JP3269969B2 - Background noise canceller - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a background noise canceling device for canceling background noise superimposed on an audio signal, and is applicable to, for example, a telephone transmitter.

[0002]

2. Description of the Related Art In a use environment of a voice processing device such as a telephone, whether it is a stationary type or a portable type, background noise often degrades the quality of an input voice signal. Background noise elimination devices for eliminating background noise superimposed thereon from a signal have already been proposed.

References: Kenichi Taniguchi, Takashi Tsumura, Kimitoshi Fukudome, "Spectral Subtraction Method for Adaptively Estimating Noise," IEICE Technical Report, Technical Report of IECE SP9
4-116 (1995-03)] Conventionally, as this type of background noise canceller, there has been one disclosed in the above-mentioned document having a configuration shown in FIG.

In FIG. 2, background noise (hereinafter simply referred to as noise) is first input to an input terminal 100 of the background noise canceling device from a microphone, a voice storage / reproduction device (not shown), or the like. After that, a signal in which an audio signal and noise are mixed is input. An input signal input from the input terminal 100 is input to an analog / digital converter (hereinafter, A / D converter).
/ D converter) 101, after being converted into a digital signal, is further subjected to a fast Fourier transform operation unit (hereinafter referred to as a
In an FFT computing unit 102, division into frames, window function computation, and fast Fourier transform are performed. An adder 105 subtracts an estimated noise signal output from a noise estimator 106, which will be described in detail later, from a signal output from the FFT operator 102, and outputs the signal after the subtraction by an inverse fast Fourier transform operator ( In the following, an inverse FFT operation unit 107 performs an inverse fast Fourier transform for each frame of data, and a digital / analog converter (hereinafter referred to as a D / A converter) 108
The signal is converted into an analog signal and output to a speaker, a telephone line, or the like (not shown) via the output terminal 109.

Next, the configuration and operation of the noise estimation unit 106 will be described in detail. The noise estimation unit 106 includes a noise spectrum storage unit 103 and a noise spectrum update unit 104.

[0006] The noise spectrum updating section 104 includes an output signal Ni, k (where k
Represents the analysis frame order, i represents the frequency) and the output signal Ei, k from the adder 105 according to the equation (1). The noise spectrum updating unit 104 calculates the background noise according to the equation (2). Is estimated and the estimated noise signal Ni,
k + 1 is given to the noise spectrum storage unit 103 and stored. Also, the above-mentioned document includes a noise spectrum updating unit 104.
And an output signal Xi, k from the FFT converter 102.
And generating an estimated noise signal Ni, k + 1 in accordance with equation (3). Note that μ and μ ′ in Equations (2) and (3) are noise tracking coefficients that determine the speed of noise estimation.

Ei, k = Xi, k−Ni, k (1) Ni, k + 1 = Ni, k + μEi, k (2) Ni, k + 1 = Ni, k + μ′Xi, k Ei, k (3) The estimated noise signal Ni, k is updated from the initial value 0 so as to follow the input noise every time the analysis proceeds. The updated estimated noise signal (noise spectrum) is output to and stored in the noise spectrum storage unit 103, and is read out and provided to the noise spectrum update unit 104 at the timing of the next analysis frame to be used for updating the noise spectrum. At the same time, the signal is supplied to the adder 105, and is used for the elimination of noise in the output signal of the FFT calculator 102 by the subtraction processing shown in the equation (1).

Here, if the input signal to the conventional background noise canceller is only noise, after the noise estimation has converged,
Xi, k = Ni, k, and Ei, k = 0.

The noise estimating section 106 estimates a noise spectrum in a section before a speech signal is input. In addition, the tracking speed of the noise estimation is selected by selecting the noise tracking coefficient μ or μ ′ so that the noise estimation unit 106 hardly updates during the voice period shorter than the noise period.

After the noise estimation period, the speech signal Si,
The noise Ni, k is superimposed on k, and at this time, the conventional background noise canceller operates as follows.

When noise is superimposed on an audio signal,
The output signal Xi, k from the FFT calculator 102 is expressed by equation (4). Here, as described above, in the noise period,
If the estimated noise signal Ni, k follows the true noise, the output signal Ei, k from the adder 105 becomes as shown in equation (5), and the noise can be reduced. The operation after the inverse FFT calculator 107 is the same in the noise period and the input period of the audio signal.

Xi, k = Ni, k + Si, k (4) Ei, k = Xi, k-Ni, k = Si, k (5)

[0013]

However, the conventional background noise canceller has the following problems.

(1) Since the noise tracking coefficients μ and μ ′ do not immediately follow the background noise, the background noise (estimated noise signal) after the update processing has a tone property in the noise period,
In addition, since the time variation is small, background noise during the noise period becomes unnatural, and may be rather annoying.

(2) It is necessary to reduce the background noise estimation coefficient (noise tracking coefficient) in order to ensure sound quality in a voice section. When the estimation coefficient of the background noise is reduced, the average effect of the background noise estimation is large, but the estimation speed becomes slow. For this reason, the initial rise characteristic of the background noise canceller is poor.

(3) When the estimation coefficient of the background noise is reduced, the estimation speed becomes slow. Therefore, when the background noise level changes suddenly or when the background noise level changes immediately after the audio signal, the sound quality is significantly deteriorated.

[0017]

In each of the first to fifth aspects of the present invention, a time axis / frequency axis conversion means for converting an input signal on the time axis into a signal on the frequency axis, A background noise estimating means for estimating background noise in the input signal based on the input signal based on the input signal and forming the estimated noise signal, and estimating the estimated noise signal from the converted input signal on the frequency axis. The present invention relates to a background noise elimination device including a noise elimination operation means for reducing noise and a frequency axis / time axis conversion means for converting a signal on the frequency axis subjected to the noise elimination operation into a signal on the time axis.

[0018]

[0019]

According to a first aspect of the present invention, based on an input signal on a time axis or a frequency axis, whether an input signal is in a noise period including only background noise or an input signal is in a voice period including a voice signal And a tracking speed switching unit that switches a tracking speed of the output estimated noise signal to an input signal so as to be faster than a voice period in a noise period, inside the background noise estimating unit. It is characterized by having.

According to such a characteristic configuration, the update speed of the estimated noise signal is changed between the speech period and the noise period, so that the estimation can be performed at a high speed in the noise period and the update can be performed at a low speed in the speech period. Even during the period, if there is noise other than the audio frequency, the estimation value can be updated by following the noise, it is possible to prevent a large erroneous update of the noise estimation from deteriorating the sound quality, and the audio component does not overlap. Noise can be effectively removed.

Further, a second aspect of the present invention is characterized in that a spectrum attenuation processing means for attenuating an output signal from the noise cancellation calculation means and outputting the signal to the frequency axis / time axis conversion means is provided.

With the function of the spectrum attenuation processing means, it is possible to obtain a noise removing effect with excellent sound quality.

Still further, according to a third aspect of the present invention, based on an input signal on a time axis or a frequency axis, the input signal is in a noise period including only background noise, or the input signal is in a voice period including a voice signal. A noise detecting means for detecting whether or not the sound signal is present in a noise period, a distortionless attenuating means for attenuating an input signal on a time axis or a frequency axis without changing a sound quality in a noise period, and the distortionless attenuating means in a noise period. Output selection means for using an output signal from the device as an output signal of the device.

