JP2014115377A - Sound processing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To highly accurately estimate a noise component of a sound signal even when the noise component fluctuates.SOLUTION: An index computation section 42 computes a confidence level CN(k,m) per unit period being an index of probability that a respective unit period of a sound signal corresponds to a noise component, in accordance with a disparity between a first index value P1(k,m) which follows time changes of power|X(k,m)|of each observation component X(k,m) of the sound signal containing a target sound component and the noise component, and a second index value P2(k,m) which follows the time changes of power|X(k,m)|at low followability compared to the first index value P1(k,m). An estimation processing section 44 generates an estimated noise component N(k,m) from the observation component X(k,m) by utilizing the confidence level CN(k,m) per unit period computed by the index computation section 42.

Description

  The present invention relates to a technique for processing an acoustic signal including a noise component, and is particularly preferably used for noise suppression in which the noise component is estimated and suppressed from the acoustic signal.

  Conventionally, a noise suppression technique for estimating a noise component included in an acoustic signal and suppressing it from the acoustic signal has been proposed. Various methods are used for estimating the noise component. For example, Non-Patent Document 1 discloses a technique (minimum statistical method) for estimating a noise component as a noise component in an acoustic signal.

Rainer Martin, "Spectral Subtraction Based on Minimum Statistics", Proc. EUSIPCO 94, p. 1182-1185

  However, in order to estimate the noise component with high accuracy under the technique of Non-Patent Document 1, it is necessary to search for the minimum value of power for a long section of about several seconds in the acoustic signal. Therefore, there is a problem that it is difficult to estimate the noise component after the change quickly and with high accuracy when the characteristic of the noise component changes. In view of the above circumstances, an object of the present invention is to estimate the noise component of an acoustic signal with high accuracy even when the noise component fluctuates.

  In order to solve the above-described problems, the acoustic processing device of the present invention has a first index value (for example, a first index value P1 (k, m) that follows a time change of an acoustic signal including a target sound component and a noise component. )) And the second index value (for example, the second index value P2 (k, m)) that follows the time change of the acoustic signal with low followability compared to the first index value. The reliability (for example, reliability CS (k, m) or reliability CN (k, m)) that is an index of accuracy corresponding to one of the target sound component and the noise component is calculated for each unit period. An index calculating means and an estimation processing means for generating an estimated noise component from the acoustic signal using the reliability of each unit period calculated by the index calculating means. In the above configuration, the estimated noise component is generated using the reliability of each unit period according to the difference between the first index value and the second index value that follow the time change of the acoustic signal with different tracking characteristics. . Therefore, even when the noise component fluctuates, the noise component of the acoustic signal can be estimated with high accuracy.

  In a preferred aspect of the present invention, the index calculation means uses a basic index (for example, reliability CS (k, m)), which is an index of accuracy that each unit period of the acoustic signal corresponds to the target sound component, as the first index value. The second index value is calculated for each unit period according to the difference, and the reliability is set to a predetermined value (for example, zero) for the unit period for which the basic index exceeds the first threshold value (for example, the threshold value TH1). In the above aspect, since the reliability is set to a predetermined value when the basic index exceeds the first threshold, the estimated noise component can be generated with high accuracy even when the target sound component is sufficiently dominant within the unit period. There are advantages. In another aspect, the index calculation means sets the reliability to a predetermined value (for example, zero) for a unit period in which the intensity of the acoustic signal is below a second threshold (for example, the threshold TH2). In the above aspect, since the reliability is set to a predetermined value when the intensity of the acoustic signal is lower than the second threshold value, for example, even when the intensity of the acoustic signal is instantaneously reduced, the estimated noise component can be generated with high accuracy. There is an advantage.

  In a preferred aspect of the present invention, the estimation processing means uses a smoothing coefficient (for example, the smoothing coefficient ω (k, m)) according to the reliability calculated by the index calculation means as an index of intensity of each unit period of the acoustic signal. An estimated noise component is generated by applying to the moving average. In the above aspect, since the estimated noise component is calculated by the exponential moving average of the intensity of each unit period of the acoustic signal, the non-patent for searching for the minimum value of the power for a long section of about several seconds in the acoustic signal. Compared with the technique of Reference 1, there is an advantage that the total number of unit periods (storage capacity necessary for holding the intensity of the acoustic signal) is reduced for the calculation of the estimated noise component. .

  In a preferred aspect of the present invention, the estimation processing means includes a smoothing coefficient (for example, a smoothing coefficient ω1 (k, m) corresponding to the reliability calculated by the index calculating means and the first coefficient (for example, the first coefficient A1). ) Is applied to the exponential moving average of the intensity of each unit period of the acoustic signal to generate the first estimated noise component (for example, the first estimated noise component Q1 (k, m)), and the reliability calculated by the index calculating means And a smoothing coefficient (for example, smoothing coefficient ω2 (k, m)) corresponding to a second coefficient different from the first coefficient (for example, the second coefficient A2) is an exponential moving average of the intensity of each unit period of the acoustic signal Is applied to generate a second estimated noise component (for example, a second estimated noise component Q2 (k, m)), and an estimated noise component is generated according to the first estimated noise component and the second estimated noise component. In the above aspect, the estimated noise component is generated according to the first estimated noise component and the second estimated noise component obtained by exponentially moving and averaging the intensity of the acoustic signal with different smoothing coefficients. The effect that it can be generated is particularly remarkable.

  The acoustic processing apparatus according to a preferred aspect of the present invention includes noise suppression means for suppressing the estimated noise component generated by the estimation processing means from the acoustic signal. According to the present invention, since the estimated noise component can be generated with high accuracy, the noise component of the acoustic signal can be suppressed with high accuracy.

  The acoustic processing device according to each of the above aspects is realized by hardware (electronic circuit) such as a DSP (Digital Signal Processor) dedicated to processing of an acoustic signal, or a general-purpose calculation such as a CPU (Central Processing Unit). This is also realized by cooperation between the processing device and the program. The program according to the present invention follows a first index value that follows a time change of an acoustic signal including a target sound component and a noise component, and follows a time change of the sound signal with lower followability than the first index value. An index calculation process for calculating, for each unit period, a reliability that is an index of accuracy in which each unit period of the acoustic signal corresponds to one of the target sound component and the noise component according to a difference from the second index value Using the reliability of each unit period calculated in the calculation process, the computer executes an estimation process for generating an estimated noise component from the acoustic signal. According to the above program, the same operation and effect as the sound processing apparatus according to the present invention are realized.

