JPH09171397A - Background noise eliminating device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a background noise eliminating device capable of surely eliminating even a narrow-bond frequency and unsteady noise component. SOLUTION: A sound containing noise component is inputted from an input terminal IN. A sound detector 8 detects the presence of the sound in the input signal for eliminating the noise component from the input signal by subtraction in a frequency area. The Foulier-transformed input signal is calculated on its power by a signal power calculation part 14. A background noise renewal part 11 estimates a background noise only for a prescribed time from the time when the sound detection by the sound detector 8 was ended, and further, a background noise holder 12 holds its estimated value. Then, an adder 15 subtracts background noise power from the signal power calculated by the signal power calculation part 14 to output it.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice input device,
In particular, it relates to a noise canceller in the transmitter section of a telephone.

[0002]

2. Description of the Related Art Usually, communication of a will with a machine is performed by a button,
Inputting means such as a mouse or keyboard and outputting means such as characters and images using a display. However, voice interaction is easy and natural for humans, and in recent years, attempts have been made to put into practical use a voice interface based on voice recognition technology for promoting voice interaction with machines, not limited to telephones. It is being appreciated.

For a voice input device in such a background, a particular problem is to separate the noise component and the voice signal from the input signal. Conventionally, for example, a spectral subtraction method has been developed as a method of adaptively estimating noise and extracting only a voice signal (Taniguchi et al. “Spectral subtraction method of adaptively estimating noise”) Report "TECHNICAL REPORT OF I
ECE.SP94-116 (1995-03) ").

[0004]

According to the article by Taniguchi et al. Mentioned above, it is possible to perform adaptive noise estimation which operates without detecting a voice section. However, if this method is directly used for voice communication, the background noise after processing has a tone-like property and the time variation is small, so
There was a problem that background noise was annoying. Also,
When the background noise estimation coefficient is reduced, the average effect of background noise estimation is large, but the estimation speed becomes slow. For this reason,
The initial startup characteristics of the device are poor. There is also a problem that the processed sound is greatly deteriorated when the background noise changes suddenly or immediately after the audio signal.

In order to avoid these problems, a method of detecting a noise level after dividing into a plurality of band signals by a channel divider is described in R. J. Proposed by Burmer et al. (R. J. Vi1mur)
Etc. "NOISE SUPPRESSION SYSTEM" (US
PAT.4,811,404)). This method is called a spectral gain conversion method, and switches the SN ratio for each channel divided by voice metric calculation.

However, even when this method is adopted, when a voice signal consisting of only a buzzer sound or a vowel for a long period is continuously input for, for example, 1 second, it is regarded as noise and is erased. There was a problem of being lost. If the frequency characteristic of the audio signal has a peak characteristic,
In particular, for a voice signal in which a component without a large frequency spread such as a vowel is continuous, this is regarded as noise and is forcibly erased. As a result, there is a problem in that the sound quality of the noise-suppressed audio signal deteriorates.

The present invention has been made in order to solve the above problems, and it is an object of the present invention to provide a background noise canceling apparatus capable of surely canceling a narrow band non-stationary noise component. To aim.

A second object of the present invention is that the initial rising response of the device is agile and even if the background noise changes suddenly,
It is an object of the present invention to provide a background noise canceling device in which processed sound is stable and does not deteriorate.

A third object of the present invention is to provide a background noise canceller which is excellent in speech quality because voice is not deleted even if voice is continuously input.

A fourth object of the present invention is to provide a background noise canceling apparatus in which the frequency characteristic of a voice signal is not canceled even if it does not have a large frequency spread.

[0011]

According to a first aspect of the present invention, there is provided a background noise canceller for removing a noise component by subtracting in a frequency domain from an input signal input when voice includes a noise component. Voice detecting means for detecting the presence or absence of voice, signal power calculating means for calculating the power of the input signal, and background noise is estimated for a predetermined time after the voice detection by the voice detecting means is completed, and the estimation thereof is performed. It is characterized by comprising background noise estimation holding means for holding a value, and calculation means for subtracting the background noise power of the background noise estimation holding means from the signal power calculated by the signal power calculation means and outputting the subtracted background noise power.

According to a second aspect of the present invention, there is provided smoothing operation means for smoothing the signal power output from the operation means, and the background noise updating means causes the apparatus to rise at an initial time and a fixed time after the end of the voice period. Is characterized in that the smoothing calculation in the smoothing calculating means is performed steeply and the others are performed gently.