With such a characteristic configuration, it is possible to attenuate the noise inputted to the distortion-free attenuating means so as not to change the component ratio of the frequency components at all. The remaining noise can be attenuated with a natural sense of hearing, and the sound quality during the noise period can be improved.

[0026] The fourth present invention are those having the characteristic configuration of the second and third of the present invention is intended to act similarly to the second and third of the present invention. Further, the fourth present invention
Is based on the input signal on the time axis or frequency axis.
If the input signal is in a noise period that contains only background noise,
Audio detection to detect if the signal is in the audio period containing the audio signal
Output means, and the sound inside the background noise estimation means.
During the voice period, updating of the value of the output estimated noise signal is stopped.
It is also characterized in that an update stop unit for stopping the operation is provided. This
With such a configuration, the background noise estimating means is updated during periods other than the noise period.
Estimation that new operation can be stopped and incorrectly updated with the signal during the audio period
The sound quality is reduced without performing noise attenuation using the noise signal.
Deterioration can be prevented beforehand.

Furthermore, a fifth aspect of the present invention, first, those having a characteristic feature of the second and third of the present invention, first,
It works in the same way as the second and third inventions.

[0028]

BEST MODE FOR CARRYING OUT THE INVENTION

(A) First Embodiment Hereinafter, a first embodiment of a background noise canceling device according to the present invention will be described in detail with reference to the drawings.

(A-1) Configuration of the First Embodiment FIG. 1 is a block diagram showing the overall configuration of the background noise elimination device of the first embodiment.

Referring to FIG. 1, the background noise canceling device includes:
The input processing system includes an input terminal 1, an A / D converter 2, a window function calculator 3, an FFT calculator 4, and a voice detector 28.

The input terminal 1 is connected to a microphone, a voice recording / reproducing device or the like (not shown). The background noise canceling device receives the input analog signal via the input terminal 1. The input signal may be only background noise, or may be an audio signal on which background noise is superimposed.

The A / D converter 2 converts an input analog signal into a digital signal, and outputs the digital signal to the voice detector 28 and the window function calculator 3. The resolution and the sampling frequency may be general values for the audio signal.

The voice detector 28 detects the period of the voice signal on which the noise is superimposed, in other words, separates the voice period from the noise-only period, and supplies the voice detection signal to each of the switches 6, 9, 18 described later. Things. As the sound detector 28, a known sound detection method may be applied. For example, a method of monitoring a difference between power average values of a long-term (for example, 500 ms) signal and a short-term (for example, 40 ms) signal, a method of calculating a correlation coefficient, and a method of calculating an LPC coefficient and detecting a change thereof For example, any method may be used as long as it can correctly detect the audio signal.

The window function calculator 3 converts the digital signal from the A / D converter 2 into a predetermined frame length (for example, 32 m).
s) and performs a window function process in order to prevent high frequency components generated in this division process, and outputs the processed signal to the FFT calculator 4. Here, it is sufficient to perform the window function processing using a well-known window function such as a Hamming window, a Hamming window, and a Blackman window.

The FFT operation unit 4 performs an FFT operation to convert a signal on the time axis into a frequency spectrum component (a signal on the frequency axis), and a switch 6, a background noise estimating unit 7 and a This is output to the adder 5. Further, the FFT calculator 4 outputs the phase information obtained as a result of the FFT calculation to a phase holder 29 described later. Note that the FFT calculation method is a known calculation method, and a description thereof will be omitted.

The background noise elimination apparatus of the first embodiment includes an adder 5 and a background noise estimator 30 as a configuration for eliminating background noise. The background noise estimating unit 30 includes an on / off switch 6, a background noise updating unit 7, and a background noise holding unit 8.

The switch 6 is supplied with the output spectrum signal of the FFT calculator 4 as an input signal.
A sound detection signal from the sound detector 28 is given as a control signal. The switch 6 is closed when the voice detection signal indicates a period of only noise (hereinafter, referred to as a noise period), and the spectral signal Xf, k output from the FFT calculator 4 (where k is (Representing the analysis frame order, f represents the frequency) and passing it to the background noise updating unit 7. On the other hand, the switch 6 is opened when the voice detection signal indicates a voice period, and prevents passage of the output spectrum signal from the FFT calculator 4. That is, the spectrum signal Xf, k from the FFT calculator 4 is supplied to the background noise updating unit 7 only in the noise period.

The background noise updating unit 7 includes a background noise holding unit 8
, The estimated noise signal N'f, k-1 one analysis frame before is also given, and the background noise updating unit 7 forms the estimated noise signal N'f, k of the current analysis frame according to the equation (6). To the background noise holding unit 8 and the adder 5. Further, the background noise holding section 8 delays the given estimated noise signal N′f, k by one analysis frame and gives the delayed signal to the background noise updating section 7. Note that a in the equation (6) is 1>a> 0
And a constant for determining the speed of noise estimation. The initial value N'f, 0 of the estimated noise signal is set to 0.

N′f, k = a · N′f, k−1 + (1−a) · Xf, k (6) From the equation (6), the analysis of the background noise proceeds, and the estimated noise When the signal N'f, k-1 = Xf, k, N'f, k =
Xf, k. That is, when there is no audio signal, the spectrum signal Xf, k from the FFT operator 4
Is the background noise spectrum Nf, k. When the noise spectrum can be correctly estimated by repeating the background noise estimation algorithm shown in equation (6), the estimated noise signal N′f, k is Noise spectrum Nf, k (= Xf, k)
It turns out that it converges to.

The estimated noise signal N ′ from the background noise updating unit 7
f and k are output to the adder 5 and the band-specific spectrum attenuation processing unit 17 described later.

The estimated noise signal N'f, k from the background noise updating unit 7 is given to the adder 5 as a subtraction input.
Further, the spectrum signal Xf, k from the FFT calculator 4 is given as an input to be subtracted, and the adder 5 performs a subtraction process according to the equation (7) and is included in the spectrum signal Xf, k. The noise component is eliminated, and a residual component (elimination residual component) Ef, k is obtained and output to an on / off switch 9 described later. Since Nf, k = Xf, k in the noise period, the residual component (erasure remaining component) Ef, k can be expressed by the equation (8).

Ef, k = Xf, k−N′f, k (7) Ef, k = Nf, k−N′f, k (8) Here, the estimated noise signal from the background noise estimation unit 30 N'f,
k is appropriately updated in the noise period by the on / off control of the switch 6, but the value at the final timing of the noise period is maintained in the voice period. The adder 5 adds the spectral signal Xf, k from the FFT calculator 4 to the equation (7) regardless of whether or not the estimated noise signal N'f, k is updated (regardless of the period). Perform an erasure operation.