  The program of the present invention is provided in a form stored in a computer-readable recording medium and installed in the computer. The recording medium is, for example, a non-transitory recording medium, and an optical recording medium (optical disk) such as a CD-ROM is a good example, but a known arbitrary one such as a semiconductor recording medium or a magnetic recording medium This type of recording medium can be included. For example, the program of the present invention can be provided in the form of distribution via a communication network and installed in a computer.

  The present invention is also specified as an acoustic processing device (index calculation device) that calculates a reliability that is an accuracy index in which each unit period of an acoustic signal corresponds to one of a target sound component and a noise component. That is, the acoustic processing device according to another aspect of the present invention has a first index value that follows a time change of an acoustic signal that includes a target sound component and a noise component, and low followability compared to the first index value. In accordance with the difference from the second index value that follows the time change of the acoustic signal, the reliability that is an index of accuracy in which each unit period of the acoustic signal corresponds to one of the target sound component and the noise component is determined for each unit period. An index calculation means for calculating is provided. According to the above configuration, the reliability of each unit period of the acoustic signal can be calculated with high accuracy even when the noise component fluctuates.

1 is a block diagram of a sound processing apparatus according to a first embodiment of the present invention. It is a block diagram of a noise estimation part. It is explanatory drawing of the effect of 1st Embodiment. It is explanatory drawing of the effect of 1st Embodiment.

<First Embodiment>
FIG. 1 is a block diagram of a sound processing apparatus 100 according to the first embodiment of the present invention. As shown in FIG. 1, a signal supply device 12 and a sound emission device 14 are connected to the sound processing device 100 of the first embodiment. The signal supply device 12 supplies a time domain acoustic signal x (t) indicating the waveform of the mixed sound of the target sound component and the noise component to the acoustic processing device 100 (t: time). The target sound component is an acoustic component such as voice or musical sound, for example, and the noise component is an additive noise acoustic component typified by an environmental sound such as an operation sound of an air conditioner or a crowded noise in a crowd. A sound collection device that collects ambient sound and generates an acoustic signal x (t), or a playback device that acquires the acoustic signal x (t) from a portable or built-in recording medium and supplies it to the acoustic processing apparatus 100 Alternatively, a communication device that receives the acoustic signal x (t) from the communication network and supplies the acoustic signal x (t) to the acoustic processing device 100 may be employed as the signal supply device 12.

  The acoustic processing device 100 is a signal processing device (noise suppression device) that generates an acoustic signal y (t) in which a noise component is suppressed (a target sound component is emphasized) from the acoustic signal x (t) supplied by the signal supply device 12. is there. The sound emitting device 14 (for example, a speaker or a headphone) emits a sound wave corresponding to the acoustic signal y (t) generated by the acoustic processing device 100. Note that a D / A converter that converts the acoustic signal y (t) from digital to analog is not shown for convenience.

  As shown in FIG. 1, the sound processing device 100 is realized by a computer system including an arithmetic processing device 22 and a storage device 24. The storage device 24 stores a program executed by the arithmetic processing device 22 and various data used by the arithmetic processing device 22. A known recording medium such as a semiconductor recording medium or a magnetic recording medium or a combination of a plurality of types of recording media can be arbitrarily employed as the storage device 24. A configuration in which the acoustic signal x (t) is stored in the storage device 24 (therefore, the signal supply device 12 is omitted) is also suitable.

  The arithmetic processing unit 22 executes a program stored in the storage device 24 to thereby generate a plurality of functions (frequency analysis unit 32, noise estimation unit) for generating the acoustic signal y (t) from the acoustic signal x (t). 34, a noise suppression unit 36, and a waveform generation unit 38). A configuration in which each function of the arithmetic processing device 22 is distributed to a plurality of devices or a configuration in which a dedicated electronic circuit (for example, a DSP) realizes a part of the functions may be employed.

  The frequency analysis unit 32 converts a component (hereinafter referred to as “observation component”) X (k, m) of an acoustic signal x (t) corresponding to each of a plurality of frequencies on the frequency axis into a unit period (frame) on the time axis. It generates sequentially every time. The symbol k means any one frequency (frequency bin) on the frequency axis, and the symbol m means any one unit period on the time axis. For calculation of each observation component (frequency spectrum) X (k, m), a known frequency analysis such as short-time Fourier transform is arbitrarily employed. Note that it is also possible to use a series (filter bank) of a plurality of bandpass filters having different passbands as the frequency analysis unit 32.

  The noise estimation unit 34 sequentially calculates noise components (hereinafter referred to as “estimated noise components”) N (k, m) included in the acoustic signal x (t) (observed components X (k, m)) for each unit period. presume. The estimated noise component N (k, m) in the first embodiment corresponds to power (power spectrum). A specific configuration and operation of the noise estimation unit 34 will be described later.

The noise suppression unit 36 suppresses the estimated noise component N (k, m) estimated by the noise estimation unit 34 from the acoustic signal x (t) (observed component X (k, m)), thereby suppressing the acoustic signal y (t). Y (k, m) for each frequency (hereinafter referred to as “noise suppression component”) is sequentially generated for each unit period. Specifically, the noise suppression unit 36 of the first embodiment calculates the estimated noise component N (k, m) from the power (power spectrum) | X (k, m) | ² of the observed component X (k, m) in the frequency domain. The noise suppression component Y (k, m) is calculated by the following equation (1) that subtracts m) (spectrum subtraction).

The symbol max {} in Equation (1) means an operation for selecting the maximum value in parentheses. The symbol j in Equation (1) means an imaginary unit, and the symbol θ (k, m) means the phase angle (phase spectrum) of the acoustic signal x (t). As understood from the equation (1), the estimated noise component N (k, m) is calculated from the power | X (k, m) | ² of the observed component X (k, m) with a predetermined constant (flooring coefficient) β as a lower limit. The square root of the result of subtracting m) corresponds to the amplitude (amplitude spectrum) | Y (k, m) | of the noise suppression component Y (k, m).