According to a third aspect of the present invention, the background noise estimation and holding means includes a background noise holder that holds the estimated background noise, and a switch that stops updating the background noise according to the detection result of the voice detector. It is characterized by

According to a fourth aspect of the present invention, the background noise updating means changes the smoothing operation coefficient of the smoothing operation means in the voice period, the device starting period and the noise period, respectively.

According to a fifth aspect of the present invention, a band divider for dividing the output of the arithmetic means into a plurality of predetermined bands and a logarithmic calculation for logarithmically converting the signal value of each band divided by the band divider. Section, an s / n calculator that calculates s / n for each region based on the logarithmic value calculated by the logarithm calculator, and a power that attenuates the power of each channel for each band according to the calculated s / n An attenuator and a smoothing calculator for smoothing the output of the power attenuator are provided, and the coefficient of the smoothing calculator is variable.

According to a sixth aspect of the present invention, the smoothing processing of the smoothing arithmetic unit is performed steeply for a certain time after the initial rising of the device and after the end of the voice period, and is gently performed for the rest.

According to a seventh aspect of the present invention, there is provided a peak frequency exploring section for determining a peak frequency of a narrow band signal, an adjacent frequency power monitoring section for calculating power at an adjacent frequency of the peak frequency from an output of the peak frequency exploring section, A narrow band noise determination unit that determines narrow band noise from the adjacent frequency power monitoring unit is provided, and narrow band noise is reduced by attenuating a component of a specific channel or a specific frequency according to the determination result of the narrow band noise determination unit. It is characterized by removing.

According to the invention of claim 8, a peak frequency multiple relation search unit for calculating a multiple relation of each peak frequency based on the search result of the peak frequency search unit, the adjacent frequency power monitoring unit and the peak frequency multiple relation. A narrow band noise determination unit that determines narrow band noise from the search unit is provided.

According to a ninth aspect of the present invention, the specific frequency is attenuated by reducing the gain created by the power attenuating section according to the determination result of the narrow band noise determining section. To do.

According to a tenth aspect of the present invention, a specific frequency component of the band divider is added to the background noise holder in accordance with the determination result of the narrow band noise determination section, and further the input signal is input from the input signal by the calculating means. It is characterized in that the specific frequency is attenuated by subtraction.

According to the invention of claim 11, an A / D converter for converting the input signal into a digital signal, and a fast Fourier converter for converting the digital signal into a frequency domain signal by a fast Fourier transform process, A fast inverse Fourier transformer for retransforming a frequency domain signal into a time domain signal,
And a D / A converter for converting the reconverted digital signal into an analog signal.

[0022]

BRIEF DESCRIPTION OF THE DRAWINGS FIG.
Embodiments of the present invention will be described in detail.

First Embodiment FIG. 1 is a block diagram showing a first embodiment of the present invention.

In FIG. 1, 1 is an analog / digital converter (hereinafter referred to as A / D converter), 2 is a window function calculator, 3 is a fast Fourier transformer (hereinafter referred to as FFT),
Reference numeral 4 is a signal phase calculator, 5 is a fast inverse Fourier transformer (hereinafter referred to as IFFT), 6 is a window function overlap calculator, and 7 is a digital / analog converter (hereinafter referred to as D / A converter).

An input signal is supplied to the A / D converter 1 from the input terminal IN as analog voice and is converted into a digital signal. The digital signal converted by the A / D converter 1 is divided into frame-based signals and input to the window function calculator 2 and the voice detector 8. In the window function calculator 2, window calculation processing, for example, hamming (hammin
g) Well-known window function calculation processing such as window calculation or hanning window calculation is performed. The output of the window function calculator 2 is output to the FFT 3, and the FFT 3 converts the input signal into a signal component in the frequency domain by a fast Fourier transform process.

Reference numeral 8 is a voice detector, and the installation position of the voice detector is, for example, after the A / D converter 1 but also after the window function calculator 2 or after the adder 15 described later. , The latter stage of the band divider 13, and the band-specific power attenuator 18
It may be provided at any one of the following stage, the smoothing calculator 19, and the ΙFFT5, and the accuracy of the voice detection can be improved by arranging them respectively at these plural positions.

In FIG. 1, 9 is a first counter and 10 is a counter.
Is a switch, 11 is a background noise updating unit, 12 is a background noise holder, 13 is a frequency band divider, 14 is a signal power calculation unit, and 15 is an adder.