Further, the background noise elimination device of the first embodiment includes a band-specific spectrum attenuation processing unit 17 for attenuating the noise elimination spectrum signal Ef, k output from the adder 5 for each band during a voice period. ing. The band-specific spectrum attenuation processing unit 17 includes a band divider 10, a band-specific attenuation calculation unit 11, a power calculation unit 12, and a logarithmic calculation unit 1.
3, a signal-to-noise ratio calculator for each band (hereinafter, referred to as a band s / n calculator) 14, an attenuation calculator 15 for each band, and a smoothing calculator 16. In order to make the band-specific spectrum attenuation processing unit 17 function effectively only during the audio period, the input stage and the output stage of the band-specific spectrum attenuation processing unit 17 are provided with an on / off switch 9 and a two-input one-output switch 18, respectively. Have been.

In the first embodiment, an example is shown in which the quality of the audio signal is improved by dividing the output signal from the adder 5 into a plurality of bands. If the number is small, the attenuation process may be performed in one entire frequency band to be processed without dividing the frequency band.

The switch 9 is supplied with a voice detection signal from the voice detector 28 as a control signal. The switch 9 is opened during a noise period to pass the output spectrum signal Ef, k from the adder 5. , And is closed during the speech period, and the output spectrum signal Ef, k from the adder 5 is supplied to the band divider 10.

The band divider 10 divides the components of the input spectrum signal Ef, k into several bands (hereinafter referred to as channels). As a dividing method, for example, a band of 300 to 3400 Hz (spectral signal Ef,
k effective band) can be applied to a small frequency band having a band spread of 310 Hz into ten equal frequency bands. Also, the division method depends on the characteristics of the noise to be applied.
The size of the divided band may be changed as appropriate, and is not limited to equal division. The band-divided audio signal from the band divider 10 is provided to the band-dependent attenuation calculator 11 and the power calculator 12.

The band splitter 10 is also provided with the estimated noise signal N'f, k from the background noise updating unit 7, and the band splitter 10 similarly divides this signal into a plurality of bands. Then, the band division noise signal is also provided to the power calculation unit 12.

The power calculator 12 calculates the audio power Pch (m, k) for each channel according to the equation (9), and calculates the noise power Nch for each channel according to the equation (10).
(M, k) is calculated. The audio power Pch (m, m) for each channel obtained by the power calculator 12
k) and the noise power Nch (m, k) are calculated by the logarithmic calculation unit 13.
Given to. Here, m represents the channel number, and the sum in Equations (9) and (10) is the sum of all the frequency components f (fme to fms) between the lower limit frequency fme and the upper limit frequency fms of the m channel. It is about.

Pch (m, k) = Σ (Ef, k) (9) Nch (m, k) = Σ (N′f, k) (10) The logarithmic converter 13 calculates the audio power for each channel. Pch
(M, k) and the noise power Nch (m, k) are subjected to logarithmic conversion, and the audio power log (Pch (m, k)) and noise power l for each channel after logarithmic conversion.
og (Nch (m, k)) is output to the band s / n calculator 14. As the logarithmic conversion method here, a known method such as a method using a lookup table or a method using linear approximation can be applied.

In the first embodiment, the logarithm is calculated to improve the calculation accuracy in the band-specific s / n calculation unit 14, which will be described later. However, the accuracy of the S / N estimation is very important. When used in an environment other than the above, the logarithmic calculation unit 13 may be omitted.

The band s / n calculator 14 includes a logarithmic calculator 13
Power signal log (Nch
(M, k)) and log (Pch (m, k)), (1
The equation (1) is applied to calculate the signal-to-noise power ratio S / Nch (m, k) expressed in logarithm for each channel and to provide the calculated result to the attenuation calculation unit 11 for each band.

S / Nch (m, k) = log (Pch (m, k)) − log (Nch (m, k)) (11) The band-specific attenuation calculator 15 calculates the logarithmic expression for each channel. Based on the obtained signal-to-noise power ratio S / Nch (m, k), the amount of attenuation for the audio spectrum signal of each channel output from the band divider 10 is calculated and provided to the band-specific attenuation calculator 11. is there. In other words, it calculates the attenuation gain L (m, k) of how much the power Pch (m, k) of the audio spectrum signal of each channel output from the band divider 10 is attenuated. The method of determining the attenuation gain L (m, k) is determined, for example, by creating an empirical rule in advance as a reference table and using the logarithmic signal-to-noise power ratio S / Nch (m, k) as a key. The attenuation gain may be referred to from the following equation. Further, an equation such as equation (12) is formed by empirical rules and the like, and the logarithmic signal-to-noise power ratio S / Nch (m, k) is calculated. The attenuation gain may be determined by applying the equation. Note that β in Expression (12) is a constant.

L (m, k) = − 10/3 · S / Nch (m, k) + 20−β (12) Here, the determination method of the attenuation gain is to use the reference table, Low S / N despite applying a conversion formula like
(For example, equation (12) is inversely proportional to S / N). That is, a channel component having a large amount of noise component is largely attenuated, and a channel having a large number of audio components is not so much attenuated.

The attenuation calculation unit 11 for each band uses the attenuation gain L (m, k) from the attenuation calculation unit 15 for each band to obtain (13)
According to the formula, the frequency components of each channel from the band divider 10 are attenuated and output to the smoothing operation unit 16. Therefore, the power of each channel is attenuated as shown in equation (14). In this specification, underlined symbols represent vectors.

P′ch (m, k) = L (m, k) · Pch (m, k) (13) where P′ch (m, k) = ｛E′fms, k,. , E '
fme, k｝ Pch (m, k) = {Efms, k,..., Efme, k}.

P′ch (m, k) = L (m, k) · Pch (m, k) (14) The smoothing operation section 16 outputs the output signal P′ch (m) from the band-based attenuation operation section 11. , K) and its own output signal Och (m, k−1) one analysis frame before, as shown in equation (15), using a weighting constant γ (1>γ> 0). As a result, a signal Och (m, k) obtained by smoothing the output signal P′ch (m, k) from the band- dependent attenuation calculation unit 11 is formed, and one input terminal b of the two-input one-output switch 18 is formed.
To give.

Och (m, k) = γ · P′ch (m, k) + (1−γ) · Och (m, k−1) (15) The switch 18 outputs the sound from the sound detector 28. The detection signal is given as a control signal, and the switch 18 selects a signal from an undistorted attenuator 23, which will be described later, provided to the input terminal a during the noise period, and switches the inverse FFT calculator 1
9, and also selects a signal from the smoothing operation unit 16 of the band-specific spectrum attenuation processing unit 17 provided to the input terminal b during an audio period and supplies the signal to the inverse FFT operation unit 19.

The background noise canceling apparatus of the first embodiment is a distortion-free apparatus that forms a signal to be supplied to the inverse FFT operator 19 from the output signal of the FFT operator 4 and the output signal of the adder 5 during the noise period. A damping unit 23 is provided. The distortionless attenuator 23 includes a power monitor 24, a power attenuation calculator 2
5, an attenuation amount smoothing operation unit 26 and a power attenuation unit 27.

The power monitor 24 calculates the spectrum power PE of the output spectrum signal (estimated residual signal) Ef, k of the adder 5 according to the equation (16), and also calculates the spectrum power PE according to the equation (17). Is calculated and given to the power attenuation calculator 25. Note that the sum Σ in Equations (16) and (17) is performed for all frequency components.

