JP2013068919A - Device for setting coefficient for noise suppression and noise suppression device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To achieve noise suppression which allows effective suppression of generation of musical noise.SOLUTION: A characteristic value calculation unit 42 calculates a shape parameter α0 according with a shape of an intensity distribution of a sound signal x(t). A first coefficient setting unit 44 variably sets a flooring coefficient η in accordance with the shape coefficient α0 calculated by the characteristic value calculation unit 42, so as to satisfy a relation formula prescribing a relation between the shape parameter α0 of the sound signal x(t) in the case where a kurtosis K of the intensity distribution of the sound signal x(t) is not changed before and after noise suppression processing and the flooring coefficient η to be applied to flooring processing of the noise suppression processing. A noise suppression unit 36 executes, on the sound signal x(t), noise suppression processing including flooring processing to which the flooring coefficient η set by the first coefficient setting unit 44 is applied.

Description

  The present invention relates to a technique for suppressing a noise component from an acoustic signal.

  In noise suppression technology that suppresses noise components from acoustic signals in the frequency domain, the generation of annoying musical noise caused by noise suppression processing becomes a problem. In Patent Document 1, a noise index value corresponding to the kurtosis of the intensity distribution of an acoustic signal is calculated as an index of the degree of occurrence of musical noise, and a suppression coefficient or flooring coefficient applied to noise suppression processing is used as the noise index value. A technique for variably setting according to this is disclosed. Non-Patent Document 1 discloses that the kurtosis ratio of the intensity distribution of an acoustic signal before and after the noise suppression process and the noise suppression rate by the noise suppression process can be formulated.

JP 2010-020012 A

T. Inoue, et al., "Theoretical analysis of iterative weak spectral subtraction via higher-order statistics", Proc. MLSP2010, p.220-225, 2010

  However, even under the techniques of Patent Document 1 and Non-Patent Document 1, it is not always easy to select an optimum coefficient that can effectively suppress the occurrence of musical noise. In view of the above circumstances, an object of the present invention is to realize noise suppression that can effectively suppress the generation of musical noise.

  The noise suppression coefficient setting apparatus according to the present invention includes a characteristic value calculating means for calculating a noise characteristic value corresponding to the shape of the intensity distribution of an input acoustic signal (for example, the acoustic signal x (t)), and a peak of the intensity distribution of the acoustic signal. When the degree of noise does not change before and after the noise suppression process, the noise characteristic value (for example, shape parameter α0) of the acoustic signal and the flooring coefficient (for example, the flooring coefficient η) applied to the flooring process of the noise suppression process First coefficient setting means for variably setting the flooring coefficient in accordance with the noise characteristic value calculated by the characteristic value calculation means so as to satisfy a relational expression that defines the relationship (for example, Expression (24)).

  In the above configuration, the flooring coefficient is variably set according to the noise characteristic value so as to satisfy the condition (first condition) that the kurtosis of the intensity distribution does not change before and after the noise suppression process. Therefore, it is possible to suppress the noise component while effectively suppressing (ideally completely preventing) the generation of musical noise.

  In a preferred aspect of the present invention, the first coefficient setting means sets the flooring coefficient so that the noise suppression rate by the noise suppression process is a positive number. In the above aspect, the flooring coefficient is set so as to satisfy the condition that the noise suppression rate is a positive number (second condition). For example, a flooring coefficient with a positive noise suppression rate is selected from a plurality of flooring coefficients that satisfy the relational expression between the noise characteristic value and the flooring coefficient. Therefore, it is possible to apply a significant flooring coefficient that can effectively realize noise suppression to the noise suppression processing.

  A noise suppression coefficient setting device according to a preferred aspect of the present invention variably sets a suppression coefficient (for example, suppression coefficient β) for controlling the noise suppression strength according to the noise characteristic value calculated by the characteristic value calculation means. Second coefficient setting means is provided. For example, a suppression coefficient that maximizes the noise suppression rate by the noise suppression process or a suppression coefficient that exceeds the target value is set. In the above aspect, since the suppression coefficient applied to the noise suppression process (subtraction process) to control the noise suppression intensity is variably set according to the noise characteristic value of the input acoustic signal, the suppression coefficient is fixed to a predetermined value. Compared with the configuration described above, appropriate noise suppression is realized. In addition, the specific example of the above aspect is later mentioned as 2nd Embodiment, for example.

  The noise suppression coefficient setting device according to a preferred aspect of the present invention sets the noise characteristic value calculated by the characteristic value calculation means and the first coefficient setting so that the accumulated value of the noise suppression rate by each noise suppression process exceeds the target value. The number of iterations setting means for variably setting the number of iterations of noise suppression processing (for example, the number of iterations Q) according to the flooring coefficient set by the means is provided. In the above aspect, since the number of iterations of the noise suppression process is variably set so that the accumulated value of the noise suppression rate exceeds the target value, the noise suppression processing is compared with a configuration in which the number of iterations is fixed to a predetermined value. It is possible to suppress excess and deficiency. In addition, the specific example of the above aspect is later mentioned as 3rd Embodiment, for example. The noise suppression process is repeated cumulatively for the input acoustic signal. The cumulative repetition of the noise suppression process is to repeat the noise suppression process for one section of the acoustic signal (that is, the acoustic signal after the execution of each noise suppression process is the target of the next noise suppression process). Means that.

  The present invention is also realized as a noise suppression device including the noise suppression coefficient setting device according to each of the above aspects. The noise suppression device according to the first aspect of the present invention performs, on an input acoustic signal, a noise suppression coefficient setting device according to each of the above aspects and noise suppression processing to which the coefficient set by the noise suppression coefficient setting device is applied. Noise suppression means to be executed. The noise suppression means includes, for example, an element that executes noise suppression processing including flooring processing to which the flooring coefficient set by the first coefficient setting means is applied to the input acoustic signal, and the floor set by the first coefficient setting means. An element for executing noise suppression processing applied to the input acoustic signal by applying the ring coefficient and the suppression coefficient set by the second coefficient setting means, and noise suppression processing for the input acoustic signal over the number of iterations set by the iteration count setting means It is an element that repeats cumulatively. According to the above noise suppression apparatus, the above-described operation and effects of the noise suppression coefficient setting apparatus of the present invention are realized.