  The waveform generation unit 38 in FIG. 1 generates a time-domain acoustic signal y (t) from the noise suppression component Y (k, m) that the noise suppression unit 36 sequentially generates for each unit period. Specifically, the waveform generation unit 38 converts the noise suppression component Y (k, m) of each unit period into the time domain by the short-time inverse Fourier transform and connects the preceding and subsequent unit periods to each other to connect the acoustic signal. Generate y (t). The acoustic signal y (t) generated by the waveform generator 38 is supplied to the sound emitting device 14 and radiated as a sound wave.

  FIG. 2 is a block diagram of the noise estimation unit 34. As shown in FIG. 2, the noise estimation unit 34 of the first embodiment includes an index calculation unit 42 and an estimation processing unit 44. The index calculation unit 42 calculates the reliability CN (k, m) of each frequency for each unit period of the acoustic signal x (t). The reliability CN (k, m) is the accuracy that the observed component X (k, m) of the acoustic signal x (t) corresponds to the noise component (the noise component is the target sound component in the observed component X (k, m)). It is an index of the degree of dominance in comparison, and is variably set within a range of 0 or more and 1 or less. As illustrated in FIG. 2, the index calculation unit 42 of the first embodiment includes a smoothing processing unit 422 and a calculation processing unit 424.

The smoothing processing unit 422 is a first index value that follows the time change of the acoustic signal x (t) (specifically, the power change of the observed component X (k, m) | X (k, m) | ² ). P1 (k, m) and the second index value P2 (k, m) are sequentially calculated for each unit period. Specifically, the smoothing processing unit 422 performs exponential movement of the power | X (k, m) | ² of the observed component X (k, m) as expressed by the following formulas (2A) and (2B): The first index value P1 (k, m) and the second index value P2 (k, m) are calculated by averaging (smoothing).

Symbol α1 in equation (2A) and symbol α2 in equation (2B) mean a smoothing coefficient (forgetting coefficient) of the exponential moving average, and are set within a range of 0 or more and 1 or less. As understood from the equations (2A) and (2B), the first index value P1 (k, m) and the second index value P2 (k, m) are the power | X of the observed component X (k, m). This corresponds to an envelope (power envelope) on the time axis of (k, m) | ² .

The smoothing coefficient α2 in the formula (2B) is lower than the smoothing coefficient α1 in the formula (2A) (0 ≦ α2 <α1 ≦ 1). Accordingly, the second index value P2 (k, m) is lower in tracking ability than the first index value P1 (k, m), and the power | X (k, m) | Follow the time change of ² . That is, the smoothing time constant τ2 according to Equation (2B) exceeds the smoothing time constant τ1 according to Equation (2A) (τ2> τ1). Note that the smoothing coefficient α1 may be set to 1 (that is, the power | X (k, m) | ² of the observed component X (k, m) is used as the first index value P1 (k, m)). Is possible.

  The calculation processing unit 424 in FIG. 2 uses the first index value P1 (k, m) and the second index value P2 (k, m) calculated by the smoothing processing unit 422 to determine the reliability (observation component X (k , m) is an accuracy corresponding to the noise component) CN (k, m). The calculation processing unit 424 of the first embodiment has a probability that the observed component X (k, m) corresponds to the target sound component (the target sound component is dominant in the observed component X (k, m) compared to the noise component. The reliability (basic index) CS (k, m) meaning a certain degree) is calculated according to the difference between the first index value P1 (k, m) and the second index value P2 (k, m). The reliability CN (k, m) is calculated by calculation using the degree CS (k, m).

Specifically, the calculation processing unit 424 calculates the reliability CS (k, m) by the calculation of the following formula (3).

That is, the calculation processing unit 424 determines a relative ratio (P1 (k, m) / P2 (k, m)) of the first index value P1 (k, m) to the second index value P2 (k, m) as a predetermined value. The reliability CS (k, m) is calculated within a range (0 ≦ CS (k, m) ≦ 1) within a value (1 in the example of the first embodiment). A configuration in which the difference between the first index value P1 (k, m) and the second index value P2 (k, m) is calculated as the reliability CS (k, m) may be employed.

  There is a general tendency that the target sound component such as voice or musical sound has a significant temporal variation compared to the noise component. Since the first index value P1 (k, m) follows the time change of the observed component X (k, m) with a higher followability than the second index value P2 (k, m), In the case where the target sound component is dominant in comparison with the noise component in the observed component X (k, m), the first index value P1 (k, m) is the second index value P2 (k, m). As a result, the reliability CS (k, m) of the observed component X (k, m) increases. Therefore, the reliability CS (k, m) can be used as an index of the accuracy with which the observed component X (k, m) corresponds to the target sound component.

The calculation processing unit 424 in FIG. 2 sets the reliability CS (k, m) calculated by the calculation of Expression (3) to a predetermined value (1 in the example of the first embodiment) as expressed by Expression (4) below. ) To calculate the reliability CN (k, m). That is, the reliability CN (k, m) is calculated so that the reliability CN (k, m) decreases as the reliability CS (k, m) increases.

As described above, the reliability CS (k, m) becomes a larger value within the range of 0 to 1 (the noise component is dominant) as the target sound component is dominant in the observed component X (k, m). As the reliability CS (k, m) becomes smaller, the noise component is more dominant in the observed component X (k, m) than the target sound component (observed component X (k, m)). The higher the accuracy corresponding to the noise component), the higher the reliability CN (k, m). Therefore, the reliability CN (k, m) can be used as an index of the accuracy with which the observed component X (k, m) corresponds to the noise component. As described above, the index calculation unit 42 in FIG. 2 differs between the first index value P1 (k, m) and the second index value P2 (k, m) (specifically, the relative ratio between the two). Accordingly, the reliability CN (k, m) of each frequency is sequentially calculated for each unit period. 2 uses the reliability CN (k, m) calculated by the index calculation unit 42 to sequentially calculate the estimated noise component N (k, m) of the acoustic signal x (t) for each unit period. To generate.