PEk = ΣEf, k (16) PXk = ΣXf, k (17) In the first embodiment, as described above, the sum of all the frequencies to be processed is obtained. , Divided into several bands and the spectral powers PEk, PX
k may be calculated.

The power attenuation calculator 25 calculates the basic value LXk of the attenuation with respect to the output signal Xf, k of the FFT calculator 4 according to the equation (18), and outputs it to the attenuation smoother 26. It is.

LXk = PEk / PXk (18) The attenuation smoothing section 26 prevents the level of the processed sound from becoming discontinuous even if the attenuation LXk fluctuates steeply and discontinuously for each frame to be processed. In other words, a smoothing coefficient δ (1>δ> 0) is used for the attenuation LXk from the power attenuation calculator 25 in order to make the attenuation LXk continuous so that the sound quality is natural and good. The smoothing operation according to equation (19) is performed to obtain a smooth gain LXsm, k, which is output to the power attenuator 27.

LXsm, k = δ · (LXsm, k−1) + (1−δ) · LXk (19) Here, the smoothing coefficient δ is a coefficient for determining the smoothing speed of the smoothing operation. What is necessary is just to determine suitably according to the use environment and noise to be performed.

As shown in equation (20), the power attenuating unit 27 calculates the output signal Xf, k of the FFT arithmetic unit 4 by the attenuation amount LXsm, k after the smoothing process from the attenuation amount smoothing operation unit 26. The attenuated signal Xlf, k is attenuated by the switch 18 described above.
Is output to one of the input terminals a.

Xlf, k = LXsm, k · Xf, k (20) Here, the spectrum signal Xlf, k after the attenuation processing has the component ratio of the frequency component of the spectrum signal Xf, k before the attenuation processing at all. It is obtained as a spectral component that reproduces a natural tone with only the power attenuated without changing.

Further, the background noise canceller of the first embodiment includes an inverse FFT calculator 19, a window function overlap processor 20, a D / A converter 21, and an output terminal 22 as an output processing system. I have. Further, as described above, the phase holder 29 for transferring the phase information in the input processing system to the output processing system is provided.

The phase holder 29 holds the phase information obtained as a result of the FFT operation by the FFT operation unit 4, and performs the inverse FFT operation.
It is delivered to the arithmetic unit 19.

In addition to the phase information, the inverse FFT calculator 19 is supplied with the output spectrum signal from the power attenuator 27 during the noise period by the selection operation of the switch 18 and from the smoothing calculator 16 during the voice period. Are provided. The inverse FFT calculator 19 performs an inverse FFT process on the output spectrum signal of the smoothing calculator 16 or the power attenuator 27 so as to have the phase held in the phase holder 29, and converts the signal on the frequency axis. It is converted into a signal on the time axis and output to the window function overlap calculator 20.

The window function overlap calculating section 20 reduces the two or three window functions by, for example, 50% on the time axis in order to cancel the influence of the window function on the time axis signal in the window function calculating section 3 described above. Overlap at the overlap rate and add to each other,
The signal (digital signal) after the addition is converted to a D / A converter 21.
Is output to Here, the overlap ratio is not limited to 50%, and the number of overlapping frames is not limited to two or three.

The D / A converter 21 converts a digital audio signal in a noise period in which noise has been reduced or a digital signal in a noise period in which power has been attenuated into an analog signal. Then, the output is output from the output terminal 22 to a desired device, for example, a speaker, a recorder, or the like (not shown).

(A-2) Operation of the First Embodiment Next, the operation of the background noise elimination device of the first embodiment composed of the above components will be described.

Here, the input signal to the background noise elimination device is a signal in a voice band, and this input signal has a voice period after a noise period first, and thereafter, a noise period and a voice period alternate. It occurs in. In the case of an input signal from a storage medium, there may be no noise period at first, but hereinafter, description will be made assuming that there is a noise period first.

The input signal (analog signal) input from the input terminal 1 to the background noise canceller regardless of the noise period or the voice period is converted into a digital signal by the A / D converter 2 and then converted into a window function. In the arithmetic unit 3, after being divided into a predetermined frame length, the input signal for each analysis frame is subjected to a window function process in order to prevent high frequency components generated in the division process, and further, The FFT operation is performed by the FFT operation unit 4 to convert the signal on the time axis into a signal (spectral signal) on the frequency axis. The phase information obtained at the time of this FFT operation is given to and held by the phase holder 29.

The output signal from the A / D converter 2 is also output to a voice detector 28, which detects whether the signal is in a noise period or a voice period.

Here, since the beginning of the input signal is a voice period, the switch 6 is closed, the switch 9 is opened, and the switch 18 is turned on the side of the non-distortion attenuator 23 by the detection signal from the voice detector 28. Connected to input terminal a.

In the background noise estimating unit 30, the switch 6 is closed during the noise period as described above.
The spectrum signal (spectrum signal of noise) from the FFT calculator 4 is input to the background noise updating unit 7, and the functions of the waveform noise updating unit 7 and the background noise holding unit 8 cause the background noise updating unit 7 to output the FFT calculator. An estimated noise signal that follows the noise spectrum signal from the signal 4 is output to the adder 5. Note that the estimated noise signal from the background noise updating unit 7 is also supplied to the band divider 10, and the processing of the band divider 10 for this signal will be described later in the operation of a speech period. In the adder 5, an estimated residual signal is obtained by subtracting the estimated noise signal from the background noise updating unit 7 from the noise spectrum signal from the FFT calculator 4, and the power monitor of the distortionless attenuation unit 23 is obtained. Give 24. In the audio period described later, the signal from the adder 5 is an audio signal from which noise has been eliminated (including an erasure residual).

In the distortion-free attenuator 23, the power monitor 24 calculates the spectrum power of the noise spectrum signal from the FFT calculator 4 and the spectrum power of the residual signal of the noise estimation, and calculates the power attenuation. The unit 25 calculates the amount of attenuation of the noise signal from the FFT calculator 4 on the basis of these two spectral powers.
The attenuation from the power attenuation calculator 25 is smoothed so that the attenuation does not affect the steep and discontinuous changes in the noise. Then, in the power attenuator 27, only the power based on the attenuation amount from the attenuation amount smoothing operation unit 26 is set so that the noise spectrum signal from the FFT calculator 4 does not change the component ratio of the frequency component at all. The noise spectrum signal which has been attenuated and thus attenuated so as to maintain a natural tone is supplied to the inverse FFT calculator 19 via the switch 18.

In the inverse FFT calculator 19, the output spectrum signal from the power attenuator 27 is subjected to inverse FFT processing so as to have the phase held in the phase holder 29, and is converted into a signal on the time axis. Further, in the window function overlap calculating unit 20, a few window functions are formed with a 50% overlap ratio on the time axis, for example, in order to cancel the influence of the window function on the time axis signal in the window function calculating unit 3. The signals are overlapped and added to each other, and the added signal (digital signal) is converted into a digital signal in the D / A converter 21 and output from the output terminal 22 to the next processing device.

As described above, when the input signal is in the noise period, an estimated noise signal is formed in the background noise estimator 30 so as to follow the input signal, and the noise is reduced in the distortionless attenuator 23. Is attenuated and output with a natural sense of hearing such that the volume does not change sharply.