The noise suppression apparatus according to the second aspect of the present invention is a Q-stage unit process that sequentially processes acoustic signals of D (D is a natural number of 2 or more) channels generated by sound collection devices arranged apart from each other. Means,
Output processing means for emphasizing a target sound component in a specific sound source direction by delay addition of the D-channel acoustic signal processed by the final stage unit processing means in the Q stage, and each of the Q stage unit processing means The noise estimation means for generating an estimated noise component by independent component analysis for the D channel acoustic signal supplied to the unit processing means, and the flooring coefficient corresponding to the noise characteristic value of the estimated noise component are variable for each channel. The noise suppression coefficient setting device of each aspect described above to be set, and the noise suppression that outputs by executing the noise suppression processing to which the flooring coefficient set for each channel by the noise suppression coefficient setting device is applied to the acoustic signal of the channel Means. In the second mode, the same effect as that of the noise suppression device of the first mode is realized. In addition, there is an advantage that highly accurate noise suppression can be realized even when the characteristics of the noise component change over time. The noise suppression device of the second mode will be described later as a fourth embodiment, for example.

  The noise suppression coefficient setting device according to each aspect described above is realized by hardware (electronic circuit) such as a DSP (Digital Signal Processor) dedicated to noise component suppression, as well as a CPU (Central Processing Unit) and the like. It is also realized by cooperation between a general-purpose arithmetic processing unit and a program (software). The program according to the present invention includes a characteristic value calculation process for calculating a noise characteristic value according to the shape of the intensity distribution of an input acoustic signal, and the acoustic signal when the kurtosis of the intensity distribution of the acoustic signal does not change before and after the noise suppression process. The flooring coefficient is set according to the noise characteristic value calculated by the characteristic value calculation process so as to satisfy the relational expression that defines the relationship between the noise characteristic value of the signal and the flooring coefficient applied to the flooring process of noise suppression processing. The computer executes the first coefficient setting process to be variably set. According to the above program, the same operation and effect as the noise suppression coefficient setting device of 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, or is provided in a form distributed via a communication network and installed in the computer.

It is a block diagram of the noise suppression apparatus which concerns on 1st Embodiment. It is a graph which shows the relationship between a noise suppression rate and kurtosis ratio. It is a graph which shows the effect of an embodiment. It is a graph which shows the effect of an embodiment. It is a block diagram of the noise suppression apparatus which concerns on 2nd Embodiment. It is a schematic diagram of a coefficient table. It is a block diagram of the noise suppression apparatus which concerns on 3rd Embodiment. It is a block diagram of the noise suppression apparatus which concerns on 4th Embodiment. It is a block diagram of a unit processing part. It is a graph which shows the effect of 4th Embodiment. It is a graph which shows the effect of 4th Embodiment. It is a graph which shows the effect of 4th Embodiment.

<First Embodiment>
FIG. 1 is a block diagram of a noise suppression device 100A according to the first embodiment of the present invention. A signal supply device 12 and a sound emission device 14 are connected to the noise suppression device 100A. The signal supply device 12 supplies the acoustic signal x (t) to the noise suppression device 100A. The acoustic signal x (t) is a time-domain signal that represents a waveform of a mixed sound of a target sound component (for example, sound such as voice or musical sound) and a noise component (environmental sound such as air conditioning equipment operation sound or hustle). Yes (t: time). A sound collection device that collects ambient sounds 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 the acoustic signal x (t) to the noise suppression device 100A Alternatively, a communication device that receives the acoustic signal x (t) from the communication network and supplies it to the noise suppression device 100A can be employed as the signal supply device 12.

  The noise suppression device 100A is an acoustic processing device that generates an acoustic signal y (t) from the acoustic signal x (t) supplied by the signal supply device 12. The acoustic signal y (t) is a time-domain signal representing a waveform of a sound (a sound in which the target sound component is emphasized) in which a noise component is suppressed from the acoustic signal x (t). The sound emitting device 14 (for example, a speaker or headphones) reproduces sound waves according to the acoustic signal y (t) generated by the noise suppressing device 100A. 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 noise suppression device 100 </ b> A is realized by a computer system including an arithmetic processing device 22 and a storage device 24. The storage device 24 stores a program PGM 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 PGM stored in the storage device 24 to thereby generate a plurality of functions (frequency analysis unit 32, noise estimation) for generating the acoustic signal y (t) from the acoustic signal x (t). Unit 34, noise suppression unit 36, waveform synthesis unit 38, characteristic value calculation unit 42, first coefficient setting unit 44). A configuration in which each function of the arithmetic processing unit 22 is distributed over a plurality of integrated circuits, or a configuration in which a dedicated electronic circuit (DSP) realizes each function may be employed.

  The frequency analysis unit 32 sequentially generates a spectrum (complex spectrum) X (f, τ) of the acoustic signal x (t) for each frame on the time axis. For the generation of the spectrum X (f, τ), a known frequency analysis such as short-time Fourier transform can be arbitrarily employed. Symbol τ is a variable that specifies a frame, and symbol f is a variable that specifies a frequency. Note that a filter bank including a plurality of bandpass filters having different passbands can also be employed as the frequency analysis unit 32.

  The noise estimation unit 34 generates a spectrum (complex spectrum) N (f, τ) of a noise component (hereinafter referred to as “estimated noise component”) n (t) estimated to be included in the acoustic signal x (t). A known technique can be arbitrarily adopted to generate the spectrum N (f, τ) of the estimated noise component n (t). For example, the noise estimation unit 34 divides the acoustic signal x (t) on the time axis into a target sound section where the target sound component exists and a noise section where the target sound component does not exist, and the spectrum of each frame in the noise section. X (f, τ) is specified as the spectrum N (f, τ) of the estimated noise component n (t). A known voice detection technique (VAD: Voice Activity Detection) is arbitrarily adopted for the classification of the target sound section and the noise section.

数式(1)の記号ｊは虚数単位を意味し、記号θx(f,τ)は音響信号ｘ(t)の位相スペクトルを意味する。また、数式(1)の記号|ＹQ(f,τ)|は、音響信号ｘ(t)のスペクトルＸ(f,τ)に対して反復回数Ｑの雑音抑圧処理を実行した時点の振幅スペクトルである。各回の雑音抑圧処理では、数式(2A)で表現される減算処理と数式(2B)で表現されるフロアリング処理とが周波数ｆ毎に択一的に実行される。

The noise suppression unit 36 performs the spectrum (complex spectrum) Y of the acoustic signal y (t) by iterative noise suppression for the spectrum X (f, τ) of the acoustic signal x (t) in each frame of the target sound section and the noise section. (f, τ) is sequentially generated for each frame. The iterative noise suppression is signal processing in which noise suppression processing for the spectrum X (f, τ) of the acoustic signal x (t) is repeated cumulatively over a predetermined number of iterations Q (Q is a natural number). The spectrum Y (f, τ) generated by the noise suppression unit 36 is expressed by the following formula (1).

The symbol j in Equation (1) means an imaginary unit, and the symbol θx (f, τ) means the phase spectrum of the acoustic signal x (t). In addition, the symbol | YQ (f, τ) | in the equation (1) is an amplitude spectrum at the time when the noise suppression processing of the number of iterations Q is performed on the spectrum X (f, τ) of the acoustic signal x (t). is there. In each noise suppression process, a subtraction process expressed by Expression (2A) and a flooring process expressed by Expression (2B) are alternatively executed for each frequency f.

