JP4533126B2 - Proximity sound separation / collection method, proximity sound separation / collection device, proximity sound separation / collection program, recording medium - Google Patents

Description

  The present invention relates to a proximity sound separation and collection method that suppresses a noise signal and collects a target sound with a high S / N ratio in an environment in which a target sound source close to the microphone and a noise source far from the microphone are simultaneously sounding. The present invention relates to a sound separating and collecting apparatus.

In an environment where the target sound and noise are heard simultaneously, as a method of suppressing the noise and emphasizing the target sound, conventionally, a single microphone is used as the target sound and the time fluctuations such as air conditioning noise are slow. A spectral subtraction method (Non-patent Document 1) has been proposed in which a noise signal spectrum is subtracted from a spectrum of a mixed signal by using a noise noise (hereinafter referred to as stationary noise).
A microphone array method (Non-Patent Document 2) that suppresses noise using a plurality of microphones has also been proposed.
Boll, SF “Suppression of Acoustic Noise in Speech Using Spectral Subtraction.” IEEE Trans. Acoust. Speech, and Signal Processing, vol. ASSP-27, no.2 pp.133-120, 1979. Y. Kaneda and J. Ohga, “Adaptive microphone-array system for noise reduction,” IEEE Trans. Acoust. Speech Signal Process, vol.ASSP-34, no.6, pp.1391-1400, 1986

  Since the noise suppression method proposed in Non-Patent Document 1 uses noise continuity, a first problem that makes it difficult to suppress non-stationary noise signals such as speech and music occurs. Further, since the noise suppression method proposed in Non-Patent Document 2 requires at least two or more microphones, there arises a second problem that the scale of the apparatus becomes large.

  According to the first embodiment of the present invention, a band dividing unit that divides each output signal of the voice input unit into a plurality of band signals within a voice band, and a band-specific feature amount calculation unit that calculates an acoustic feature amount of the band signal. Based on the feature value for each band calculated by the feature value calculation unit for each band, it is determined for each band whether the signal is a signal mainly composed of the signal of the target sound source or a signal mainly composed of noise. Based on the determination result determined by the signal determination unit, the signal determination unit for each band, the weight value determination unit for each band for determining the weight value for each band, and the weight value determined by the weight value determination unit for each band It is characterized by comprising band-by-band weight value multiplying means for multiplying the band signal and signal synthesis means for returning the signal weighted by the band-by-band weight value multiplying means to a time waveform.

According to the second embodiment of the present invention, in the proximity sound separating and collecting apparatus proposed in the first embodiment, the band-specific feature amount calculating means calculates the power value of each band signal, and the band-specific signal determining means is The power value of each band signal is determined to be a band signal whose main component is a target sound signal when the power value is higher than a preset threshold value, and is determined to be a band signal whose main component is noise.
According to the third embodiment of the present invention, in the proximity sound separating and collecting apparatus proposed in the first embodiment, the band-specific feature value calculating unit calculates the sharpness as the feature value of each band signal, and determines the band signal. The means determines that the sharpness of each band signal is equal to or higher than a preset threshold value as a band signal whose main component is the target sound signal, and determines that the sharpness of each band signal is equal to or lower than the threshold value as a band signal whose main component is noise. It is characterized by that.

According to the fourth embodiment of the present invention, in any of the proximity sound separation and collection apparatuses proposed in the second embodiment or the third embodiment described above, the feature amount calculated by the band-wise feature amount calculation means is used. A threshold value calculation means for calculating a threshold value from the value is added, and the determination by the band-specific signal determination means is executed according to the threshold value calculated by the threshold value calculation means.
In each embodiment, the target sound is voice, and the target sound is used only under the condition that the target sound is closer to the microphone than the noise source. In addition, paying attention to the sparseness of the voice signal (the property that a frequency with a large power is localized in a specific band), the target voice is estimated from a signal mixed with noise.