The signal converted by the FFT 3 is output to the signal phase calculator 4 and the signal power calculator 14.
The first counter 9 is connected to the voice detector 8, counts the time from the start of the voice detection operation, and sets the update permission condition for the background noise updating unit 11. The switch 10 is also connected to the voice detector 8, and the voice detector 8
The switch 10 is controlled by the detection result of 1), and when the detection result of the voice detector 8 is sound, the switch 10 is opened, and the switch 10 is closed during the other period. The background noise holder 12 is connected to the background noise updating unit 11 to form a background noise updating and holding unit. The signal power calculation unit 14 is connected to the background noise updating unit 11 via the switch 10 and also connected to the + input terminal of the adder 15. The-input terminal of the adder 15 is connected to the background noise holder 12, and the output of the adder 15 is input to the frequency band divider 13 and a peak frequency search unit 21 described later.

Further, in FIG. 1, 16 is a logarithmic calculation unit,
17 is a signal-to-noise ratio (hereinafter referred to as s / n) calculator,
Reference numeral 18 is a band-based power attenuator, 19 is a smoothing calculator, and 20 is a second counter.

The calculation result of the adder 15 is input to the frequency band divider 13, where the output of the adder 15 is divided into a plurality of predetermined bands. The output from the background noise holder 12 is also input to the band divider 13. The frequency band divider 13 is connected to the logarithmic calculator 16 and the band-specific power attenuator 18. The logarithmic calculation unit 16 further adds s
It is connected to the band-specific power attenuator 18 via the / n calculator 17. The output of the band-based power attenuator 18 is input to the smoothing calculator 19. In addition to this, the detection output from the voice detector 8 and the output signal from the second counter 20 are input to the smoothing calculator 19. In addition, the second counter 2
The detection output of the voice detector 8 is input to 0.

Further, in FIG. 1, 21 is a peak frequency search unit, 22 is an adjacent frequency power monitoring unit, 23 is a narrow band noise determination unit, and 24 is a peak frequency multiple relation search unit.

The peak frequency search unit 21 is for calculating the peak frequency of the narrow band signal from the output signal of the adder 15, and the calculated peak frequency is the adjacent frequency power monitoring unit 22 and the peak frequency multiple relation searching unit. 24 is output. Further, the narrow band noise determination unit 23 determines whether the peak frequency is generated by noise or an audio signal, and outputs an attenuation control signal to the power attenuating unit 18 for each band according to the result.

The noise suppressing operation of the first embodiment will be described below.

In the background noise canceling apparatus to which the present invention is applied, at least immediately after the start of the apparatus (this is referred to as "apparatus starting period"), there is only noise and no voice (this is referred to as "noise period"). Is present, and thereafter, a period in which a voice including noise is received and the noise is canceled (this period is referred to as a “voice period”) is continued. Of course,
A noise period may be inserted in the middle of the voice period. Further, when the background noise is noise of a narrow band component, this is called a "narrow band noise period".

First, the signal of the noise period is input terminal IN.
The following description will be made on the assumption that the data is input from the A / D converter 1 to the A / D converter 1. In the first embodiment, as a method of separating the frames of the input signal, one frame is composed of 256 samples of the digital signal, but this value is not limited to 256 samples if properly determined.

The output of the A / D converter 1 is input to the voice detector 8 and the window function calculator 2. The method of detecting the voice in the voice detector 8 may be any method as long as it is a method of distinguishing the presence or absence of voice, such as a method of using a time average for a long time or a short time.

The detection result determined by the voice detector 8 is input to the smoothing calculator 19 via the second counter 20 or directly, and the processing operation is controlled. Details of the processing operation of the smoothing calculator 19 will be described later.

In the FFT 3, the output of the window function calculator 2 is converted into a signal component in the frequency domain by a fast Fourier transform process. At this time, an imaginary number component and a real number component are generated as the frequency domain output of the FFT 3, and these signal components are output to the signal phase calculator 4 and the signal power calculator 14, respectively. The signal phase calculator 4 calculates and holds a value related to the phase of the obtained signal. The value related to this phase has the form required by IFFT5, such as cosine component = (real component) ² / (signal power P) ² sine component = (imaginary component) ² / (signal power P) ^2. It is preferable to store at.

The signal power calculator 14 calculates the signal power P, for example, the root mean square power, from the Fourier-transformed signal components in the frequency domain. That is, the power p (k) according to the following equation (1) can be obtained.

P (k) = {real number component (k) ² + imaginary number component (k) ² } ^1/2 (1) Since the switch 10 is closed in the initial state when the apparatus is started, the signal power calculation unit 14 The power p (k), which is the calculation result in the above, is output to the background noise updating unit 11 via this switch 10. As described above, the switch 10
Is controlled to be opened and closed by the voice detection result of the voice detector 8. When the detection result of the voice detector 8 is sound, that is, the switch 10 is opened during the voice period. Further, the switch 10 is controlled to be closed during a noise period when the detection result of the voice detector 8 is silent.