The amplitude spectrum | Yi (f, τ) | in the equations (2A) and (2B) means an amplitude spectrum at the time when the i-th noise suppression process is completed. The amplitude spectrum | X (f, τ) | of the acoustic signal x (t) is used as the initial value | Y0 (f, τ) | of the amplitude spectrum | Yi (f, τ) | | Y1 (f, τ) |). The symbol Eτ [| Ni-1 (f, τ) | ² ] in Equation (2A) is the time average of the power | Ni-1 (f, τ) | ² of the estimated noise component n (t) over a plurality of frames. means. Note that the amplitude spectrum | Yi (f, τ) | (formula (2A)) is calculated by calculating the amplitude spectrum | Yi-1 (f, τ) | is applied as the amplitude spectrum | Ni-1 (f, τ) | of the estimated noise component n (t). That is, the estimated noise component n (t) is updated for each noise suppression process.

As can be understood from the equation (2A), the subtraction processing is performed by calculating the time average (ie, estimated noise) of the power spectrum | Ni-1 (f, τ) | ² of the estimated noise component n (t) and the suppression coefficient β. The amplitude spectrum | Yi (f, τ) | is calculated by subtracting the multiplication value from the power spectrum | Yi-1 (f, τ) | ² after the noise suppression processing immediately before ((i−1) th). It is processing. Therefore, the suppression coefficient (subtraction coefficient) β functions as a variable for controlling the suppression intensity of the noise component. In the first embodiment, the suppression coefficient β is fixed to a predetermined value (for example, 2 or 4).

  On the other hand, the flooring process of the equation (2B) executed when the numerical value after the subtraction of the equation (2A) becomes a negative number is the amplitude spectrum | Yi after the noise suppression process of the immediately previous ((i-1)) th time. This is a process for calculating the amplitude spectrum | Yi (f, τ) | by multiplying -1 (f, τ) | by the flooring coefficient η. That is, the flooring coefficient η functions as a variable that defines the lower limit value of the amplitude spectrum | Yi (f, τ) |. As described in Equation (1), the phase spectrum θx (f, τ) of the acoustic signal x (t) is added to the amplitude spectrum | YQ (f, τ) | after completion of the Qth noise suppression process. The spectrum Y [f, t] is supplied to the waveform synthesis unit 38 for each frame.

  1 generates a time domain acoustic signal y (t) from the spectrum Y (f, τ) generated by the noise suppression unit 36 for each frame. Specifically, the waveform synthesizer 38 converts the spectrum Y (f, τ) of each frame into a time domain signal by inverse Fourier transform, and connects the preceding and subsequent frames to each other to connect the acoustic signal y (t). Is generated. The acoustic signal y (t) generated by the waveform synthesizer 38 is supplied to the sound emitting device 14 and reproduced as a sound wave.

The characteristic value calculation unit 42 in FIG. 1 calculates a shape parameter α 0 corresponding to the characteristic of the noise component (estimated noise component n (t)) in the acoustic signal x (t) from the acoustic signal x (t). Calculate. The shape parameter α 0 is the power | X (f, τ) | ² (that is, the power of the estimated noise component n (t) | N (f, τ) | ² ) A variable that changes according to the shape of the frequency distribution (hereinafter referred to as “intensity distribution”). The characteristic value calculator 42 of the first embodiment calculates a shape parameter α 0 of a probability distribution that approximates the intensity distribution of the acoustic signal x (t). A typical example of the probability distribution is a Gaussian distribution. The probability density function P (x) of the gamma distribution that approximates the intensity distribution of the acoustic signal x (t) with the power x (x = | X (f, τ) | ² ) of the acoustic signal x (t) as a random variable is This is expressed by Equation (3).

The shape parameter α in the equation (3) is defined by the following equations (4A) and (4B), and the scaling parameter θ in the equation (3) is defined by the following equation (5). In addition, the symbol Γ (α) in Equation (3) means a gamma function defined by Equation (6) below. The symbol E [] means an average value (expected value). The characteristic value calculation unit 42 in FIG. 1 applies the power | X (f, τ) | ² of the acoustic signal x (t) in the noise interval to the random variable x of the equation (4B) and calculates the equation (4A). The shape parameter α to be calculated is calculated as the shape parameter α0 of the acoustic signal x (t) before execution of the first noise suppression processing.

  The first coefficient setting unit 44 in FIG. 1 variably changes the flooring coefficient η that the noise suppression unit 36 applies to the flooring process of Equation (2B) according to the shape parameter α0 calculated by the characteristic value calculation unit 42. Set. Specifically, the flooring coefficient η is calculated so that the musical noise of the acoustic signal y (t) resulting from the noise suppression process is minimized and the noise component is surely suppressed by the noise suppression process. .

  In order to specify the condition that the flooring coefficient η should satisfy, the relationship between the degree of occurrence of musical noise and the degree of suppression of noise components will be examined by focusing attention on a single noise suppression process for convenience. First, considering that the musical noise caused by noise suppression processing is non-Gaussian noise, the kurtosis that is an index of Gaussianity of the intensity distribution (probability density function) was caused by the noise suppression processing. This is used as a quantitative indicator of the amount of musical noise generated. Specifically, considering the tendency that musical noise becomes more apparent as the change in kurtosis before and after the noise suppression process becomes larger, the relative kurtosis KB after execution to the kurtosis KA before execution of the noise suppression process The ratio (hereinafter referred to as “kurtosis ratio”) κ is used as an index of the generation amount of musical noise (κ = KB / KA). The larger the kurtosis ratio κ is, the more musical noise is. When the kurtosis ratio κ is 1 (when the kurtosis does not change before and after the execution of the noise suppression process), it can be evaluated that no musical noise has occurred. The correlation between kurtosis (kurtosis ratio κ) and musical noise is also described in detail in Patent Document 1.

数式(7)の記号μm（μ2，μ4）は、確率密度関数Ｐ(x)の原点回りのｍ次モーメントを意味し、以下の数式(8)で定義される。

The kurtosis K of the probability density function P (x) is expressed by the following equation (7).

The symbol μm (μ2, μ4) in the equation (7) means an m-th moment around the origin of the probability density function P (x), and is defined by the following equation (8).

As can be understood from the description in Non-Patent Document 1, the m-th moment μm after the noise suppression processing is expressed by the following formula (9). The function M (α, β, η, m) of Equation (9) is defined by Equation (10).

The symbol Γ (b, a) in Equation (10) is a first-type incomplete gamma function defined by the following Equation (11), and the symbol γ (b, a) in Equation (10) is It is a second type incomplete gamma function defined by Equation (12).