Specifically, the estimation processing unit 44 expresses the exponential moving average of the power | X (k, m) | ² of the observed component X (k, m) in each unit period as expressed by the following formula (5). Is calculated as an estimated noise component N (k, m). That is, the estimated noise component N (k, m) is an estimate of the power | X (k, m) | ² of the observed component X (k, m) in the current unit period and the past unit period (immediate unit period). In other words, the weighted sum with the noise component N (k, m-1).

The smoothing coefficient (weighted value) ω (k, m) in Expression (5) is set according to the reliability CN (k, m) calculated by the index calculation unit 42. Specifically, the smoothing coefficient ω (k, m) of the first embodiment is a product of the reliability CN (k, m) and a predetermined coefficient A as expressed by the following equation (6). It is.

The coefficient A is set to an appropriate numerical value within a range of 0 or more and 1 or less. Specifically, the coefficient A is variably set according to the continuity (degree of temporal variation) of the noise component assumed as a suppression target. For example, when a noise component with a relatively large temporal variation is assumed, the coefficient A is set to a value of about 0.2 to 0.5, and a noise component with a small temporal variation is assumed. The coefficient A is set to a value of about 0.02. It is also possible to variably set the coefficient A according to an instruction from the user.

The estimation processing unit 44 calculates the estimated noise component N (k, m) by executing the calculation of Expression (5) to which the reliability CN (k, m) calculated by the index calculation unit 42 for each unit period is applied. Updates every unit period according to the reliability CN (k, m). The initial value (N (k, 0)) of the estimated noise component N (k, m) is, for example, the observed component X (k over the M0 unit periods in the noise interval, as expressed by the following equation (7). , m) power | X (k, m) | ² is set to an average value (simple average).

The noise section is a section in which the target sound component is estimated not to exist in the acoustic signal x (t). For example, a section over a predetermined time length from the start point of noise suppression (for example, a section of 500 milliseconds from the start point of the acoustic signal x (t)) is suitable as the noise section.

As understood from the equations (5) and (6), the larger the reliability CN (k, m) (the larger the smoothing coefficient ω (k, m)), the observed component X (k , m) power | X (k, m) | ² (the degree to which the observed component X (k, m) is reflected in the estimated noise component N (k, m)) increases. That is, the observed component X (k, m) with the dominant noise component is preferentially reflected in the estimated noise component N (k, m) (the weight of the observed component X (k, m) with the dominant noise component (mixed) The estimated noise component N (k, m) is updated every unit period so that the ratio) increases. Therefore, according to the first embodiment, it is possible to estimate the noise component with high accuracy even when the noise component of the acoustic signal x (t) fluctuates with time. The noise suppression unit 36 in FIG. 1 suppresses the estimated noise component N (k, m) of each unit period estimated by the noise estimation unit 34 by the above procedure from the observed component X (k, m) of that unit period ( Formula (1)). Therefore, according to the first embodiment, there is an advantage that it is possible to generate an acoustic signal y (t) in which a noise component is suppressed with high accuracy (that is, an acoustic signal in which a target sound component is emphasized with high accuracy).

In a region F31 of FIG. 3, a spectrogram of the acoustic signal x (t) including only the noise component is shown. In the spectrogram of FIG. 3 and FIG. 4 to be described later, a point having a lower display gradation (a point close to black) means that the intensity is lower. In FIG. 3, it is assumed that one type of noise component starts at the origin on the time axis, and another type of noise component is added to the acoustic signal x (t) at time T0 when about 3 seconds have elapsed from the origin. ing. That is, the acoustic characteristic of the noise component of the acoustic signal x (t) varies at time T0. A region F32 in FIG. 3 shows a spectrogram of the estimated noise component N (k, m) estimated from the acoustic signal x (t) in the region F31 by the method of the first embodiment. Also, the power | X (k, m) of the observed component X (k, m) in the noise interval (interval of 500 milliseconds from the origin on the time axis) in the acoustic signal x (t) in the region F31 in FIG. The spectrogram of the estimated noise component when the average of ² is used as the noise component estimation result (hereinafter referred to as “proportional”) is shown in a region F33 of FIG. In contrast, the fluctuation of the noise component at time T0 is not reflected in the estimation result. On the other hand, according to the first embodiment, it is clear from FIG. 3 that the noise component of the acoustic signal x (t) is estimated with high accuracy, and the noise component after the fluctuation can be quickly estimated when the noise component fluctuates. To be grasped.

  In the region F41 of FIG. 4, a spectrogram of the acoustic signal x (t) in which the noise component exemplified in the region F31 of FIG. 3 is superimposed on the target sound component (speech) is shown. Also, in the area F42 in FIG. 4, the acoustic signal y ((), the estimated noise component N (k, m) (area F32 in FIG. 3) generated by the method of the first embodiment is suppressed from the acoustic signal x (t). A spectrogram of t) is illustrated, and a spectrogram after suppression when the estimated noise component of the region F33 of FIG. 3 is suppressed from the acoustic signal x (t) is illustrated in a region F43 of FIG. In contrast, in particular, the noise component after time T0 remains even after suppression, whereas according to the first embodiment, it can be confirmed from FIG. 4 that the noise component is well suppressed before and after time T0.

In the first embodiment, the smoothing coefficient ω (k, m) corresponding to the reliability CN (k, m) is used as the exponent of the power | X (k, m) | ² of the observed component X (k, m). The estimated noise component N (k, m) is calculated by applying to the moving average. Therefore, in comparison with the technique of Non-Patent Document 1 that searches for a minimum value of power for a long section of about several seconds in an acoustic signal, the power | for calculating the estimated noise component N (k, m) Advantage of reducing the total number of unit periods in which X (k, m) | ² should be held (the storage capacity necessary to hold the power | X (k, m) | ² of the observed component X (k, m) | There is.