In the band dividing means, the band is finely divided to the extent that the signal of each band is mainly composed of one acoustic signal component (to the extent that the spectrum of the target sound can be separated). As a specific example, the target sound is converted into noise. It is assumed that the target sound signal is received by the microphone larger than the noise signal because it is closer to the microphone.
The received sound signal is band-divided by the band dividing means, and the acoustic feature quantity for each band is calculated by the band-specific feature quantity calculating means. The weight value determining means for each band determines whether each frequency component is a component of a target sound source close to the microphone or a noise source component coming from a distance based on the feature amount calculated for each band. A weight value α (ω ₁ ) is determined based on the determination. For example, when the power of each band is used as the feature amount, the signal of the band in which the power is larger than a predetermined threshold is used as the target signal by using the fact that the signal power of the target sound source is larger than the signal power of noise. Determination is made, and a weight value to be multiplied by the band is determined to be, for example, α (ω ₁ ) = 1.0. The band where the power is smaller than the threshold is determined as a noise signal component, and the weight value α (ω ₁ ) (0 <α (ω ₁ ) <1) close to zero is determined.

In addition, when using the sharpness of a signal (detailed description will be described in the embodiment) as a feature amount, the sharpness of a nearby sound source is large and the sharpness of a distant sound source is small. If it is less than a certain threshold value, it is determined as a noise signal component, and a weight value close to zero is determined as, for example, α (ω ₁ ) (0 <α (ω ₁ ) <1).
The band-by-band weight value multiplication means multiplies each band signal X (ω ₁ ) by the determined weight value α (ω ₁ ). The weighted signal is returned to the time waveform by the signal synthesis means.

According to the configuration of the present invention, it is possible to recover the target speech from a signal mixed with noise without using the nature (stationarity) of noise. Therefore, it is possible to cope with a non-stationary signal such as voice or music as a noise source. That is, the first problem described above can be solved.
In addition, since the present invention can be realized with a single microphone, the apparatus scale can be reduced. Thereby, the second problem described above can also be solved.

The proximity sound separation and collection apparatus according to the present invention can be configured entirely by hardware. From this, a proximity sound separation and collection program written in a computer-readable program language is installed in the computer, and the computer is installed. The embodiment that functions as the proximity sound separation and collection device is the best embodiment.
When the computer is caused to function as a proximity sound separating and collecting apparatus according to the present invention, the computer includes a band dividing unit, a band-specific feature amount calculating unit, a band-specific signal determining unit, a band-specific weight value determining unit, a band-specific weight value multiplying unit, A signal synthesis means is constructed and functions as a proximity separation device.

FIG. 1 shows an embodiment of a proximity sound separating and collecting apparatus proposed in claim 5 of the present invention. The input means 1 is a microphone, for example. The signal of the target sound source M is S (t), and the signal of the noise source N is n (t). In order to simplify the description, the description will be made assuming that the noise source N is one, but in general, a plurality of noise sources N may be provided.
The band dividing means 2 divides the voice band into a plurality of bands by, for example, fast Fourier transform. At this time, each band signal X (ω ₁ ), X (ω ₂ ),. . . X (ω _N ) is subdivided so as to be mainly composed of one acoustic signal component. Here, one acoustic signal component refers to one spectrum included in the signals S (t) and n (t), and it is only necessary to divide each spectrum into fine parts that can be separated. (For further details, see Japanese Patent No. 3355598).

The band-specific feature value calculation means 3 calculates the acoustic feature value (τ (ω ₁ )) of the signal for each frequency band. This feature amount is, for example, signal power or sharpness. Here, description will be made assuming that the power of the signal proposed in claim 6 of the present invention is used as the feature amount. Therefore, the band-specific feature amount calculation means 3 uses the power values 20log ₁₀ | X (ω ₁ ) |, 20 log ₁₀ | X (ω ₂ ) of the band signals X (ω ₁ ), X (ω ₂ ),... X (ω _N ). ) |,... 20 log ₁₀ | X (ω _N ) |
The band-specific signal determination means 4 determines the attribute of each band signal X (ω ₁ ), X (ω ₂ ),... X (ω _N ) based on the power value of each band. Here, since noise comes from a distance from the target sound, it can be assumed that the noise signal n (t) is received smaller than the target sound signal S (t). That is, the band-divided band signals X (ω ₁ ), X (ω ₂ ),... X (ω _N ) are considered to have a spectrum as shown in FIG. Therefore, as shown in FIG. 2, it is estimated that the main component of the band where the power exceeds the threshold (T) is the target signal S (t), and the main component of the band below the threshold T is the noise signal n (t). It is estimated that. The band-specific signal determination means 4 applies this determination algorithm to determine the attributes of the respective band signals X (ω ₁ ), X (ω ₂ ),... X (ω _N ), and determines the determination result as a weight value for each band. Deliver to means 5.