したがって、雑音抑圧処理の前後にわたる尖度比κは、以下の数式(15)で表現される。

From Expression (7) and Expression (9), the following Expression (13) indicating the kurtosis K after execution of the noise suppression process is derived. Further, the kurtosis K before execution of the noise suppression process is expressed by the following formula (14) in which the suppression coefficient β and the flooring coefficient η are set to 0 in formula (13).

Therefore, the kurtosis ratio κ before and after the noise suppression process is expressed by the following formula (15).

On the other hand, a noise reduction rate R is introduced as an index value of the degree of suppression of noise components by noise suppression processing. As understood from Non-Patent Document 1, the noise suppression rate R is obtained by calculating the shape parameter α of the intensity distribution of the acoustic signal x (t) and the function M (α, β, η, m) of Equation (10). Including the following expression (16). The noise suppression rate R in Equation (16) means the difference (decibel value) between the SN (Signal to Noise) ratio after execution of the noise suppression processing and the SN ratio before execution of the noise suppression processing. It functions as an index of the amount of noise component suppression in the suppression process.

  FIG. 2 is a graph showing the relationship between the kurtosis ratio κ (vertical axis) in Expression (15) and the noise suppression rate R (horizontal axis) in Expression (16). In FIG. 2, it is assumed that the shape parameter α of the noise component is 1.0 (white Gaussian noise), and the flooring coefficient η is set to 0.0 (dashed line) and 1.0. The correlation between the kurtosis ratio κ and the noise suppression rate R when the suppression coefficient β is changed for each (solid line) is illustrated. The arrow in FIG. 2 means the direction in which the suppression coefficient β increases.

  When the flooring coefficient η is set to 0.0, both the kurtosis ratio κ and the noise suppression rate R monotonously increase as the suppression coefficient β increases. Therefore, musical noise increases while the effect of noise suppression increases. On the other hand, when the flooring coefficient η is set to 1.0, the relationship between the kurtosis ratio κ and the noise suppression rate R draws a hysteresis loop with respect to the change of the suppression factor β, and the noise suppression rate R becomes 0. There is a specific point p where the kurtosis ratio κ is 1 except for the point (when the noise component is not suppressed). The fact that the kurtosis ratio κ is maintained at 1 (that is, the kurtosis K does not change before and after the noise suppression process) means that no musical noise is logically generated before and after the noise suppression process. That is, it can be confirmed from FIG. 2 that there is a combination of a suppression coefficient β and a flooring coefficient η that can effectively suppress a noise component without generating musical noise. Assuming the case where the suppression coefficient β is fixed to a predetermined value, the noise component can be suppressed without generating musical noise by setting the flooring coefficient η to an appropriate value.

  Based on the above knowledge, it is considered to establish both the first condition for preventing the generation of musical noise and the second condition for suppressing the noise component. The first condition is a condition that the kurtosis K does not change before and after the noise suppression process, and the second condition is a condition that the noise suppression rate R by the noise suppression process is a positive number. Each of the first condition and the second condition will be described in detail below.

[First condition]
The shape parameter α of the intensity distribution of the kurtosis K (αi, β, η) after execution of the i-th (i = 0, 1, 2,...) Noise suppression processing is calculated as the (i + 1) -th. When the shape parameter αi + 1 of the intensity distribution subject to noise suppression processing is used, the following formula (17) is established between the kurtosis K (αi, β, η) and the shape parameter αi + 1. To do. The derivation of Equation (17) is described in detail in Non-Patent Document 1.

The kurtosis K (α0,0,0) of the intensity distribution (that is, the intensity distribution of the acoustic signal x (t)) before execution of the first (i = 0) noise suppression processing is expressed by the following equation (18). The

When the first condition is satisfied, the kurtosis K (αi, β, η) at an arbitrary time point is expressed as the kurtosis K (α0, 0, 0 before noise suppression processing) as expressed by the following formula (19). ).

The function M (α, β, η, m) of the equation (10) is changed into a function S (α, β, m) and a function F (α that do not include the flooring coefficient η, as shown in the following equation (20). , β, m), the kurtosis K (α 0, β, η) immediately after the first noise suppression processing is expressed by Equation (21).

フロアリング係数ηの４乗を変数Ｈで置換すると、数式(22)は以下の数式(23)に変形される。

From the formula (19) and the formula (21), the following formula (22) expressing the first condition is derived.

When the fourth power of the flooring coefficient η is replaced with the variable H, Equation (22) is transformed into the following Equation (23).

第１実施形態では抑圧係数βは所定値に固定される。したがって、数式(24)は、音響信号ｘ(t)の強度分布の尖度が雑音抑圧処理（１回の雑音抑圧処理およびＱ回にわたる雑音抑圧処理の反復）の前後で変化しない場合における形状母数αとフロアリング係数ηとの関係を規定する関係式として利用可能である。 By arranging Equation (23) with respect to the variable H, the following Equation (24) defining the flooring coefficient η (η = H ^1/4 ) that satisfies the first condition is derived.

In the first embodiment, the suppression coefficient β is fixed to a predetermined value. Therefore, Equation (24) is the shape matrix in the case where the kurtosis of the intensity distribution of the acoustic signal x (t) does not change before and after the noise suppression process (one noise suppression process and Q noise suppression processes repeated). It can be used as a relational expression that defines the relationship between the number α and the flooring coefficient η.

[Second condition]
The second condition is a condition that the noise suppression rate R in Expression (16), which represents the amount of noise component suppression in one noise suppression process, is a positive number. Therefore, by using the above-described function S (α, β, m) and function F (α, β, m), Equation (25) expressing the second condition is derived.

Taking into account that the flooring coefficient η is a positive number, the following formula (26) that defines the range of the flooring coefficient η that satisfies the second condition is derived.

The first coefficient setting unit 44 in FIG. 1 calculates the flooring coefficient η according to the shape parameter α0 calculated by the characteristic value calculation unit 42 so as to satisfy the expressions (24) and (26). Specifically, the first coefficient setting unit 44 calculates the variable H by executing the calculation of the equation (24) with respect to the shape parameter α0, and calculates the fourth root of the variable H as the flooring coefficient η (η = H ^1/4 ). From the equation (24), a plurality of flooring coefficients η are calculated. The first coefficient setting unit 44 selects one flooring coefficient η that satisfies Expression (26) among the plurality of flooring coefficients η. The flooring coefficient η set by the first coefficient setting unit 44 is applied to the noise suppression processing (flooring processing of Expression (2B)) by the noise suppression unit 36.

  As described above, in the first embodiment, the flooring coefficient η is the estimated noise component n (t) so as to satisfy the first condition (Formula (24)) that the kurtosis K does not change before and after the noise suppression process. Is variably set according to the shape parameter α0. Therefore, as described in detail below, it is possible to suppress the noise component of the acoustic signal x (t) while effectively suppressing (ideally completely preventing) the generation of musical noise.