In the signal power calculation unit 14, (1)
Although the square root calculation shown in the formula is executed, p (k) ² may be input to the background noise updating unit 11 without calculating the square root.

FIG. 2 is a block diagram showing a concrete configuration of the background noise updating section 11. In this figure, 201,
Reference numeral 203 is a multiplier, 202 is an adder, and 204 is a delay register (D) constituting the background noise holder 12. The signal p (k) passed through the switch 10 is supplied to the background noise updating unit 11
Input to the multiplier 201. In addition, this multiplier 201
Is input with a coefficient value α that determines the averaging follow-up speed, and the multiplier 203 is input with a coefficient value (1−α).
As a result, as the output value of the adder 202, the power N ′ (k) of the estimated background noise expressed by the following equation (2)
Is required. Here, the variable k corresponds to the sampling point on the time axis, and N ′ (k) represents the kth update result.

N ′ (k) = α (p (k)) + (1−α) N ′ (k−1) (2) Further, α, which is a coefficient that determines the averaging follow-up speed, At the time of initial rising, or for a while from the time when the voice detection result was voiced at the end, a large value α
It is set to 1, for example 1.0, but is otherwise set to a small value α2, for example 0.1. The monitoring of this time is performed by the first counter 9, and the coefficient is set by the following procedure.

First, the first counter 9 counts the time t from the beginning of the operation of the device, or the time t of a certain period from the background noise state change. By counting the time t by the first counter 9, generally, the coefficient α is set to a large value within the range of α ≦ 1.0 in a period in which the preset threshold value t1, for example, 500 ms is not exceeded. . In this embodiment, the time threshold t1 is set to 64 ms, and at timings below this value,
α is set to 1.0.

Further, when the time exceeds t1, that is, t ≧ t1, a value α1 smaller than α described above is set as α. At this time, the relationship of α1 <α is established. In this embodiment, α1 = 0.1.

By setting α in this way, it is possible to improve the initial rising characteristics of the device quickly, eliminate the background noise disturbance that occurs immediately after the voice period, and improve the sound quality.

On the other hand, N '(k) which is the output of the background noise holder 12 is input to the-input terminal of the adder 15.
The signal power calculation unit 1 is connected to the + input terminal of the adder 15.
The power p (k) from 4 is supplied as an input, and the adder 15 performs subtraction between N '(k) of the background noise holder 12 and p (k) of the signal power calculator 14.
That is, the following expression (3) is executed.

S (k) = p (k) −N ′ (k) (3) The subtraction result s (k) calculated by the equation (3) is S
When (k) is the voice signal power, Ν (k) is the background noise power, and e (k) is the power of the residual noise, when the input signal to the input terminal IN is a mixed signal of the voice signal and the noise, , S (k) = S (k) + N (k) −N ′ (k) = S (k) + e (k) (4). Further, the subtraction result s (k) of the equation (3) becomes s (k) = N (k) −N ′ (k) = e (k) (5) when the input signal is only noise. . Therefore, if the estimated background noise power N ′ (k) is sufficiently converged, the actual background noise power N
Since it is equal to (k), the power of the residual noise e (k)
Becomes 0. If the input signal to the input terminal IN is a mixed signal of the audio signal and the noise, the subtraction result s (k) of the equation (3) is equal to the power of the signal obtained by removing the noise component from the input signal. .

Next, the operation of the frequency band divider 13 will be described.

The signal s (k) which has been subjected to the subtraction processing of the equation (3).
Is input from the adder 15 to the frequency band divider 13,
Band split. Here, the band division means that the audio signal to be subjected to signal processing is limited to some predetermined small regions represented on the frequency axis.
Therefore, the conventional R.S. J. It is not necessary to separately prepare frequency signals obtained by dividing all frequency components of the actual input signal into divided bands (hereinafter referred to as channels) as in the “noise suppression method” of Burmer et al.

Here, the entire frequency band (all channels) to be processed, for example, 250 to 3500 Hz is 2
It may be equally divided for every 50 Hz, or the range in which the frequencies included in the audio signal to be handled are concentrated may be divided into areas with emphasis, and the range of the remaining frequency region may be roughly divided.
In short, this dividing method may be any method as long as it is appropriately determined depending on the system to which it is applied.