The band-specific weight value determining means 5 determines the weight value α (ω _i ) as, for example, α (ω _i ) = 1.0 for the band determined as the target sound signal S (t). For the band determined as the noise signal n (t), the weight value α (ω _i ) is determined as 0 ≦ α (ω _i ) ≦ 1, for example. The weight value 0 ≦ α (ω _i ) ≦ 1 specified for the band determined to be noise is set to a value close to 0 as much as possible. The weight value α (ω _i ) = 1 designated for the band determined as the target sound signal does not necessarily have to be 1, as long as it is larger than the weight value given to the noise band.
The weight values α (ω _i ), α (ω ₂ ),... Α (ω _N ) of each band determined by the band-specific weight value determining means 5 are given to the band-specific weight value multiplying means 6, and this band-specific weight value Each band signal X (ω ₁ ), X (ω ₂ ),... X (ω _N ) is multiplied and weighted by the multiplication means 6 and each band signal α (ω _i ) · X (ω ₁ ), α (ω ₂ ) · X (ω ₂ ),... Α (ω _N ) · X (ω _N ) are input to the signal synthesizing means 7, and the signal synthesizing means 7 returns the time signal using, for example, inverse Fourier transform. Since a weight value close to 0 is designated for the band determined to be noise, the noise signal component contained in this time signal is small, and the SN ratio of the target sound signal S (t) is improved.

FIG. 3 shows an embodiment of the proximity sound separating and collecting apparatus proposed in claim 7 of the present invention. In this embodiment, the feature amount calculated by the feature amount calculation means 3 is shown as a sharpness J (ω ₁ ), J (ω ₂ ),... J (ω _N ).
Let y (n) be the linear prediction residual signal of the signal x (n). The sharpness (n) of the signal y (n) is defined by (1) below, E is the average value in parentheses
J (n) = E {y ⁴ (n)} / E ² {y ² (n)}-3 (1)
It is known that the sharpness of the signal y (n) has a large value in the case of a sound source signal close to the microphone and decreases as the distance from the microphone increases. Consider applying this property to a band signal X (ω ₁ ) obtained by band division. The sharpness of the band-divided band signal X (ω ₁ ) is measured, and when the sharpness of each band exceeds a predetermined threshold T, it is determined as a target sound signal, and the band below the threshold is a noise signal component Is determined. Here, in the case of the time waveform x (n), it is necessary to linearly predict the signal once, obtain the residual signal y (n), and measure the sharpness of the residual signal y (n). This is to remove the envelope information of speech by linear prediction. However, since the envelope information of the voice has not already been left in each band-divided component, in the present invention, the sharpness J ~ (ω _i , J) of the band-divided signal X (ω ₁ ) is expressed by Equation (2). Define and use it to determine the attributes of the signal components in each band.

J ~ (ω _i , _j ) = E {x ⁴ (ω _i , _j )} / E ² {x ² (ω _i , _j )}-3 (2)
Here, the index i is a band index, and j is a frame index.
The band-specific feature amount calculation means 3 calculates the sharpness J ~ (ω _i , _j ) defined by the equation (2) for each band. The band-specific signal determination unit 4 determines that a band having a sharpness equal to or higher than a threshold is a target sound signal component, and determines a band having a sharpness not higher than the threshold to be a noise signal.
Similarly to the case of FIG. 1, the band-specific weight value determining means 5 determines the weight value α (ω _i ) as α (ω _i ) = 1.0 for the band determined as the target sound signal component, and the noise signal component For the band determined to be, the weight value α (ω _i ) is determined as α (ω _i ) (0 ≦ α (ω _i ) ≦ 1). Each band signal X (ω _i ) is multiplied by the determined weight value α (ω _i ) of each band, and the weighted band signals α (ω _i ) · X (ω _i ) are time-signaled by the signal synthesis means 7. By returning to, the target sound signal from which the noise component has been removed can be obtained.

