient is applied to one of the two acoustic signals or a synthesized acoustic signal to generate an acoustic signal in which a noise component is suppressed.

Images