In this embodiment, for example, as shown in FIG. 3, the band of 250 Hz to 3500 Hz is divided into 13 equal parts for every 250 Hz. The output from the background noise holder 12 is also input to the band divider 13. In this case, the background noise band division method is also the same as the division method for dividing the signal output from the adder 15.

The signal, the processing area of which has been determined by the band divider 13, is input to the logarithmic calculator 16. The logarithmic calculation unit 16 inputs the residual signal e (k) and the estimated background noise N ′.
(K) is logarithmically converted for each divided channel. This logarithmic conversion method may be a known method, for example, a method using a predetermined conversion table, or a method of obtaining logarithm by linear approximation or the like may be used. The signal subjected to logarithmic conversion is input to the s / n calculation unit 17.

The s / n calculator 17 is composed of an area s / n calculator for each channel, and based on the frequency signal input to the corresponding s / n calculator, the s / n estimated value for each channel is calculated. Is calculated as S / Ν (m) as follows. That is, if the signal power of the m-th channel resulting from the k-th update result is E (k, m) and its background noise power is Ν '(k, m), then E (k, m) = ΣE ( k, f) (6) N '(k, m) = ΣN' (k, f) (7) S / Ν (m) = logE (m) -logN '(m) (8). In these equations (6) and (7), Σ means addition in the frequency region in the range of the channel start frequency sf and the channel end frequency ef. For example,
Now, if the first channel is set to 250 Hz to 500 Hz, then sf = 250, ef = 500, and m = 1.

In this embodiment, the logarithmic calculation unit 1
6 is used to calculate s / n by subtraction in the logarithmic domain, but if there is no problem even with a simpler method, directly use the results of equations (6) and (7) to obtain E ( m) / N '
The division of (m) may be performed.

As described above, when there is a sound, s
Since (k) ≈S (K) and N ′ (k) = N (k), the above equation (8) is expressed as S / Ν (m) ≈logE (m) −IogN (m). (8 '). This is a logarithmic expression of voice signal power / estimated background noise power≈voice signal power / background noise power. From the above description, it can be seen that the s / n of each channel is correctly obtained as the S / N (m) in the s / n calculator 17 even if the input signal has background noise mixed in the audio signal. .

Next, the operation of the band-specific power attenuator 18 will be described.

The s / n estimated value of each channel calculated by the s / n calculator 17 is input to the band-specific power attenuator 18. The band power attenuator 18 determines the amount of attenuation of the signal output from the band divider 13 according to the output from the s / n calculator 17, that is, the s / n estimated value for each channel. Attenuating. A publicly known reference table may be used for each channel to determine the attenuation amount, or an evaluation formula is set, whereby the attenuation is given according to S / N (m) which is the s / n estimated value. It may be. This Embodiment 1
Then, when s / n is a variable x and the attenuation amount is a function value y, y = − (2/3) x + 20: x ≦ 30 (9) y = 0: x> 30 (10) I have decided. Therefore, (4),
The residual noise e (k) expressed by the equation (5) is attenuated according to s / n of each channel. As a result, only noise frequency channels with small s / n are further attenuated,
The sound quality is improved as a whole.

Next, the operation of the smoothing calculator 19 will be described.

The smoothing calculator 19 has a variable filter coefficient β.
Is composed of a plurality of smoothing filters (not shown). The smoothing calculator 19 includes a power attenuator 18 for each band.
Is input. In addition to this, the detection output from the voice detector 8 and the output signal from the second counter 20 are input. The detection output of the voice detector 8 is input to the second counter 20. The second counter 20 counts the elapsed time t3 after the initial start-up of the device or after the voice detector 8 finally determines that there is sound, that is, after the voice is cut off. In the second counter 20, the count result t3 is the threshold value n set in advance.
When it becomes larger than v3, a coefficient control signal is output to the smoothing calculator 19 to change the coefficient β of the smoothing filter to β1. At other timings, the coefficient β of the smoothing filter is set in the range of β ≦ 1.0, for example, β2 as below. When the detection result of the voice detector 8 is in the voice period, the coefficient β is set to β3 as below, for example. The relationship between these coefficient values β1, β2, and β3 is defined by the following equation (11).

1.0 ≧ β2 ≧ β3 ≧ β1 ≧ 0.0 (11) In this embodiment, nv3 = 48 ms, β2 = 1.0, β3 = 0.3, and β1 = 0.1. However, these β values may be set to those optimum for the configuration to which the system is applied, and are not limited to the numerical values of the embodiment.