  By the way, in the above-described embodiment, the attribute of each band signal is determined using a predetermined threshold T for the determination by the band-specific signal determination means 4, but when this determination method is employed, noise is detected with respect to the target sound signal. This is effective when the signal is sufficiently small, but as the noise signal increases, the threshold T needs to be set larger. On the other hand, the sound signal generally has the property that the power of the signal decreases as the frequency becomes higher (see FIG. 5). Therefore, when the noise increases, it is necessary to set a large threshold value T in order to suppress the influence of the low frequency component of the noise, and as a result, there arises a problem that the high frequency component of the target sound signal is suppressed (FIG. 5). reference).

FIG. 4 shows an embodiment (corresponding to claim 8) for solving this problem. In this embodiment, a threshold value calculation means 8 for calculating an appropriate threshold value for each of a plurality of bands is provided, and using the threshold value calculated by the threshold value calculation means 8, the signal attribute determination means 4 for each band appropriately determines signal attributes. It is what.
That is, in this embodiment, the sound signal has several (usually about three) formant frequencies (see FIG. 5), and further, the power attenuates as the frequency increases. Even when the noise is large to some extent, the received signal s (t) + n (t) is separated into a plurality of, for example, about three bands, and the threshold calculation unit 8 calculates a threshold suitable for each band. The present invention can be applied. However, the condition that the power of the noise signal is smaller than the power of the target sound signal is necessary.

  A specific method will be described below. As shown in FIG. 5, the power of an audio signal is usually attenuated as it goes up. For this reason, when the noise is large to some extent, if an attempt is made to remove the noise component of the entire band with one threshold (T), the threshold T is set higher to remove the low frequency component of the noise signal. As a result, the target signal in the high range is attenuated. Therefore, the signal is divided into a plurality of (for example, three) bands, and threshold values (T1, T2, T3) suitable for each band are calculated by the threshold value calculation means 8. As a band dividing method, for example, the formant frequency (first formant frequency f1, second formant frequency f2, third formant frequency f3) of an average audio signal is used, and the band below f2 is set to the first band, and f2 to f3. The lower band is the second band, and the band above f3 is the third band.

A method of calculating threshold values (T1, T2, T3) in each band will be described by taking T1 as an example. In the first band, the frequency component X (ω _Max1 ) having the largest power among the received signals is selected. This band X (ω _Max1 ) can be determined to be highly likely to be a component of the target sound signal. Therefore, the power _{20log 10} | X (ω _Max1 ) | of X (ω _Max1 ) is calculated, and a value that is, for example, 20 dB smaller than the power value (other values (10 dB, 15 dB, etc.) may be used as the threshold value T1. That is, T1 = _{20log 10} | X (ω _Max1 ) | −20. In this way, a signal component that is 20 dB or more smaller than the frequency component having the maximum power in the first band is determined as a noise component and suppressed.

Similarly, for the threshold T2, the power of the frequency component 20log ₁₀ | X (ω _Max2 ) | having the largest power in the second band is calculated, and the threshold T2 is calculated as T2 = _{20log 10} | X (ω _Max1 ) |- Set to 20. The same applies to the threshold value T3. With the above method, the threshold value calculation means 8 obtains a threshold value suitable for each band. The calculation result is input to the band-specific signal determination means 4. The band-specific signal determination means 4 determines the attribute of each band signal using the threshold value calculated for each band. Therefore, even when the noise signal is increased to some extent, the noise component for each band is compared with the sixth aspect. Can be accurately determined. From the above description, it can be understood that a target signal can be extracted from a received sound signal in which a noise signal from a distance is mixed.

  The band dividing means 2, the band-specific feature amount calculating means 3, the band-specific signal determining means 4, the band-specific weight value determining means 5, the band-specific weight value multiplying means 6, the signal synthesizing means 7, the threshold values described in the above embodiments Each of the calculating means 8 can be made to function from the computer by installing a proximity sound distribution program described in a computer-readable program language in the computer, causing the CPU provided in the computer to decode and execute the program. It can function as a proximity sound separating and collecting device. The proximity sound separation and sound collection program is recorded on a recording medium such as a magnetic disk or CD-ROM that can be read by a computer, and can be installed in the computer from these recording media or can be installed through a communication line.

  The proximity sound separating and collecting apparatus according to the present invention is utilized in, for example, a hands-free audio conference system.