When the noise period ends and the input signal enters the voice period, the voice detector 28 detects that, and
The switch 6 is opened, the switch 9 is closed, and the switch 18 is connected to the input terminal b on the band-specific spectrum attenuation processing unit 17 side.

In this voice period, the distortionless attenuator 23 operates because there is an input signal. However, since the output side thereof is disconnected by the switch 18, it is equal to absent.

In the background noise estimating unit 30, as described above, the switch 6 is opened during the voice period, and the output spectrum signal from the FFT calculator 4 is not supplied to the background noise updating unit 7, so that the background noise updating unit 7 is updated to the immediately preceding value even after the updating operation (that is, equivalent to not performing the updating operation).
The estimated noise signal at the end of the noise period is repeatedly supplied to the adder 5 and the band divider 10 of the band-specific spectrum attenuation processing unit 17 in each analysis frame.

In this voice period, the adder 5 includes the FFT
A speech spectrum signal on which noise is superimposed is given from a computing unit 4. In the adder 5, an estimated noise signal fixed to a value at the end of the noise period is subtracted from the speech spectrum signal, and noise is added. The components are eliminated (including the erasure residual) and, via switch 9 in the closed state,
The voice spectrum signal after the noise cancellation is provided to the band divider 10 of the band-specific spectrum attenuation processing unit 17.

In the band-specific spectrum attenuation processing unit 17, first, in the band divider 10, the speech spectrum signal after noise elimination and the estimated noise signal at the end of the noise period from the background noise updating unit 7 are respectively plural. Divided into bands.

Thereafter, the power of the estimated noise signal and the power of the speech spectrum signal after noise elimination are calculated in the power calculation unit 12 for each of the divided bands, and the logarithmic calculation unit 13
, The logarithmic value of the power is obtained, and the band s / n
The calculation unit 14 calculates (estimates) the S / N ratio based on the two types of logarithmic values, and the band-specific attenuation calculation unit 15 calculates, based on the S / N ratio, a band having more noise components. The amount of attenuation for the speech spectrum signal after noise elimination for each band, which greatly attenuates, is calculated.

Then, the voice spectrum signal after noise removal for each band output from the band divider 10 is converted into the attenuation amount for each band output from the attenuation calculation unit 15 for each band by the attenuation calculation unit 11 for each band. The spectrum signal after the attenuation process is smoothed by the smoothing operation unit 16, and even if the signal from the band-based attenuation operation unit 11 includes a steep change, it is weakened.

The voice spectrum signal from the band-specific spectrum attenuating unit 17 that has been subjected to the band-specific attenuating process as described above is used as the switch 18 for selecting the band-specific spectral attenuating unit 17 based on the voice detection signal. To the inverse FFT calculator 19.

Thereafter, in the same manner as in the noise period, in the inverse FFT calculator 19, the output spectrum signal from the smoothing calculator 16 is subjected to inverse FFT processing so as to have the phase held in the phase holder 29. Then, the window function is converted into a signal on the time axis, and further, in the window function overlap calculator 20, in order to cancel the influence of the window function on the time axis signal in the window function calculator 3, two or three window functions are, for example, 5 on the time axis
The signals are overlapped with each other at an overlapping rate of 0% and added together, and the added signal (digital signal) is converted into a digital signal in the D / A converter 21 and output from the output terminal 22 to the next processing unit. .

As described above, when the input signal is in the voice period, the updating of the estimated noise signal to a new value in the background noise estimating unit 30 is stopped, and the voice is determined based on the fixed estimated noise signal. The noise superimposed on the signal is eliminated, and the speech spectrum signal after the noise elimination is subjected to
The s / n ratio is estimated for each band, and the band with the smaller estimated s / n ratio is intensively attenuated and then smoothed to prevent the sound volume from fluctuating sharply and becoming unnatural. The sound quality is natural and clear.

When the voice period ends and the next noise period starts, the above operation in the noise period is executed.
The operation in the above-described voice period is also performed in the voice period next to the noise period.

(A-3) Effects of the First Embodiment According to the background noise elimination device of the first embodiment, the switch 6 that operates according to the detection result from the voice detector 28 is provided, and the background noise is excluded except during the noise period. Since the updating operation of the noise estimating unit 7 is stopped, the noise is not attenuated by using the estimated noise signal erroneously updated with the signal in the voice period, thereby preventing the sound quality from being deteriorated. it can.

Further, according to the background noise canceling apparatus of the first embodiment, the distortionless attenuator 23 which operates effectively during the noise period is provided, and the frequency component of the noise spectrum signal input to the distortionless attenuator 23 is provided. The remaining component noise, which was previously distorted and very unnatural due to distortion, can be attenuated with a natural sense of perception. Sound quality can be improved.

In this way, in the distortion-free attenuation section 23, the attenuation amount is smoothed using the attenuation amount smoothing operation section 26 and then applied to the attenuation processing, so that the attenuation amount changes sharply. Can be prevented, and the noise volume can be prevented from changing sharply and becoming unnatural.

Further, according to the background noise canceller of the first embodiment, the band-specific spectrum attenuation processing unit 17 estimates the S / N during the voice period and concentrates the low S / N channels (bands). That is, since the channel component having a large amount of noise component is largely attenuated, and the channel having a large amount of audio component is not attenuated so much, it is possible to obtain a noise removing effect having excellent sound quality.

In this case, a smoothing operation unit 16 is provided in the band-specific spectrum attenuation processing unit 17 so as to prevent a sharp change in the amount of attenuation. It is possible to prevent the sound volume from fluctuating sharply and become unnatural, and it is possible to obtain a natural and clear sound quality from which a noise component has been removed during a voice period.

(B) Second Embodiment Next, a second embodiment of the background noise canceller according to the present invention will be described in detail with reference to the drawings.

FIG. 3 is a block diagram showing the overall configuration of the background noise canceling apparatus according to the second embodiment, in which the same or corresponding parts as those in FIG. 1 are denoted by the same reference numerals.

As is clear from the comparison between FIG. 3 and FIG. 1, the background noise canceling apparatus of the second embodiment has the switch 9 in the background noise canceling apparatus of the first embodiment, The difference is that they are replaced.

The coefficient determination unit 42 also has the voice detector 28
Is supplied to the coefficient determination unit 42.
Is the speech period and the noise period, which are used again by the background noise updating unit 7 for the update operation of the estimated noise signal N'f, k.
This is to switch the noise estimation speed determination coefficient a in the equation (6).

N′f, k = a · N′f, k−1 + (1−a) · Xf, k (6) That is, in the first embodiment, the estimated noise signal N ′ in the voice period. The update of f, k was stopped, but this second
In the embodiment, the estimated noise signal N'f, k is updated using the noise estimation speed determination coefficient a different from the noise period also in the speech period. More specifically, during the speech period, the noise estimation speed determination coefficient a (hereinafter, referred to as “the noise estimation speed determination coefficient a” (hereinafter, referred to as “the estimated noise signal N′f, k”) is updated so that the speech signal is not reflected in the background noise update process. In the noise period, the noise estimation speed determination coefficient a is set so that the noise can be quickly followed.
(Hereinafter, the coefficient in the noise period is represented by a0).