The smoothing calculator 19 creates a smoothing signal sth (k) as follows and outputs it to the IFFT 5.

Sth (k) = β × e (f, k) + (1-β) × sth (k−1) (12) where β is a constant.

Β = 1.0: t3 ≦ nv3 (13) β = 0.1: t3> nv3 (14) β = 0.8: voice period In this way, the smoothing calculator 19 smoothes the signal as time averaged at each frequency. The signal sth (k) is output. The smoothed signal sth (k) is input to the IFFT5. IF
In FT5, a signal in the frequency domain is converted into a signal in the time domain, but after the signal is converted into the signal in the time domain, the output of the fast inverse Fourier transform is passed through D /
It is output to the A converter 7. The output of the fast inverse Fourier transform may be directly output to the D / A converter 7,
By overlapping the window function with a window function such as a hamming window, the connection of the output signals can be smoothed. In this embodiment, a well-known Hamming window function is used, and a window function overlap computing unit 6 in which window functions overlap by 50% is used. The output from the window function overlap calculator 6 is output to the D / A converter 7. D / A
The converter 7 converts the digital signal into an analog signal,
Output to the output terminal OUT.

Next, the narrow band noise determination operation will be described. The output of the adder 15 is input to the peak frequency search unit 21, and the peak frequency search unit 21 determines the frequency indicating the peak included in the signal in the frequency domain. Therefore, one peak detection method in the peak frequency search unit 21 will be described.

FIG. 3 is a diagram showing an example of frequency characteristics having a peak. The horizontal axis represents frequency and the vertical axis represents root mean square power.

First, the average value avl1 of power is obtained over the entire frequency range. The average value avl1 is compared with the power E (m) of each channel. Then, the peak detection threshold thpeek preset in the peak frequency search unit 21.
If the value obtained by adding the average value avl1 and the average value avl1 has the following relationship, it is determined that there is a peak, and the channel m is determined to be a channel including peak characteristics.

E (m) ≧ avl1 + thpeek channel with peak E (m) <avl1 + thpeek channel without peak FIG. 4 shows a second method for obtaining peak characteristics in the peak frequency search unit 21. In the second method, the difference code def (k) of the power f is obtained at a predetermined frequency interval from the following expression (15) over the entire frequency band. That is, def (k) = f (k + 1) -f (k) (15) Here, k is an integer and is an integer that maximizes {(the number of samples at the time of fast Fourier transform) -1}. As can be seen from FIG. 4, when the calculation by the equation (15) is performed from the direction of smaller frequency, def (k)> 0 is always satisfied while the power f is increased, and def (k) is satisfied when the power is decreased. k) <0. However, in the case of def (k) = 0, the same handling as in the case of increasing the power is performed.

Next, in order to calculate the peak, the peak code p is calculated based on the following equations (16-1) and (16-2).
Calculate eek (k).

That is, when def (k) ≧ 0, peak (k) = def (k) × def (k + 1) (16-1), and when def (k) <0, peak (k) = const : Const = positive constant (16-2).

By computing these equations (16-1) and (16-2), it is determined that there is no peak when peak (k) ≧ 0, and there is a peak when peak (K) <0. , The peak frequency is determined to be f (k). The above determination operation is performed inside the peak frequency search unit 21. FIG. 4 shows an example of the differential code def (k) and the peak code when fl and f2 have peak components. In this embodiment, the peak frequency search is executed by the second method, that is, the method based on the equations (16-1) and (16-2).

Next, a method of determining narrow band noise using the result of such a peak frequency search method will be described.

When the peak frequency search section 21 has analyzed the presence / absence of a peak component and the peak frequency, it outputs the result to the adjacent frequency power monitoring section 22. In the adjacent frequency power monitoring unit 22, the frequency detected by the peak frequency exploration unit 21 is centered on the upper and lower Δ values other than the peak frequency.
power of frequency f, eg, frequency of 20 Hz above and below
Calculate f3. The calculation result is further input to the narrow band noise determination unit 23. Then, the narrow band noise determination unit 2
3, the relationship between this power pf3 and the preset threshold th2 is used to determine whether or not the output signal s (k) of the adder 15 having the detected peak property contains narrow band noise as follows. Making a decision. Here, the following difference between the voice signal and the power component of the narrow band noise is utilized. That is, if it is the peak of the audio signal, the peak is gradual and there is some power component around the peak, but if it is the power component of the narrow band noise, the peak is steep and the peak frequency This is because there is no power as large as the power at the peak frequency in the vicinity.