Second Embodiment
A second embodiment of the present invention will be described below. In addition, about the element which an effect | action and function are the same as that of 1st Embodiment in each form illustrated below, the reference | standard referred by description of 1st Embodiment is diverted, and each detailed description is abbreviate | omitted suitably.

The index calculation unit 42 (calculation processing unit 424) of the second embodiment performs the calculation of the following equation (8) instead of the above-described equation (4), so that the reliability CS (k, m) The degree CN (k, m) is calculated.

  As understood from Equation (8), when the reliability CS (k, m) exceeds the threshold TH1 (CS (k, m)> TH1), the index calculation unit 42 calculates the reliability CN (k, m). Set to zero. The threshold value TH1 is close to a numerical value of the reliability CS (k, m) (for example, about 0.6 to 0.9) that can be evaluated that the target sound component in the observed component X (k, m) is sufficiently dominant. A positive number).

  When the target sound component is dominant in the observed component X (k, m) as the reliability CS (k, m) exceeds the threshold TH1 under the first embodiment (when the noise component is extremely small), The reliability CN (k, m) calculated by the equation (4) is set to a positive number (CN (k, m)> 0) although it is a sufficiently small value. Since the target sound component has a sufficiently high power (energy) compared to the noise component, even if the reliability CN (k, m) is set to a sufficiently small value, the purpose of the observed component X (k, m) The sound component is reflected in the estimated noise component N (k, m). Accordingly, there is a possibility that the estimation accuracy of the noise component is lowered. On the other hand, in the second embodiment, if the target sound component is dominant in the observed component X (k, m) as the reliability CS (k, m) exceeds the threshold TH1, the reliability CN (k, m) Is set to zero, and as a result, the observed component X (k, m) is not reflected in the estimated noise component N (k, m). Therefore, there is an advantage that the estimated noise component N (k, m) can be estimated with high accuracy even when the target sound component is sufficiently dominant in the observed component X (k, m).

Further, as understood from Equation (8), when the power | X (k, m) | ² of the observed component X (k, m) in the current unit period is lower than the threshold value TH2 (| X (k, m) Similarly, in ² <TH2), the index calculation unit 42 sets the reliability CN (k, m) to zero. The threshold TH2 is set to a numerical value of power | X (k, m) | ^{2 that} can be evaluated to be sufficiently small. Specifically, the threshold value TH2 is set to a multiplication value (and hence a variable value) of the predetermined coefficient λ and the estimated noise component N (k, m−1) of the immediately preceding unit period (TH2 = λ · N ( k, m-1)). The coefficient λ is set to a predetermined positive number (for example, a numerical value of 0.1 or less).

A noise component can be evaluated as a steady acoustic component when observed over a long period of time, but can be evaluated as an acoustic component that fluctuates rapidly in a short time when attention is focused on a short time of about a unit period. Therefore, there is a possibility that the power | X (k, m) | ^{2 of the} observed component X (k, m) may instantaneously decrease in a small number of unit periods even if the noise component can be evaluated to be stationary in the long term. is there. As described above, when the instantaneous decrease in the power | X (k, m) | ² is reflected in the estimated noise component N (k, m), the long-term estimated noise component N (k, m) is actually There is a possibility of deviating from the noise component. In the second embodiment, the reliability CN (k, m) is set to zero when the power | X (k, m) | ² of the observed component X (k, m) decreases as it falls below the threshold value TH2. As a result, the observed component X (k, m) is not reflected in the estimated noise component N (k, m). That is, the instantaneous decrease in the power | X (k, m) | ² of the observed component X (k, m) does not affect the estimated noise component N (k, m). Therefore, the estimated noise component N (k, m) is reduced as compared with the first embodiment in which the instantaneous power | X (k, m) | ² is also reflected in the estimated noise component N (k, m). The effect of being able to estimate with high accuracy is particularly remarkable.

As understood from the equation (8), the reliability CS (k, m) is equal to or less than the threshold TH1 (CS (k, m) ≦ TH1) and the power of the observed component X (k, m) | When X (k, m) | ² is equal to or less than the threshold value TH2 (| X (k, m) | ² ≦ TH2), the reliability CN (k, m) is the same as that in the first embodiment. It is set to a numerical value corresponding to CS (k, m).

<Third Embodiment>
The estimation processing unit 44 of the third embodiment sequentially calculates the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m) for each unit period. The first estimated noise component Q1 (k, m) is represented by the following expression (9A), and the power | X (of the observed component X (k, m) to which the smoothing coefficient ω1 (k, m) is applied. k, m) | is an exponential moving average of ² . Similarly, the second estimated noise component Q2 (k, m) is the power of the observed component X (k, m) to which the smoothing coefficient ω2 (k, m) is applied, as expressed by the following equation (9B). | X (k, m) | is an exponential moving average of ² .

The smoothing coefficient ω1 (k, m) of the equation (9A) is set according to the reliability CN (k, m) calculated by the index calculation unit 42 (calculation processing unit 424) and the predetermined first coefficient A1. The smoothing coefficient ω2 (k, m) in Equation (9B) is set according to the reliability CN (k, m) and the predetermined second coefficient A2. For example, the smoothing coefficient ω1 (k, m) is a multiplication value of the reliability CN (k, m) and the first coefficient A1 as in the following formula (10A), and the smoothing coefficient ω2 (k, m) ) Is a multiplication value of the reliability CN (k, m) and the second coefficient A2 as in the following formula (10B). The first coefficient A1 and the second coefficient A2 are set to different numerical values within a range of 0 or more and 1 or less (0 ≦ A1 <A2 ≦ 1).

  As described above, the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m) are generated according to the reliability CN (k, m). Similar to the noise component N (k, m), this corresponds to the estimation result of the noise component of the acoustic signal x (t). However, the first coefficient A1 is lower than the second coefficient A2 (A1 <A2). That is, the smoothing coefficient ω1 (k, m) is lower than the smoothing coefficient ω2 (k, m) (ω1 (k, m) <ω2 (k, m)). Therefore, the first estimated noise component Q1 (k, m) follows the time change of the noise component of the acoustic signal x (t) with lower followability than the second estimated noise component Q2 (k, m).