The coefficients a0 and a1 (1> a) in each of these periods
0> 0, 1>a1> 0) may be arbitrarily selected if a0 <a1 is satisfied. For example, a0 = 0.1 in a noise period and a1 = 0.9 in a voice period. You can do it.

As described above, in the case of the second embodiment, the background noise updating unit 7 is provided with the noise estimation speed determining coefficient a0 or a1 different from the coefficient determining unit 42 in the noise period and the speech period. The noise updating unit 7 determines (2
The estimated noise signal N'f, k is formed according to the equation (1), and the estimated noise signal N'f, k is formed according to the equation (22) during the voice period.

N′f, k = a0 · N′f, k−1 + (1−a0) · Xf, k (21) N′f, k = a1 · N′f, k−1 + (1 −a1) · Xf, k (22) Here, in the noise period, the output spectrum signal Xf, k from the FFT calculator 4 is the background noise spectrum Nf, k itself, and the background noise updating unit 7 When the spectrum is correctly estimated, the estimated value by the background noise estimation algorithm automatically converges. When the convergence is made in this manner by the repetitive calculation of the equation (21) for each analysis frame, N ′
i, k-1 = Xf, k, and as a result, N'f, k = Xf, k.

Then, when the noise period ends and the sound period starts, in the second embodiment, the coefficient is switched from a0 to a1 and the noise estimation is continued as described above. As a result, even during the speech period, an estimated noise signal that follows noise fluctuation can be formed, and the speech quality can be effectively secured when the noise to be removed is separated from the frequency line segment of the speech signal. Become.

In the second embodiment as well, the operation of each unit other than the background noise estimating unit 30 is the same as in the first embodiment, and the description of the operation will be omitted.

Therefore, according to the background noise canceling apparatus of the second embodiment, the update speed of the estimated value of the background noise is changed according to the detection result from the voice detector 28, and the estimation is performed at high speed in the noise period. In the audio period, the update is performed at a low speed, so that even in the audio period, if there is noise other than the audio frequency, the estimation value can be updated by following the noise, and a large erroneous update of the noise estimation occurs, thereby deteriorating the sound quality. This can be prevented beforehand, and noise that does not overlap with the audio component can be effectively removed.

Further, the background noise canceller of the second embodiment also includes the distortionless attenuator 23 and the band-specific spectrum attenuator 17 so that these devices are provided similarly to the first embodiment. Of course, the effect of this is achieved.

(C) Third Embodiment Next, a third embodiment of the background noise canceller according to the present invention will be described in detail with reference to the drawings.

FIG. 4 is a block diagram showing the overall configuration of a background noise canceling apparatus according to the third embodiment, in which the same or corresponding parts as those in FIG. 1 are denoted by the same reference numerals.

As is clear from the comparison between FIG. 4 and FIG. 1, the background noise canceller of the third embodiment is different from the background noise canceller of the first embodiment in that the input of the spectrum-attenuation processing unit 17 for each band is provided. The switch 9 provided in the stage is eliminated, and instead, the spectrum attenuation processing unit 1 for each band is used.
7, which controls the passing / non-passing of the spectrum signal (represented by Xlf, k in the fourth embodiment) outputted from the smoothing operation unit 16. Power monitor 24
In that a switch 51 is provided.

The voice detection signal output from the voice detector 28 is also supplied to the switch 51, and the switch 51 closes during the noise period and passes the spectrum signal Xlf, k output from the smoothing operation unit 16. The signal is supplied to the power monitor 24 and is opened during the voice period to prevent the spectral signal Xlf, k output from the smoothing operation unit 16 from being supplied to the power monitor 24.

In the third embodiment, since the switch 9 is not provided at the input stage of the band-specific spectrum attenuation processing section 17, the band-specific spectrum attenuation processing section 17 operates effectively even in the noise period. . However, since the switch 18 selects the spectrum signal from the distortionless attenuating section 23 during the noise period, the output spectrum signal Xlf, k from the smoothing operation section 16 (band-specific spectrum attenuation processing section 17) is subjected to the inverse FFT operation. It is not input to the device 19.

On the other hand, as described above, the output spectrum signal Xlf, k from the smoothing operation section 16
It is provided to the power monitor 24 via the switch 51.

The power monitor 24 of the third embodiment
Calculates the spectrum power PXl of the spectrum signal Xlf, k from the smoothing operation unit 16 according to the equation (23), and calculates the spectrum power PX of the output spectrum signal Xf, k of the FFT calculator 4 according to the equation (24). Calculated and given to the power attenuation calculator 25. Note that the sum Σ in Equations (23) and (24) is performed for all frequency components.

PXlk = ΣXlf, k (23) PXk = ΣXf, k (24) The power attenuation amount calculation unit 25 of the third embodiment is configured as follows:
The basic value LXk of the amount of attenuation with respect to the output signal Xf, k of the FFT operation unit 4 is calculated according to the equation, and is output to the attenuation amount smoothing operation unit 26.

LXk = PXlk / PXk (25) Also in the third embodiment, the operations of the attenuation smoothing calculation unit 22 and the power attenuation unit 27 in the distortionless attenuation unit 23 are the same as in the first embodiment. is there.

Also, as described above, except that the band-specific spectrum attenuation processing unit 17 operates effectively even in the noise period, the operation of each unit other than the distortionless attenuation unit 23 is the first operation.
This is the same as the embodiment.

As described above, in the first embodiment and the third embodiment, the estimated residual signal in the noise period given to the power monitoring unit 24 of the distortionless attenuation unit 23 is output from the adder 5. Although there is a difference between using the signal output from the adder 5 and the signal output from the adder 5 after performing band-specific spectrum attenuation processing, the noiseless The third embodiment is similar to the first embodiment in that the amount of attenuation is determined based on the estimated residual signal.

According to the background noise canceling apparatus of the third embodiment, whether the estimated residual signal in the noise period applied to the power monitoring unit 24 of the distortionless attenuating unit 23 is output from the adder 5 is used. Although there is a difference as to whether a signal output from the adder 5 is subjected to spectral attenuation processing for each band, the overall processing is the same as that of the first embodiment. The same effect as described above can be obtained.

(D) Fourth Embodiment Next, a fourth embodiment of the background noise canceller according to the present invention will be described in detail with reference to the drawings.

FIG. 5 is a block diagram showing the overall configuration of a background noise canceling apparatus according to the fourth embodiment, in which the same or corresponding parts as those in FIG. 1 are denoted by the same reference numerals.

As is clear from the comparison between FIG. 5 and FIG. 1, in the background noise elimination apparatus of the fourth embodiment, the distortionless attenuation section 23 performs the distortionless attenuation processing on the noise at the signal stage on the time axis. The point of execution is significantly different from the background noise canceller of the first embodiment.