Pf3 ≦ th2 narrow band noise (17) pf3> th2 audio signal (18) where th2 is a predetermined constant for narrow band detection, and in this embodiment, the power pf3 of the frequencies of 20 Hz above and below. According to the calculated value, th2 = pf3 / 2.0 is set. However, the value of th2 can be appropriately determined according to the signal to which the invention is applied and the circuit configuration, and is not limited to this numerical value.

When the condition of the expression (17) is detected, the narrow band noise judging section 23 sends a control signal to the power attenuating section 18 for each band so as to attenuate the power of the channel including the peak. Output. Power attenuator 18 for each band
Then, the power of the channel including this peak is selectively attenuated. Therefore, the power attenuator 18 for each band is
The output of is a narrowband signal, in particular, an output from which stationary narrowband noise with low power is removed in adjacent channels.

The output of the peak frequency search unit 21 is
It is also output to the peak frequency multiple relationship search unit 24. In the peak frequency multiple relation search unit 24, the peak frequency search unit 2
With respect to the peak frequency found in 1, it is detected whether or not a peak exists in a frequency component that is not in a multiple relationship.

FIG. 5 is a diagram showing the frequency characteristic of the input signal when narrow band noise having a non-stationary peak is mixed in the audio signal. FIG. 5 corresponds to the characteristic of the output signal s (k) of the adder 15 when the peak noise and the voice signal are input. The operation of the peak frequency multiple relationship search unit 24 will be described below.

In FIG. 5, the horizontal axis represents frequency. The peak frequency search unit 2 is included in the peak frequency multiple relation search unit 24.
It is input from 1 that the detection result has peak characteristics of frequencies fl to f7. Therefore, the peak frequency multiple relationship search unit 24 determines the respective peak frequencies fl to f7.
Then, a search is made as to whether or not the peak is detected as a frequency having a multiple relationship. As a search method, a remainder calculation is performed between detected frequencies. This will be described with reference to the example of FIG. 5. The frequencies other than f6 have a multiple relationship with each other. For example, for f1 and f2, mod (f2, f1) = 0. Here, mod (a, b) represents the remainder when a is divided by b.

In this way, the peak frequency multiple relation search unit 24 detects independent narrow band signals that are not in multiple relation with each other. The narrow band noise determination unit 23 determines that the detection result of the peak frequency multiple relation search unit 24 is mod (a,
A frequency in which b) does not become 0 is determined as narrow band noise. In the example of FIG. 5, the frequency of f6 is determined to be narrow band noise.
The determination result is output to the band-based power attenuating unit 18, and the band-based power attenuating unit 18 outputs the frequency f6 of the determination result described above.
Alternatively, the power of the channel including this frequency f6 is attenuated. In this way, it is possible to obtain an audio signal from which a narrow-band noise component that changes with time is removed. The attenuated signal is input to the smoothing calculator 19, and the subsequent processing is as described above.

Second Embodiment FIG. 6 is a block diagram showing a second embodiment of the present invention. The blocks corresponding to those in the first embodiment (FIG. 1) are designated by the same reference numerals, and their description will be omitted.

In the second embodiment, the frequency determined as narrow band noise by the narrow band noise determination section 23 is output to the band divider 13, and the band divider 13 further determines the narrow band as the determination result. The signal is output to the background noise holder 12, and the adder 15 subtracts the power of the narrow band noise frequency as background noise. That is, the background noise holder 12 regards the output narrow band signal as noise and outputs it to the adder 15. In the adder 15, a new background noise estimated value updated based on the narrow band signal output from the background noise holder 12 is subtracted from the signal power s (k). Therefore, the output of the adder 15 becomes a signal in which the power component of the narrow band noise is deleted. The output from the band divider 13 to the background noise holder 12 is output only when the narrow band noise determination unit 23 outputs.

Although the present invention has been described here as a background noise canceling device in the transmitter section of a telephone, it can be applied to a device involving voice input such as a voice input device and a recorder in addition to the telephone.

[0083]

Since the present invention is configured as described above, it is possible to provide a background noise canceling device having the following effects.

First, since the background noise power estimated and held by the background noise holder is always subtracted from the power calculated by the signal power calculator, the time which has been a problem in the conventional technique is reduced. Background noise with small fluctuations can be removed.

Secondly, since the filter coefficient of the smoothing calculator is made variable to make the initial rising response of the device agile,
Even if the background noise changes suddenly, the processed sound can be taken out stably and is not deteriorated.

Third, by calculating the multiple relationship of each peak frequency, even if the voice is continuously input, it is not erased, and therefore the call quality is excellent.