The estimation processing unit 44 of the third embodiment uses an estimated noise component N (k, m) corresponding to the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m) as a unit period. It generates sequentially every time. Specifically, as represented by the following mathematical expression (11), the estimation processing unit 44 calculates the minimum value of the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m). Is selected as the estimated noise component N (k, m).

In the third embodiment, the same effect as in the first embodiment is realized. In the third embodiment, the first estimated noise obtained by moving and averaging the power | X (k, m) | ² by applying different smoothing coefficients (ω1 (k, m), ω2 (k, m)). An estimated noise component N (k, m) is calculated according to the component Q1 (k, m) and the second estimated noise component Q2 (k, m). As assumed in the minimum statistical method of Non-Patent Document 1, the noise component tends to be lower in power than the target sound component, so the first estimated noise component Q1 (k, m) and the second estimated noise According to the third embodiment in which the minimum value of the component Q2 (k, m) is selected as the estimated noise component N (k, m), the noise of the acoustic signal x (t) is more accurately compared with the first embodiment. There is an advantage that the component can be estimated.

<Modification>
Each of the above-described embodiments can be variously modified. Specific modifications are exemplified below. Two or more aspects arbitrarily selected from the following examples can be appropriately combined.

(1) The method for calculating the estimated noise component N (k, m) according to the reliability CN (k, m) is arbitrary, and is not limited to the methods exemplified in the above embodiments. Other calculation methods of the estimated noise component N (k, m) are exemplified below.

[Example 1] The calculation method of the smoothing coefficient ω (k, m) applied to the calculation of the estimated noise component N (k, m) is arbitrary, and is not limited to the calculation of Equation (6). For example, a configuration in which the smoothing coefficient ω (k, m) is calculated according to the ρ power (ρ> 1) of the reliability CN (k, m) as in the following formula (12A) can be adopted. It is also possible to calculate the power of the coefficient A with the reciprocal of the reliability CN (k, m) as the power exponent as the smoothing coefficient ω (k, m) as in the following formula (12B).

  Regardless of which formula (12A) or formula (12B) is employed, the smoothing coefficient ω (k, m) increases as the reliability CN (k, m) increases, as in the above-described embodiment. That is, the observed component X (k, m) is set to a numerical value that increases the degree to which the estimated noise component N (k, m) is reflected. Note that the smoothing coefficient ω1 (k, m) and the smoothing coefficient ω2 (k, m) of the third embodiment are similar to the smoothing coefficient ω (k, m) of the first embodiment and the second embodiment. The calculation method is also arbitrary, and the same calculation as the formula (12A) and the formula (12B) can be adopted.

[Example 2] Formula (11) exemplified in the third embodiment can be replaced by the following formula (13).

That is, the minimum value (min {Q1 (k, m), Q2 (k, m)}) of the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m) sets the threshold σ1. In the unit period exceeding, the estimated noise component N (k, m) is set to the threshold σ1. That is, the estimated noise component N (k, m) is set with the threshold value σ1 as the upper limit value. The threshold σ1 is set to, for example, the power | X (k, m) | ² of the observed component X (k, m). In the configuration employing Equation (13), since the upper limit value of the estimated noise component N (k, m) is limited to the threshold value σ1, it is possible to reduce musical noise due to noise suppression in the frequency domain. .

[Example 3] In the third embodiment, the minimum value of the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m) is selected as the estimated noise component N (k, m). (Expression (11)), as expressed by the following expression (14), the maximum values of the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m) are estimated noise components. It is also possible to select as N (k, m).

Compared with the configuration in which the estimated noise component N (k, m) is calculated by the calculation of Equation (11), the configuration employing Equation (14) quickly follows the time variation of the noise component in the acoustic signal x (t). Thus, there is an advantage that the estimated noise component N (k, m) can be calculated.

It is also possible to replace Equation (14) with the following Equation (15).

That is, the maximum value (max {Q1 (k, m), Q2 (k, m)}) of the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m) sets the threshold σ2. In the unit period exceeding, the estimated noise component N (k, m) is set to the threshold σ2. That is, the estimated noise component N (k, m) is set with the threshold σ2 as the upper limit value. The threshold σ2 is set to the power | X (k, m) | ² of the observed component X (k, m), for example. In the configuration employing Equation (15), since the upper limit value of the estimated noise component N (k, m) is limited to the threshold σ2, the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 ( Even when the maximum value of k, m) becomes an extremely large value due to erroneous estimation or the like, it is possible to suppress the estimated noise component N (k, m) within an appropriate range.

[Example 4] In Example 2 and Example 3 of the third embodiment and this modification, one of the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m) is selected. It is also possible to calculate the estimated noise component N (k, m) in consideration of both the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m). For example, a configuration is preferably employed in which a weighted sum of the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m) is calculated as the estimated noise component N (k, m). Each weight value of the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m) is variably set according to an instruction from the user, for example.

[Example 5] The estimated noise component N (k, m) can be calculated by the following equation (16) using the estimated prior SNRξ (m).

That is, the minimum value (min {Q1 (k, m), Q2 (k, m)}) of the first estimated noise component Q1 (k, m) and the second estimated noise component Q2 (k, m) and the observed component X The estimated noise component N (k, m) is calculated by performing weighted averaging of the power | X (k, m) | ^{2 of} (k, m) with a weighting value corresponding to the estimated prior SNRξ (m). The coefficient η is a variable for controlling the contribution of the estimated prior SNRξ (m).

The estimated a priori SNRξ (m) is calculated by, for example, the following formula (17A) to which the estimated posterior SNRγ (k, m) is applied, and the estimated posterior SNRγ (k, m) of the formula (17A) is, for example, ). The symbol max {, 0} in Equation (17A) means an operation that limits the operation result to a positive number range. The calculation of subtracting 1 from the estimated posterior SNRγ (k, m) (average value over K frequencies) in the equation (17A) corresponds to the calculation for calculating the estimated prior SNR.

The symbol ε in the equation (17A) is set to a predetermined numerical value. Note that since the logarithm of zero is negative infinity, when the numerical value ε is set to zero, the calculation (log ₁₀ ε) of Expression (17A) may become unstable. Therefore, the numerical value ε is set to a very small positive number (ε> 0).