  FIG. 3 is a graph showing the kurtosis ratio κ before and after the noise suppression process, and FIG. 4 is a graph showing the cepstrum distortion d of the acoustic signal y (t) after the noise suppression process. The cepstrum distortion d is an index value indicating an error between the cepstrum of the target sound component and the cepstrum of the acoustic signal y (t) after the noise suppression processing. That is, it can be evaluated that the noise suppression performance is higher as the cepstrum distortion d is smaller (that is, the spectral envelope of the target sound component is faithfully reproduced).

  3 and 4, when the flooring coefficient η is fixed to a predetermined value and the noise suppression process is executed only once (comparative 1 / One-Shot SS), the noise suppression process using the Wiener filter is executed. When performing noise suppression using MMSE-STSA (Minimum Mean-Square Error-Short-Time Spectral Amplitude) (Comparative 3 / MMSE-STSA) One embodiment is shown in comparison. 3 and 4, when white Gaussian noise (α0 = 0.97) is used as the noise component of the acoustic signal x (t) (White Gaussian Noise) and the conversational sound (α0 = 0.21) are represented by the acoustic signal x ( It is assumed that the noise component is t).

  As can be understood from FIG. 4, according to the first embodiment, when the noise component is a conversational sound, noise suppression performance exceeding 1 to 3 is realized. Further, when the noise component is white Gaussian noise, a noise suppression performance that exceeds Comparative 1 and Comparative 2 and is comparable to Comparative 3 is realized. As can be understood from FIG. 3, according to the first embodiment, the noise components before and after the noise suppression are compared with the proportional 1 to the proportional 3 regardless of whether the noise component is white Gaussian noise or conversational sound. The kurtosis ratio κ over the range is reduced to a value close to 1. That is, it can be understood that the musical noise can be effectively reduced according to the first embodiment even when compared with any one of the proportionality 1 to the proportionality 3. As described above, according to the first embodiment, it is possible to realize advanced noise suppression while effectively reducing musical noise.

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

  FIG. 5 is a block diagram of a noise suppression device 100B according to the second embodiment. As shown in FIG. 5, the noise suppression device 100B of the second embodiment has a configuration in which a second coefficient setting unit 46 is added to the noise suppression device 100A of the first embodiment. The second coefficient setting unit 46 subtracts each noise suppression process according to the shape parameter α 0 calculated by the characteristic value calculation unit 42 and the flooring coefficient η set by the first coefficient setting unit 44 (formula (2A )) Is set to be variable. For setting the suppression coefficient β, the coefficient table TBL of FIG. 6 stored in advance in the storage device 24 is used.

  The coefficient table TBL in FIG. 6 shows the numerical value of the suppression coefficient β for each combination of the numerical values of the shape parameter α0 (α0_1, α0_2,...) And the numerical values of the flooring coefficient η (η_1, η_2,. This is a data table in which (β11, β12,...) Is associated. The numerical value of the suppression coefficient β corresponding to the combination of the shape parameter α0 and the flooring coefficient η is selected in advance so that high noise suppression performance is realized under the shape parameter α0 and the flooring coefficient η. .

係数テーブルＴBLのうち各形状母数α0とフロアリング係数ηとの組合せに対応する抑圧係数βは、その形状母数α0およびフロアリング係数ηのもとで反復回数Ｑにわたる雑音抑圧処理を実行した場合の数式(27)の雑音抑圧率Ｒ(αQ,β,η)に応じて設定される。具体的には、抑圧係数βは、雑音抑圧率Ｒ(αQ,β,η)が最大となる数値や雑音抑圧率Ｒ(αQ,β,η)が所定の目標値Ｒtarを上回る数値に設定される。目標値Ｒtarは、例えば入力装置（図示略）に対する利用者からの指示に応じて可変に設定される。 Since the noise suppression rate R in Expression (16) is a logarithmic value of the power ratio before and after the noise suppression process, it is accumulated (added) for each noise suppression process. Therefore, the noise suppression rate R (αi + 1, β, η) at the time when the (i + 1) th noise suppression process is executed is expressed by the following equation (27).

In the coefficient table TBL, the suppression coefficient β corresponding to the combination of each shape parameter α0 and the flooring coefficient η was subjected to noise suppression processing over the number of iterations Q under the shape parameter α0 and the flooring coefficient η. It is set according to the noise suppression rate R (αQ, β, η) of the case (27). Specifically, the suppression coefficient β is set to a value at which the noise suppression rate R (αQ, β, η) is maximum or a value at which the noise suppression rate R (αQ, β, η) exceeds a predetermined target value Rtar. The The target value Rtar is variably set according to an instruction from the user to an input device (not shown), for example.

  The second coefficient setting unit 46 searches the coefficient table TBL for a suppression coefficient β corresponding to the combination of the shape parameter α 0 calculated by the characteristic value calculation unit 42 and the flooring coefficient η set by the first coefficient setting unit 44. Obtaining and instructing the noise suppression unit 36. The noise suppression unit 36 repeats the noise suppression process using the flooring coefficient η specified by the first coefficient setting unit 44 and the suppression coefficient β specified by the second coefficient setting unit 46 over a predetermined number of iterations Q. Thus, the spectrum Y (f, τ) is generated.

  In the second embodiment, the same effect as in the first embodiment is realized. In the second embodiment, the suppression coefficient β is variably set according to the shape parameter α 0 calculated by the characteristic value calculation unit 42 and the flooring coefficient η set by the first coefficient setting unit 44. Compared with a configuration in which the coefficient β is fixed to a predetermined value, the noise suppression performance (noise suppression rate R) can be improved.

  In the above description, the suppression coefficient β is acquired from the coefficient table TBL, but the method of setting the suppression coefficient β is changed as appropriate. For example, the second coefficient setting unit 46 applies each of the plurality of candidate values b of the suppression coefficient β to the equation (27) together with the shape parameter α 0 and the flooring coefficient η, so that the noise suppression rate R ( It is also possible to calculate αQ, β, η) and select a candidate value b that maximizes the noise suppression rate R (αQ, β, η) as the suppression coefficient β. In addition, a configuration in which the coefficient table TBL and the calculation of Expression (27) are used together can be employed. For example, the second coefficient setting unit 46 acquires the suppression coefficient β corresponding to the combination of the shape parameter α0 and the flooring coefficient η from the coefficient table TBL, and a plurality of candidate values b within a predetermined range including the suppression coefficient β. The noise suppression rate R (αQ, β, η) is calculated for each of the above by the calculation of Equation (27). Then, the candidate value b that maximizes the noise suppression rate R (αQ, β, η) is selected as the suppression coefficient β.