Fourth, since the adjacent power of the peak signal is monitored, even if the frequency characteristic of the voice signal is a component that does not have a large spread in frequency (such as a vowel), this is regarded as noise and forced. There is no need to erase it.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a block diagram showing a specific configuration of a background noise updating unit.

FIG. 3 is a diagram showing an example of frequency characteristics having a peak.

FIG. 4 is a diagram showing a second method of obtaining peak characteristics by a peak frequency search unit.

FIG. 5 is a diagram showing frequency characteristics of an input signal when narrow band noise having a non-stationary peak is mixed in a voice signal.

FIG. 6 is a block diagram showing a second embodiment of the present invention.

[Explanation of symbols]

1 analog / digital converter, 2 window function calculator,
3 fast Fourier transformer, 4 signal phase calculator, 5 fast inverse Fourier transformer, 6 window function overlap calculator,
7 Digital / analog converter.

Claims (11)

[Claims] 1. A background noise canceller for removing a noise component by subtracting in the frequency domain from an input signal in which the voice includes a noise component, and a voice detecting means for detecting the presence or absence of voice in the input signal. A signal power calculation means for calculating the power of the input signal, a background noise estimation holding means for estimating background noise for a predetermined time after the voice detection by the voice detecting means is finished, and holding the estimated value. A background noise canceling apparatus comprising: a calculation unit that subtracts the background noise power of the background noise estimation and holding unit from the signal power calculated by the signal power calculation unit and outputs the subtracted background noise power. 2. A smoothing calculating means for smoothing the signal power output from the calculating means is provided, and the background noise updating means causes the smoothing calculating means to perform an initial rise of the apparatus and a fixed time after the end of the voice period. 2. The background noise canceller according to claim 1, wherein the smoothing calculation in step 1 is performed steeply and the others are performed gently. 3. The background noise estimation and holding means includes a background noise holder that holds the estimated background noise, and a switch that stops updating the background noise according to the detection result of the voice detector. The background noise canceller according to claim 1. 4. The background noise canceller according to claim 2, wherein the background noise updating means changes the smoothing operation coefficient of the smoothing operation means in the voice period, the device starting period and the noise period, respectively. . 5. A band divider that divides the output of the arithmetic means into a plurality of predetermined bands, a logarithmic calculation unit that logarithmically converts the signal value of each band divided by the band divider, and the logarithm. An s / n calculator that calculates s / n for each region based on the logarithmic value calculated by the calculator, and a power attenuator that attenuates the power of each channel for each band according to the calculated s / n, The background noise canceller according to any one of claims 1 to 4, further comprising: a smoothing calculator for smoothing an output of the power attenuator, wherein the coefficient of the smoothing calculator is variable. 6. The background noise according to claim 5, wherein the smoothing process of the smoothing computing unit is performed sharply for the initial rise of the apparatus and for a fixed time after the end of the voice period, and is gently performed for others. Erase device. 7. A peak frequency exploring unit for determining a peak frequency of a narrow band signal, an adjacent frequency power monitoring unit for calculating power at an adjacent frequency of the peak frequency from the output of the peak frequency exploring unit, and the adjacent frequency power monitoring unit. And a narrow band noise judging section for judging narrow band noise from the narrow band noise judging section, and according to the judgment result of the narrow band noise judging section, a narrow band noise is removed by attenuating a component of a specific channel or a specific frequency. The background noise canceller according to any one of claims 1 to 6. 8. A peak frequency multiple relation search unit that calculates a multiple relation of each peak frequency based on the search result of the peak frequency search unit, an adjacent frequency power monitoring unit, and the peak frequency multiple relation search unit, The background noise canceller according to claim 7, further comprising a narrow band noise determination unit that determines narrow band noise. 9. The specific frequency is attenuated by reducing the gain using the gain created by the power attenuator in accordance with the result of the determination by the narrow band noise determiner. The background noise canceller according to claim 8. 10. The specific frequency component of the band divider is added to the background noise holder according to the determination result of the narrow band noise determination unit, and further subtracted from the input signal by the arithmetic means, 9. The background noise canceller according to claim 7, wherein the specific frequency is attenuated. 11. An A / D converter for converting the input signal into a digital signal, a fast Fourier transformer for converting the digital signal into a frequency domain signal by a fast Fourier transform process, and a frequency domain signal 11. A fast inverse Fourier transformer that reconverts the signal into a time domain signal, and a D / A converter that converts the reconverted digital signal into an analog signal. Background noise canceller.