That is, in the background noise canceller of the fourth embodiment, specifically, the output signal of the A / D converter 2 is supplied to the power monitor 24 of the distortionless attenuator 23, and the window function is The output signal of the lap processing unit 20 is supplied to the power monitor 24 of the distortionless attenuation unit 23, and the output signal from the power attenuation unit 27 of the distortionless attenuation unit 23 is supplied to a D / A converter via a switch 62 described later. 21. In the audio period, the signal processed by the band-specific spectrum attenuation processing unit 17 is used as an output signal from the device, and in the noise period, the signal processed by the distortionless attenuation unit 23 is used as the output signal of the device. In order to obtain an output signal from the first embodiment, the switches 9 and 18 of the first embodiment are omitted, and an on / off switch inserted between the window function overlap processing unit 20 and the D / A converter 21 is used instead. 61 and an on / off switch 62 inserted between the power attenuator 27D / A converter 21.

The switches 61 and 62 also have the sound detector 2
8 is provided. The switch 61
Window function overlap processing unit 2 closed during voice period
The output signal from 0 is passed to the D / A converter 21 and opened during the noise period to prevent the output signal from the window function overlap processing unit 20 from passing. On the other hand, the switch 62 is closed during the noise period, passes the output signal from the distortionless attenuator 23, and hence the output signal from the power attenuator 27, and supplies it to the D / A converter 21. The output signal from the section 27 is blocked.

Also in the fourth embodiment, since the switch 9 is not provided at the input stage of the band-specific spectrum attenuation processing unit 17, the band-specific spectrum attenuation processing unit 17 operates effectively even in the noise period. . However, since the switch 61 provided on the output side of the window function overlap processing unit 20 is closed only during the audio period, it is output from the band-specific spectrum attenuation processing unit 17 during the noise period, and then the inverse FFT calculator 19 The signal processed by the window function overlap processor 20 is not provided to the D / A converter 21.

However, in the fourth embodiment,
In the noise period, the signal on the time axis output from the band-specific spectrum attenuation processing unit 17 and then processed by the inverse FFT calculator 19 and the window function overlap processing unit 20 is as follows:
The power is supplied to the power monitor 24 of the distortionless attenuator 23. As described above, the power monitor 24 is also provided with a signal on the time axis from the A / D converter 2. In addition, since the switch 62 of the output stage is closed only during the noise period,
It works effectively only in the noise period.

As described above, unlike the first to third embodiments, the signal on the time axis is input to the distortionless attenuator 23 of the fourth embodiment. An operation different from that of the distortionless attenuation unit 23 of the embodiment is performed.

The power monitor 24 calculates the power Pxt, k of the output signal xt, k from the A / D converter 2 according to, for example, equation (26), and also calculates the window function overlap processor 2
0, the power Pwt, k of the output signal wt, k is, for example, (2
The power is calculated in accordance with the equation (7), and both powers are output to the power attenuation calculator 25. Note that the sum Σ in the equations (26) and (27) is the number of time-hold samples 1 for power monitoring, including the time t, and the number of samples 1 may be arbitrarily selected. However, for example, 32 samples can be selected.

[0129] Pxt, k = Σxt-l, k 2 ... (26) Pwt, k = Σwt-1, k 2 ... (27) power attenuation amount calculation unit 25, the output signal from the A / D converter 2 xt , k is calculated according to, for example, equation (28) and provided to the attenuation smoothing operation unit 26.

Lxk = Pwt, k / Pxt, k (28) The attenuation smoothing operation unit 26 calculates the steepness of each processing target frame.
Even if the attenuation amount Lx fluctuates discontinuously, the attenuation amount Lx is made continuous so that the level of the processed sound (noise) does not become discontinuous and the sound feeling becomes natural and good. For example, (29)
The smoothing operation shown in the equation is performed to obtain a smooth gain Lxsm, which is provided to the power attenuator 27. Here, ε in the equation (29) is a smoothing coefficient for determining a smoothing speed of the smoothing operation, and may be appropriately determined according to a use environment and noise to be used.

Lxsm, k = ε · Lxsm, k−1 + (1−ε) · Lxk (29) The power attenuator 27 outputs the output signal x from the A / D converter 2.
Attenuation processing is performed on t, k by using the attenuation amount Lxsm, k after the smoothing processing from the attenuation amount smoothing operation unit 26 as shown in Expression (30), and the attenuated signal xlt, k is obtained. Switch 62
Output to

Xlt, k = Lxsm, k · xt, k (30) Here, the attenuated signal xlt, k is a signal in which only the amplitude is reduced on the time axis, and also in the fourth embodiment, It is obtained as a spectral component that reproduces a natural tone with only the power attenuated without changing the component ratio of the frequency component of the input signal xt, k to the distortionless attenuator 23 at all.

In the fourth embodiment, the operation of each unit other than the distortionless attenuator 23 is the same as that of the first embodiment except that the band-specific spectrum attenuator 17 operates effectively even in the noise period. Same as the form.

As described above, the difference between the first embodiment and the fourth embodiment is whether the distortionless attenuator 23 processes a signal on the frequency axis or a signal on the time axis. Other than that, the third embodiment is the same as the first embodiment.

In the background noise canceller of the fourth embodiment as well, there is a difference whether the distortionless attenuator 23 processes a signal on the frequency axis or a signal on the time axis.
Since the entire processing is the same as in the first embodiment, the same effects as those in the first embodiment can be obtained.

(E) Other Embodiments Although various embodiments have been described in the above embodiments, the present invention is not limited to the above embodiments. Other embodiments other than the modified embodiment described above are also permitted. Some examples are as follows.

The conversion between the signal on the time axis and the signal on the frequency axis may be performed by means other than the FFT and the inverse FFT operation. Also, sound detection may be performed using a signal on the frequency axis.

The intervening position of the switch for switching the processing system of the output signal between the voice period and the noise period is not limited to the above embodiment. For example, in the case of the distortionless attenuator, a switch may be provided on the input side.

In the above embodiments, the method of updating the estimated noise signal in the speech period in the background noise estimating section, the provision of the band-specific spectrum attenuation processing section, and the provision of the distortionless attenuation section are described.
It was characterized by points, but one of them
It goes without saying that the background noise canceller may be configured to have the above.

[0140]

As described above, according to the present invention, the updating of the estimated noise signal is stopped in the speech period or the updating is performed at a speed following the input signal that is slower than the noise period. Noise is not attenuated using the estimated noise signal erroneously updated with the signal of the period, so that it is possible to prevent sound quality from being deteriorated.

Further, according to the present invention, since the spectrum attenuation processing means for attenuating the output signal after the noise elimination operation from the noise elimination operation means and outputting it to the frequency axis / time axis conversion means is provided, the sound quality is excellent. The noise removal effect can be obtained.

Further, according to the present invention, in the noise period, the distortionless attenuating means for attenuating the input signal on the time axis or the frequency axis without changing the sound quality, and in the noise period,
Since output selection means for providing an output signal from the distortionless attenuating means as an output signal of the apparatus is provided, it is possible to attenuate the noise inputted to the distortionless attenuating means so as not to change the component ratio of the frequency components at all. In contrast, residual noise that has been conventionally distorted and extremely unpleasant and unnatural can be attenuated with a natural sense of hearing, and the sound quality during the noise period can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating an overall configuration of a first embodiment.

FIG. 2 is a block diagram showing a conventional configuration.

FIG. 3 is a block diagram illustrating an overall configuration of a second embodiment.