<Third Embodiment>
FIG. 7 is a block diagram of a noise suppression device 100C according to the third embodiment. As shown in FIG. 7, the noise suppression device 100C of the third embodiment has a configuration in which an iteration number setting unit 48 is added to the noise suppression device 100A of the first embodiment. The iteration number setting unit 48 sets the iteration number Q of the noise suppression processing by the noise suppression unit 36 according to the shape parameter α 0 calculated by the characteristic value calculation unit 42 and the flooring coefficient η set by the first coefficient setting unit 44. Set to variable.

  Specifically, the iteration number setting unit 48 sets the shape parameter α 0 calculated by the characteristic value calculation unit 42 and the flooring coefficient set by the first coefficient setting unit 44 for each of the plurality of candidate values q of the iteration number Q. The noise suppression rate R (αq, β, η) when the noise suppression processing is repeated as many times as the candidate value q based on η and a predetermined suppression coefficient β is calculated by the calculation of Equation (27), The minimum candidate value q in which the suppression rate R (αq, β, η) exceeds the target value Rtar is determined as the number of iterations Q. The target value Rtar is variably set according to an instruction from the user to an input device (not shown), for example. The noise suppression unit 36 repeats the noise suppression process over the number of iterations Q set by the iteration number setting unit 48.

  In the third embodiment, the same effect as in the first embodiment is realized. In the third embodiment, the number of iterations Q of the noise suppression process is variably set according to the shape parameter α 0 calculated by the characteristic value calculation unit 42 and the flooring coefficient η set by the first coefficient setting unit 44. Therefore, it is possible to prevent excessive and insufficient noise suppression processing. Specifically, there is an advantage that it is possible to reduce the amount of calculation of the arithmetic processing unit 22 by preventing the noise suppression process from being insufficient and reliably achieving the target value Rtar and preventing the noise suppression process from being excessive.

  In the above description, the number of iterations Q is set by the calculation of Equation (27), but the number of iterations Q can also be set using a table prepared in advance. For example, a table in which the number of iterations Q that can achieve the target value Rtar for each combination of each value of the shape parameter α0 and each value of the flooring coefficient η is stored in the storage device 24 in advance, and the characteristic value calculation unit The iteration number setting unit 48 searches the table for the number of iterations Q corresponding to the shape parameter α 0 calculated by 42 and the flooring coefficient η set by the first coefficient setting unit 44 and instructs the noise suppression unit 36.

<Fourth embodiment>
FIG. 8 is a block diagram of a noise suppression device 100D according to the fourth embodiment. As shown in FIG. 8, the noise suppression device 100D of the fourth embodiment includes a sound collection unit 50, a frequency analysis unit 52, a signal processing unit 54, an output processing unit 56, and a waveform synthesis unit 58. The sound collection unit 50 is a microphone array composed of D (D is a natural number of 2 or more) sound collection devices 51 that are spaced apart from each other, and D-channel sound generated by each sound collection device 51. The signals x1 (t) to xD (t) are supplied to the frequency analysis unit 52. Each acoustic signal xd (t) (d = 1 to D) is a time domain signal indicating a waveform of a mixed sound of a target sound component and a noise component. The target sound component is an acoustic component that arrives from a specific sound source direction. The noise suppression device 100D generates an acoustic signal y (t) in which noise components are suppressed from the acoustic signals x1 (t) to xD (t).

The frequency analysis unit 52 performs spectrum Xd ^[0] (f, τ) (X1 ^[0] (f, τ) to XD ^[0] (f) of each of the D-channel acoustic signals x1 (t) to xD (t). , τ)) are generated sequentially for each frame. As shown in FIG. 8, a D-dimensional vector (hereinafter referred to as “observation vector”) V ^[0] (f, τ) whose elements are the spectra X1 ^[0] (f, τ) to XD ^[0] (f, τ). τ) is output from the frequency analysis unit 52 (V ^[0] (f, τ) = [X1 ^[0] (f, τ), X2 ^[0] (f, τ),..., XD ^[0] (f, τ)] ^T ). The symbol T means transposition of the matrix.

The signal processing unit 54 uses the D channel spectrum X1 ^[0] (f, τ) to XD ^[0] (f, τ) (observation vector V ^[0] (f, τ)) generated by the frequency analysis unit 52. The D channel spectrum Y1 (f, τ) to YD (f, τ) is generated. The spectrum Yd (f, τ) is a spectrum obtained by suppressing a noise component from the spectrum Xd ^[0] (f, τ). As shown in FIG. 8, the signal processing units 54 are connected to each other in cascade, and D-channel acoustic signals x1 (t) to xD (t) (X1 ^[0] (f, τ) to XD ^[0] ( f, τ)) are sequentially processed, and Q unit processing units U [1] to U [Q] are included.

FIG. 9 is a block diagram of the unit processing unit U [q] at the q-th stage (q = 1 to Q) of the signal processing unit 54. Each unit processing unit U [q] are front observation vector V ^[q-1] supplied from (the (q-1) stage) (f, tau) in the noise suppressing process observation vectors V for the ^[q] (f , τ). The observation vector V ^[0] (f, τ) generated by the frequency analysis unit 52 is supplied to the unit processing unit U [1] in the first stage. The observation vector V ^[Q] (f, τ) generated by the unit processing unit U [Q] at the Q-th stage (final stage) corresponds to the spectrum Y1 (f, τ) to YD (f, τ) of the D channel. (V ^[Q] (f, τ) = [Y1 (f, τ), Y2 (f, τ),..., YD (f, τ)] ^T ).

As shown in FIG. 9, each unit processing unit U [q] includes a noise estimation unit 62, a noise suppression coefficient setting unit 64, and a noise suppression unit 66. The noise estimation unit 62 generates an estimated noise vector Z ^[q] (f, τ) of the estimated noise component by independent component analysis with respect to the observed vector V ^[q-1] (f, τ). The noise estimation unit 62 according to the fourth embodiment includes an independent component analysis unit 622 and a reverse projection unit 624.

数式(28)の記号Ｗ^[q](f)は分離行列を意味し、公知の更新式の演算を累積的に反復することで算定される。 The independent component analysis unit 622 performs sound source separation using frequency domain independent component analysis (FD-ICA) on the observation vector V ^[q-1] (f, τ). Thus, the separation vector G ^[q] (f, τ) is generated. The separation vector G ^[q] (f, τ) is composed of D elements g1 ^[q] (f, τ) to gD ^[q] (f, τ), and is expressed by the following equation (28). .

Symbol W ^[q] (f) in Equation (28) means a separation matrix, and is calculated by cumulatively repeating the calculation of a known update equation.