[Example 6] In each embodiment described above, the power | X (k, m) of the observed component X (k, m) to which the smoothing coefficient ω (k, m) corresponding to the reliability CN (k, m) is applied. The exponential moving average of ² was calculated as the estimated noise component N (k, m). As expressed by the following equation (18), the power of the observed component X (k, m) | X (k, m) It is also possible to calculate the estimated noise component N (k, m) with a weighted moving average of ² .

As understood from the equation (18), the power | X (k, m) | ² of the observed component X (k, m) of each of the M unit periods with the m-th unit period as the tail, Weighted relative ratio (CN (k, m) / ΣCN (k, mn)) of reliability CN (k, m) of unit period to total value of reliability CN (k, m) over M unit periods The estimated noise component N (k, m) is calculated by weighted average (weighted addition) as a value.

  As understood from the above examples, the estimation processing unit 44 estimates from the acoustic signal x (t) (observed component X (k, m)) using the reliability CN (k, m) of each unit period. Concretely included as an element (estimation processing means) for generating the noise component N (k, m), and generating the estimated noise component N (k, m) from the reliability CN (k, m) and the acoustic signal x (t) It doesn't matter what the method is.

(2) In the above-described embodiments, the reliability CN (k, m), which is an index of the accuracy that the observed component X (k, m) corresponds to the noise component, is applied to the calculation of the estimated noise component N (k, m). However, it is also possible to apply the reliability CS (k, m), which is an index of the accuracy that the observed component X (k, m) corresponds to the target sound component, to calculate the estimated noise component N (k, m). is there. For example, the mathematical formulas (5) and (6) described above can be replaced with the following mathematical formulas (5A) and (6A). Calculation of the reliability CN (k, m) is omitted.

In the configuration using Equation (5A) and Equation (6A), the reliability CS (k, m) is small (the noise component is dominant in the observed component X (k, m) compared to the target sound component). As the smoothing coefficient ω (k, m) becomes smaller, the influence of the power | X (k, m) | ² of the observed component X (k, m) in the current unit period increases. That is, the relationship between the superiority or inferiority of the noise component in the observed component X (k, m) and the degree to which the observed component X (k, m) is reflected in the estimated noise component N (k, m) is the same as that in the first embodiment. It is controlled similarly. As understood from the above description, the index calculation unit 42 in each of the above-described embodiments is a reliability (an index of accuracy in which each unit period of the acoustic signal x (t) corresponds to one of the target sound component and the noise component ( CS (k, m), CN (k, m)) is included as an element (index calculation means) for calculating each unit period.

(3) In each of the above-described embodiments, the exponential moving average of the power | X (k, m) | ² of the observed component X (k, m) is expressed as the first index value P1 (k, m) and the second index value P2 ( k, m) (Formula (2A), Formula (2B)), but the method of calculating the first index value P1 (k, m) and the second index value P2 (k, m) has been changed as appropriate. The For example, the first index value P1 (k, m) is a simple moving average of the power | X (k, m) | ² of the observed component X (k, m) as in the following equations (19A) and (19B). It is also possible to calculate the second index value P2 (k, m).

As understood from the equation (19A), the first index value P1 (k, m) is the observed component X over M1 (M1 is a natural number of 1 or more) unit periods that end with the mth unit period. This is the average of the power | X (k, m) | ² of (k, m). Further, as understood from the equation (19B), the second index value P2 (k, m) is the observed component X (k, m) over M2 unit periods with the mth unit period at the end. Average of power | X (k, m) | ² . The average number M2 of the formula (19B) exceeds the average number M1 of the formula (19A) (M2> M1). Therefore, the smoothing time constant τ2 according to Expression (19B) exceeds the smoothing time constant τ1 according to Expression (19A) (τ2> τ1). That is, as in the first embodiment, the second index value P2 (k, m) is lower in followability than the first index value P1 (k, m) and the power of the observed component X (k, m). Follow the time change of | X (k, m) | ² . Note that although the simple moving average is exemplified in the equations (19A) and (19B), the weighted moving average of the power | X (k, m) | ² over a plurality of unit sections is used as the first index value P1 (k, m). It is also possible to calculate as the second index value P2 (k, m).

As understood from the above description, the first index value P1 (k, m) is the time change of the acoustic signal x (t) (the power | X (k, m) | ^{2 of the} observed component X (k, m). The second index value P2 (k, m) is lower than the first index value P1 (k, m) and the time of the acoustic signal x (t) is lower than the first index value P1 (k, m). It is included as a numerical value that follows changes.

(4) In each of the above-described embodiments, the noise component is suppressed by subtracting the estimated noise component N (k, m) from the observed component X (k, m) in the frequency domain, but the estimated noise component N (k, m The content of the process of suppressing the noise component from the acoustic signal x (t) using m) is appropriately changed. For example, the present invention is also applied to multiplicative noise suppression that suppresses a noise component by multiplying an observed component X (k, m) by an adjustment value (spectral gain) corresponding to the estimated noise component N (k, m). Can be done. Specifically, noise suppression using a Wiener filter, noise suppression using the MMSE-STSA method or MMSE-LSA method, or any noise suppression technique represented by speech enhancement using MAP estimation, etc. It is possible to apply the estimated noise component N (k, m) calculated in each form.

(5) In each of the above-described embodiments, calculation and suppression of the estimated noise component N (k, m) are performed in parallel for each of a plurality of observed components X (k, m) obtained by dividing the acoustic signal x (t) on the frequency axis. However, calculation and suppression of the estimated noise component N (k, m) for each frequency are not essential requirements for the present invention. For example, a configuration for calculating and suppressing an estimated noise component N (k, m) for any one of a plurality of observed components X (k, m) obtained by dividing the acoustic signal x (t) on the frequency axis, A configuration that calculates and suppresses the estimated noise component N (m) without dividing x (t) into a plurality of observed components X (k, m) (that is, collectively over the entire band to be processed) is also employed. obtain.