FIG. 4 is a block diagram illustrating an overall configuration of a third embodiment.

FIG. 5 is a block diagram showing an overall configuration of a fourth embodiment.

[Explanation of symbols]

4: Fast Fourier transform operation unit (FFT operation unit), 5: Adder, 6, 9, 18, 51, 61, 62 ... Switch, 7
... Background noise updating section, 8... Background noise holding section, 10... Frequency band splitter, 11, attenuation calculation section for each band, 12... Power calculation section, 13 .logarithm calculation section, 14. (S / n calculation unit for each band), 15: attenuation calculation unit for each band, 16: smoothing operation unit, 17: spectrum attenuation processing unit for each band, 19: inverse fast Fourier transform operation unit (inverse FFT operation unit), 23 ... No distortion attenuator, 24 ... Power monitor, 25 ...
Power attenuation calculation unit, 26 ... Attenuation smoothing operation unit, 27 ...
Power attenuator, 28: voice detector, 29: phase holder,
30: background noise estimation unit; 42: coefficient determination unit.

──────────────────────────────────────────────────続 き Continued on front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 21/02

Claims (9)

(57) [Claims] 1. A time axis / frequency axis converting means for converting an input signal on a time axis into a signal on a frequency axis, and inputting background noise in the input signal based on the converted input signal on the frequency axis. Background noise estimating means for estimating while following the signal to form an estimated noise signal, noise canceling means for subtracting the estimated noise signal from the converted input signal on the frequency axis, and a frequency on which the noise canceling operation has been performed. A background noise canceling device having a frequency axis / time axis converting means for converting a signal on the axis into a signal on the time axis, wherein the input signal detects only background noise based on the input signal on the time axis or the frequency axis. And a voice detection unit for detecting whether the input signal is in a noise period including the voice signal or a voice period including the voice signal. The following speed, background noise erasing apparatus characterized in that a follow-up speed switching unit for switching to be faster than the speech period is noise period. 2. A time axis / frequency axis converting means for converting an input signal on a time axis into a signal on a frequency axis, and inputting background noise in the input signal based on the converted input signal on the frequency axis. Background noise estimating means for estimating while following the signal to form an estimated noise signal, noise canceling means for subtracting the estimated noise signal from the converted input signal on the frequency axis, and a frequency on which the noise canceling operation has been performed. A background noise canceling device having a frequency axis / time axis converting means for converting a signal on the axis into a signal on the time axis, wherein the output signal from the noise canceling arithmetic means is attenuated and the frequency axis / time axis converting means 2. A background noise canceling device comprising a spectrum attenuation processing means for outputting a signal to a background noise canceller. 3. A time axis / frequency axis converting means for converting an input signal on the time axis into a signal on the frequency axis, and inputting background noise in the input signal based on the converted input signal on the frequency axis. Background noise estimating means for estimating while following the signal to form an estimated noise signal, noise canceling means for subtracting the estimated noise signal from the converted input signal on the frequency axis, and a frequency on which the noise canceling operation has been performed. A background noise canceling device having a frequency axis / time axis converting means for converting a signal on the axis into a signal on the time axis, wherein the input signal detects only background noise based on the input signal on the time axis or the frequency axis. A voice detection unit for detecting whether the input signal is in a noise period including the voice signal or a voice period including the voice signal; and in the noise period, attenuates the input signal on the time axis or the frequency axis without changing the sound quality. A distortion damping means, in the noise period, background noise erasing apparatus characterized by an output signal from the distortion-free attenuating means is provided and an output selection means to the output signal of the apparatus. 4. A time axis / frequency axis converting means for converting an input signal on a time axis into a signal on a frequency axis, and inputting background noise in the input signal based on the converted input signal on the frequency axis. Background noise estimating means for estimating while following the signal to form an estimated noise signal, noise canceling means for subtracting the estimated noise signal from the converted input signal on the frequency axis, and a frequency on which the noise canceling operation has been performed. A background noise canceling device having a frequency axis / time axis converting means for converting a signal on the axis to a signal on the time axis, wherein the input signal is based on the input signal on the time axis or the frequency axis, and A voice detection unit for detecting whether the input signal is in a noise period including the voice signal or a voice period including the voice signal; and attenuating an output signal from the noise elimination calculation unit and outputting the signal to the frequency axis / time axis conversion unit. Spectrum attenuating processing means, and a distortion-free attenuating means for attenuating an input signal on a time axis or a frequency axis in a noise period without changing a sound quality; And output selection means for setting a signal subjected to the processing by the spectrum attenuation processing means to an output signal of the apparatus during a voice period. A background noise canceling device, further comprising an update stopping unit for stopping updating of the value of the estimated noise signal to be output during a voice period. 5. A time axis / frequency axis converting means for converting an input signal on a time axis into a signal on a frequency axis, and inputting background noise in the input signal based on the converted input signal on the frequency axis. Background noise estimating means for estimating while following the signal to form an estimated noise signal, noise canceling means for subtracting the estimated noise signal from the converted input signal on the frequency axis, and a frequency on which the noise canceling operation has been performed. A background noise canceling device having a frequency axis / time axis converting means for converting a signal on the axis into a signal on the time axis, wherein the input signal detects only background noise based on the input signal on the time axis or the frequency axis. A voice detection unit for detecting whether the input signal is in a noise period including the voice signal or a voice period including the voice signal; and attenuating an output signal from the noise elimination calculation unit and outputting the signal to the frequency axis / time axis conversion unit. Spectrum attenuating means for attenuating an input signal on a time axis or a frequency axis during a noise period without changing the sound quality; and an output signal from the distortionless attenuating means during a noise period. And output selection means for setting a signal subjected to the processing by the spectrum attenuation processing means to an output signal of the apparatus during a voice period. A background noise canceling device, further comprising: a tracking speed switching unit that switches a tracking speed of an output estimated noise signal to an input signal so that the tracking speed is higher in a noise period than in a voice period. 6. An S / N calculator for calculating an S / N of an input signal to the spectrum attenuation processing means, and an attenuation for determining an attenuation amount according to the obtained S / N. 6. The background noise canceling apparatus according to claim 2 , further comprising an amount calculating unit, and an attenuation calculating unit configured to attenuate an input signal to the spectrum attenuation processing unit with the determined amount of attenuation. 7. The spectrum attenuation processing means further includes a band divider for dividing an input signal to the spectrum attenuation processing means into a plurality of bands, the S / N calculator, the attenuation calculator, 7. The apparatus according to claim 6, wherein each of the attenuation calculation units performs processing for each divided band.
2. The background noise canceller according to item 1. 8. The background noise canceling apparatus according to claim 7 , wherein said spectrum attenuation processing means has a smoothing operation section for smoothing an output signal for each band from said attenuation operation section. 9. The non-distortion attenuating means includes: a signal in a noise period before the noise canceling operation is performed; a power monitor for monitoring the power after the noise canceling operation is performed; A power attenuation calculator that determines the attenuation based on the attenuation, an attenuation smoothing calculator that smoothes the determined attenuation, and an input signal to the distortionless attenuation unit based on the smoothed attenuation. The background noise canceling device according to any one of claims 3 to 8 , further comprising a power attenuating unit that attenuates.