The block diagram for demonstrating the Example of the proximity sound separation sound collection apparatus proposed by Claim 5 and 6 of this invention. The graph for demonstrating the operation | movement of FIG. The block diagram for demonstrating the Example of the proximity sound isolation | separation sound collection apparatus proposed by Claim 7 of this invention. The block diagram for demonstrating the Example of the proximity sound isolation | separation sound collection apparatus proposed by Claim 8 of this invention. The graph for demonstrating the operation | movement of FIG.

Explanation of symbols

N Noise source 4 Band-specific signal determination means
M target sound source 5 band-specific weight value determining means n (t) noise signal 6 band-specific weight value multiplying means s (t) target sound signal 7 signal synthesizing means
1 Input means 8 Threshold value calculation means
2 Band division means
3 Band-specific feature value calculation means

Claims (6)

A proximity sound separation and collection method for collecting sound by suppressing noise in an environment where a target sound source and a noise source are present using at least one voice input means, and emphasizing the target sound signal,
Band division processing for dividing each output signal of the voice input means into a plurality of band signals within a voice band;
A band-specific feature amount calculation process for calculating a sound source feature amount of each band signal;
Based on the feature value for each band calculated in the feature value calculation process for each band, it is determined for each band whether the signal is mainly a signal of a target sound source or a signal of a noise source. Signal determination processing;
Based on the determination result determined in the signal determination process for each band, a weight value determination process for each band for determining a weight value for each band;
A band-by-band weight value multiplication process for multiplying each band signal by the weight value determined in the band-by-band weight value determination process;
A signal synthesis process for returning a signal weighted by the weight value multiplication process for each band to a time waveform;
Only including,
The band-specific feature amount calculation processing calculates the sharpness indicating the degree of peak in the time domain as the feature amount of each band signal, and the band-specific signal determination processing determines that the sharpness of each band signal is equal to or greater than a preset threshold value. A proximity sound separation and collection method, wherein the target sound signal is determined to be a band signal having a main component, and the sharpness of each band signal is determined to be a band signal having noise as a main component if the sharpness of each band signal is equal to or less than a threshold value . Oite proximity sound separation sound collection how according to claim 1, adding the threshold calculation process for calculating the threshold value from the value of the characteristic amount calculated in the band feature quantity calculation process, calculated by the threshold calculation process A proximity sound separation and collection method, wherein the band-specific signal determination processing is executed according to a threshold value. A proximity sound separating and collecting apparatus that suppresses noise in an environment where a target sound source and a noise source exist using at least one voice input unit and emphasizes the target sound signal to collect sound,
Band dividing means for dividing each output signal of the voice input means into a plurality of band signals within a voice band;
A band-specific feature amount calculating means for calculating an acoustic feature amount of the band signal;
Band-specific signal determination means for determining whether the signal is a signal mainly composed of the signal of the target sound source based on the characteristic value for each band calculated by the band-specific feature value calculation means;
Based on the determination result determined by the band-specific signal determination unit, a band-specific weight value determination unit that determines a weight value for each band;
Weight-by-band weight value multiplying means for multiplying each band signal by the weight value determined by the weight-by-band weight value determining means;
Signal synthesis means for returning the signal weighted by the weight-by-band weight value multiplication means to a time waveform;
In the near Se'on separating and collecting apparatus for Ru provided with,
The band-specific feature amount calculation means calculates the sharpness indicating the degree of peak in the time domain as the feature amount of each band signal, and the band-specific signal determination means determines that the sharpness of each band signal is greater than or equal to a preset threshold value. A proximity sound separating and collecting apparatus, wherein the target sound signal is determined to be a band signal having a main component, and the sharpness of each band signal is determined to be a band signal having noise as a main component if the threshold is not more than a threshold value. Oite proximity sound separation sound collection equipment according to claim 3, by adding a threshold value calculating means for calculating the threshold value from the value of the calculated features in the band feature quantity calculating means, calculated by the threshold calculating unit threshold The proximity sound separating and collecting apparatus according to claim 1, wherein the determination by the band-specific signal determining means is executed. 5. A proximity sound separation / collection program that is described in a computer-readable program language and causes the computer to function as at least the proximity sound separation / collection device according to claim 3 or 4 . 6. A recording medium comprising a computer-readable recording medium and recording the proximity sound separating and collecting program according to claim 5 on the recording medium.

Images