(6) The generation and suppression of the estimated noise component N (k, m) exemplified in the above embodiments can be applied to other known acoustic processing techniques. For example, by generating and suppressing the estimated noise component N (k, m) for the acoustic signals x (t) of each channel generated by a plurality of sound collection devices (microphone arrays) in the same manner as the above-described embodiments. An acoustic signal y (t) is generated for each channel, and is adjusted to a known sound source direction by directivity control processing (for example, beam formation of a delay addition type or blind spot control type) using the acoustic signal y (t) of each channel. It is also possible to form a sound beam (region with high sound collection sensitivity).

(7) The power | X (k, m) | ² of the acoustic signal x (t) (observed component X (k, m)) applied to the various computations exemplified in the above embodiments is expressed as the amplitude | X (k , m) | and amplitude | X (k, m) | can be replaced with a power (for example, fourth power or half power). The power of the amplitude | X (k, m) | of the acoustic signal x (t) is included as the intensity of the acoustic signal x (t). The power | X (k, m) | ² and the amplitude | X (k, m) | of the acoustic signal x (t) are typical examples of the intensity of the acoustic signal x (t).

(8) In each of the above-described embodiments, the acoustic processing device 100 (noise suppression device) that executes both generation and suppression of the estimated noise component N (k, m) has been illustrated, but the estimated noise is calculated from the acoustic signal x (t). The present invention can also be applied as an acoustic processing device (noise estimation device) for calculating the component N (k, m). That is, the noise suppression unit 36 in FIG. 1 can be omitted.

  In addition, reliability CS (k, m), which is an index of the accuracy with which each unit period of the acoustic signal x (t) corresponds to the target sound component, or each unit period of the acoustic signal x (t) corresponds to the noise component. The present invention can also be applied to an acoustic processing device (index calculation device) that calculates the reliability CN (k, m) that is an index of the accuracy to be performed. A typical example of the method using the reliability CS (k, m) and the reliability CN (k, m) is the estimation of the noise component exemplified in the above-described embodiments (calculation of the estimated noise component N (k, m)). However, the reliability CS (k, m) or the reliability CN (k, m) can be used for applications other than the estimation of the noise component. Specifically, a configuration is assumed in which the acoustic signal x (t) is distinguished into a noise interval and a target sound interval on the time axis according to the reliability CS (k, m) or the reliability CN (k, m). The For example, a configuration in which a unit period in which the average value of reliability CN (k, m) over a plurality of frequencies exceeds a predetermined threshold is selected as a noise interval, or the average value of reliability CS (k, m) over a plurality of frequencies is A configuration in which a unit period exceeding a predetermined threshold is selected as a target sound section is preferably employed. As can be understood from the above description, the generation of the estimated noise component N (k, m) by the estimation processing unit 44 can be omitted.

(9) The sound processing device 100 (or noise suppression device) can be realized by a server device that communicates with a terminal device such as a mobile phone. For example, the acoustic processing device 100 generates an acoustic signal y (t) from the acoustic signal x (t) received from the terminal device and transmits the acoustic signal y (t) to the terminal device. In the configuration in which the acoustic processing device 100 receives each observation component X (k, m) of the acoustic signal x (t) from the terminal device (a configuration in which the terminal device includes the frequency analysis unit 32), the frequency analysis unit 32 is omitted. In the configuration in which each noise suppression component Y (k, m) is transmitted from the sound processing apparatus 100 to the terminal device (the configuration in which the terminal device includes the waveform generation unit 38), the waveform generation unit 38 is omitted. In addition, the acoustic processing device 100 generates an estimated noise component N (k, m) from the acoustic signal x (t) received from the terminal device and transmits the estimated noise component N (k, m) to the terminal device (for example, the terminal device includes the noise suppression unit 36). In the configuration, the noise suppression unit 36 can be omitted.

DESCRIPTION OF SYMBOLS 100 ... Acoustic processing device, 12 ... Signal supply device, 14 ... Sound emission device, 22 ... Arithmetic processing device, 24 ... Memory | storage device, 32 ... Frequency analysis part, 34 ... Noise estimation part, 36 ... ... noise suppression unit, 38 ... waveform generation unit, 42 ... index calculation unit, 422 ... smoothing processing unit, 424 ... calculation processing unit, 44 ... estimation processing unit.

Claims (5)

A first index value that follows the time change of the acoustic signal including the target sound component and the noise component, and a second index value that follows the time change of the sound signal with lower followability than the first index value. In accordance with the difference, the index calculation means for calculating, for each unit period, reliability, which is an accuracy index corresponding to one of the target sound component and the noise component, in each unit period of the acoustic signal,
An acoustic processing apparatus comprising: an estimation processing unit that generates an estimated noise component from the acoustic signal using the reliability of each unit period calculated by the index calculation unit. The index calculating means determines a basic index, which is an index of accuracy that each unit period of the acoustic signal corresponds to the target sound component, for each unit period according to a difference between the first index value and the second index value. The reliability is set to a predetermined value for a unit period in which the basic index exceeds the first threshold or a unit period in which the intensity of the acoustic signal is lower than the second threshold, while other unit periods are set as the basic index. The sound processing apparatus according to claim 1, wherein the reliability is calculated accordingly. 2. The estimation processing means generates the estimated noise component by applying a smoothing coefficient corresponding to the reliability calculated by the index calculation means to an exponential moving average of the intensity of each unit period of the acoustic signal. Or the acoustic processing apparatus of Claim 2. The estimation processing means applies a smoothing coefficient corresponding to the reliability and the first coefficient calculated by the index calculation means to the exponential moving average of the intensity of each unit period of the acoustic signal, thereby to obtain a first estimated noise component. And applying a smoothing coefficient according to the reliability calculated by the index calculation means and the second coefficient different from the first coefficient to the exponential moving average of the intensity of each unit period of the acoustic signal. The acoustic processing apparatus according to any one of claims 1 to 3, wherein a second estimated noise component is generated and the estimated noise component is generated according to the first estimated noise component and the second estimated noise component. The acoustic processing apparatus according to claim 1, further comprising: a noise suppression unit that suppresses an estimated noise component generated by the estimation processing unit from the acoustic signal.