The separation vector G _(noise) obtained by replacing the element (gs ^[q] (f, τ)) corresponding to the target sound component in the separation vector G ^[q] (f, τ) generated by the independent component analysis unit 622 with zero. ^[q] (f, τ) is supplied to the reverse projection unit 624. That is, the separation vector G _(noise) ^[q] (f, τ) is expressed by the following equation (29). Element gs ^[q] (f, τ) is the target sound component estimated by the independent component analysis.

推定雑音ベクトルＺ^[q](f,τ)は、各チャネルに対応するＤ個の推定雑音成分ｚ1^[q](f,τ)〜ｚD^[q](f,τ)を要素とするベクトルである（Ｚ^[q](f,τ)＝［ｚ1^[q](f,τ)，ｚ2^[q](f,τ)，……，ｚD^[q](f,τ)］^T）。 The inverse projection unit 624 applies an inverse projection (projection back) for solving the scaling problem of independent component analysis to the separation vector G _(noise) ^[q] (f, τ) to thereby estimate the noise vector Z ^[q]. Generate (f, τ). Specifically, the estimated noise vector Z ^[q] (f, τ) is generated by the calculation of the following equation (30).

The estimated noise vector Z ^[q] (f, τ) is a vector having D estimated noise components z1 ^[q] (f, τ) to zD ^[q] (f, τ) corresponding to each channel as elements. (Z ^[q] (f, τ) = [z1 ^[q] (f, τ), z2 ^[q] (f, τ),..., ZD ^[q] (f, τ)] ^T ).

The noise suppression coefficient setting unit 64 in FIG. 9 is an element that sets the flooring coefficients η [1] to η [D] of each channel, and includes a characteristic value calculation unit 642 and a first coefficient setting unit 644. Composed. The characteristic value calculator 642 calculates the shape parameters α 0 [1] to α 0 [D] of each channel from the estimated noise vector Z ^[q] (f, τ). Shape parameter .alpha.0 [d] is a variable corresponding to the intensity distribution of the estimated noise vector ^{Z [q] (f, τ} ) of the estimated noise component ^{zd [q] (f, τ} ). The characteristic value calculation unit 642 calculates the shape parameter α0 [d] from the estimated noise component zd ^[q] (f, τ). The characteristic value calculation unit 42 of the first embodiment uses the estimated noise component n (t). This is the same as the method of calculating the shape parameter α0 from

  The first coefficient setting unit 644 calculates the flooring coefficients η [1] to η [D] for each channel from the shape parameters α0 [1] to α0 [D] calculated by the characteristic value calculation unit 642. Specifically, the first coefficient setting unit 644 is similar to the first coefficient setting unit 44 of the first embodiment in that the mathematical expression (24) expressing the first condition and the mathematical expression (26) expressing the second condition are: The flooring coefficient η [d] corresponding to the shape parameter α0 [d] is calculated so as to satisfy the above.

The noise suppression unit 66 in FIG. 9 performs the noise suppression process using the flooring coefficients η [1] to η [D] calculated by the first coefficient setting unit 644 as the observation vector V ^[q-1] (f, τ). To generate an observation vector V ^[q] (f, τ). Noise suppression processing is performed individually for each channel. That is, the noise suppression unit 66 performs a noise suppression process using the flooring coefficient η [d] on the spectrum Xd ^[q-1] (f, τ) of the acoustic signal xd (t), thereby performing the spectrum Xd. ^[q] (f, τ) is generated. Specifically, the noise suppression unit 66 alternatively executes a subtraction process expressed by the following formula (31A) and a flooring process expressed by the formula (31B) for each frequency f. That is, in the signal processing unit 54 of the fourth embodiment, Q times of noise suppression processing are cumulatively repeated for the spectrum Xd ^[0] (f, τ) generated by the frequency analysis unit 52.

  The output processing unit 56 in FIG. 8 generates an acoustic signal from the spectrums Y1 (f, τ) to YD (f, τ) of the D channel generated by the signal processing unit 54 (the Q-th unit processing unit U [Q]). A spectrum Y (f, τ) of y (t) is generated for each frame. Specifically, the output processing unit 56 is a delay-and-add type beamformer that forms a collected beam (a region with high sensitivity) in the direction of the target sound component, and includes a delay unit 562 and an addition unit 562. Is done.

The delay unit 562 delays each of the spectra Y1 (f, τ) to YD (f, τ) by a delay amount corresponding to the sound source direction φ of the target sound component. The sound source direction φ is one unit processing unit U [q] selected from the Q unit processing units U [1] to U [Q] of the signal processing unit 54 (for example, the unit processing unit U [Q at the final stage). ]) From the separation matrix W ^[q] (f) estimated by the independent component analysis unit 622.

  The adding unit 564 generates the spectrum Y (f, τ) of the acoustic signal y (t) by adding the spectrums Y1 (f, τ) to YD (f, τ) of the D channel delayed by the delay unit 562. To do. As understood from the above description, the spectrum Y (f, τ) generated by the output processing unit 56 is a signal in which the target sound component in the sound source direction φ is emphasized.

  The waveform synthesizer 58 in FIG. 8 generates an acoustic signal y (t) in the time domain from the spectrum Y (f, τ) generated for each frame by the output processor 56, similarly to the waveform synthesizer 38 of the first embodiment. Generate. The acoustic signal y (t) is supplied to, for example, a sound emitting device (not shown) and reproduced as a sound wave.

Also in the fourth embodiment described above, the same effect as in the first embodiment is realized. In the fourth embodiment, since the estimated noise component zd ^[q] (f, τ) is estimated by independent component analysis for each noise suppression process (for each unit processing unit U [q]), the acoustic signal xd (t ), There is an advantage that highly accurate noise suppression can be realized even when the characteristics of the noise component fluctuate in the middle. Further, since the spectrums Y1 (f, τ) to YD (f, τ) of the D channel are added by the output processing unit 56, even if musical noise is temporarily generated in each spectrum Yd (f, τ) It is possible to make it difficult to perceive with y (t).

  10 to 12 are graphs showing the evaluation results of noise suppression. The evaluation result of noise suppression in the fourth embodiment and the evaluation result of noise suppression (hereinafter referred to as “Comparison 4”) using a known BSSA (Blind Spatial Subtraction Array) are shown in comparison. FIG. 10 is a graph showing the kurtosis ratio κ before and after noise suppression, and FIG. 11 is a graph showing the cepstrum distortion d of the acoustic signal y (t) after noise suppression. In FIG. 10 and FIG. 11, the sound component obtained by convolving an impulse response with a predetermined sound is set as a target sound component, and a noise component (station noise) recorded in a station premises and a noise component (people mixed sound) recorded in a crowd are used. The kurtosis ratio κ and the cepstrum distortion d are shown for each case added to the target sound component. FIG. 12 shows the result of subjective evaluation of the sound quality of the acoustic signal y (t) after noise suppression. Specifically, FIG. 12 shows the ratio between the subject who evaluated that the result of noise suppression according to the fourth embodiment is high sound quality and the subject who evaluated that the result of contrast 4 is high sound quality.

  As shown in FIG. 10, in the fourth embodiment, it is understood that the musical noise can be effectively suppressed because the kurtosis ratio κ before and after the noise suppression is reduced to a value close to 1 compared to the proportional 4. The On the other hand, the tendency that the spectral distortion d of the acoustic signal y (t) after the processing of the fourth embodiment is larger than that of the proportional 4 can be grasped from FIG. However, according to the subjective evaluation of FIG. 12, about 90% of the subjects evaluate that the acoustic signal y (t) after the processing according to the fourth embodiment is higher in sound quality than the comparative 4. That is, according to the fourth embodiment, an effect that noise suppression can be realized while maintaining the sound quality perceived by the listener at a high level is realized.

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

(1) In the subtraction processing of each form described above, the power spectrum | Ni of the estimated noise component n (t) from the power spectrum | Yi-1 (f, τ) | ² after the (i-1) th noise suppression. -1 (f, τ) | ² is subtracted, but as shown in the following formula (2A ′), the amplitude spectrum | Yi-1 (f, τ) | and the amplitude spectrum | Ni-1 (f , τ) | can be generalized as any positive number K. For generalization of the power index K in the subtraction process, see, for example, Inoue et al., “Mathematical analysis of the amount of musical noise generated in the generalized spectral subtraction method, Acoustical Society of Japan, 3-5-4, p.759-762, It is also described in detail in March 2010 (hereinafter referred to as “Non-Patent Document 2”).

  According to Non-Patent Document 2, it is possible to generate an acoustically natural acoustic signal y (t) by setting the power index K to a small value. Therefore, ideally, within the range of the arithmetic processing of the arithmetic processing unit 22 (for example, within a limit where a significant numerical value can be obtained by avoiding underflow under a floating point number that can be calculated by the arithmetic processing unit 22). The power index K is set to the minimum value. Specifically, the power index K can be set to a numerical value of 1 or less (more preferably, a numerical value lower than 0.5).

(2) The specific contents of the noise suppression processing are not limited to the above-described formula (2A) (or formula (2A ′)) and formula (2B). For example, the Wiener filter expressed by the following equations (32A) and (32B) can be alternatively executed for each frequency f by one noise suppression process.

したがって、前述の数式(24)および数式(26)と同様に、第１条件および第２条件の双方を満たす（すなわち理論的にはミュージカルノイズを発生させない）フロアリング係数ηを算定することが可能である。 For example, T. Inoue, et al., “Theoretical analysis of musical noise in Wiener filtering family via higher-order statistics”, Proc As disclosed in ICASSP2011, p.5076-5079, 2011, it is expressed by the following equation (33). The function MWF (α, β, η, m) of Equation (33) is defined by Equation (34).

Therefore, similarly to the above-described equations (24) and (26), it is possible to calculate the flooring coefficient η that satisfies both the first condition and the second condition (that is, theoretically does not generate musical noise). It is.

(3) In the fourth embodiment, the noise estimation unit 62 of each unit processing unit U [q] generates estimated noise components z1 ^[q] (f, τ) to zD ^[q] (f, τ) for each channel. However, it is also possible for the noise estimation unit 62 to generate a common estimated noise component for the D channel by independent component analysis. The noise suppression unit 66 has the spectrum X1 ^[q-1] (f, τ) to XD ^[q-1] (f, τ) (observation vector V ^[q-1] (f, τ)) of each D channel. The noise suppression processing for suppressing the common estimated noise component is executed.

(4) In each of the above-described embodiments, the first coefficient setting unit 44 calculates the flooring coefficient η by the calculation of Expression (24), but the flooring coefficient η corresponds to each numerical value of the shape parameter α0 and the suppression coefficient β. A configuration in which the first coefficient setting unit 44 sets the flooring coefficient η by referring to the attached coefficient table may be employed. In the coefficient table for setting the flooring coefficient η, the combination of the shape parameter α0 and the suppression coefficient β is associated with the flooring coefficient η so that the relationship (first condition) of Expression (24) is satisfied. As understood from the above description, the first coefficient setting unit 44 is included as an element that variably sets the flooring coefficient η according to the shape parameter α0 so as to satisfy the relational expression that defines the first condition. It does not matter whether the flooring coefficient η is calculated by actual calculation or obtained from the coefficient table.

(5) In each of the above-described embodiments, the noise suppression unit 36 performs repetitive noise suppression. However, the noise suppression unit 36 performs the noise suppression process only once for each frame of the acoustic signal x (t). The present invention applies. That is, iterative noise suppression processing is not essential in the present invention. Even in the configuration in which the noise suppression process is executed only once for each frame, the flooring coefficient η is set so as to satisfy the first condition and the second condition, so that it is compared with the configuration in which the flooring coefficient η is fixed to a predetermined value. For example, the desired effect of suppressing the noise component of the acoustic signal x (t) while suppressing the generation of musical noise is realized.

(6) A configuration including both the second coefficient setting unit 46 of the second embodiment and the iteration number setting unit 48 of the third embodiment may be employed. The iteration number setting unit 48 applies the flooring coefficient η set by the first coefficient setting unit 44 and the suppression coefficient β set by the second coefficient setting unit 46 to each of the plurality of candidate values q of the iteration number Q. The noise suppression rate R (αq, β, η) when the noise suppression processing is repeated q times is calculated by the calculation of Equation (27), and the noise suppression rate R (αq, β, η) is the target value Rtar. The minimum candidate value q above is determined as the number of iterations Q.

(7) In the second embodiment, the suppression coefficient β is set according to both the shape parameter α 0 and the flooring coefficient η. However, the method for controlling the suppression coefficient β is not limited to the above examples. For example, the suppression coefficient β (a numerical value that does not depend on the flooring coefficient η) can be set according to the shape parameter α0.

(8) In each embodiment described above, the shape parameter α0 of the probability density function P (x) approximating the intensity distribution of the acoustic signal x (t) is used as an index of the characteristic of the estimated noise component n (t) (noise characteristic value). However, the noise characteristic value is not limited to the shape parameter α0. For example, a statistic directly calculated from the intensity distribution of the acoustic signal x (t) (for example, a higher-order statistic such as kurtosis) or the frequency | X (f, τ) | of the acoustic signal x (t) A statistic corresponding to the distribution (for example, a shape parameter of a probability density function approximating a frequency distribution of amplitude | X (f, τ) |) can be used as a noise characteristic value. That is, the noise characteristic value is included as a numerical value (typically, a numerical value corresponding to the shape of the intensity distribution) that changes according to the characteristic of the acoustic signal x (t) (particularly the characteristic of the estimated noise component n (t)). The

(9) In each of the above-described embodiments, the noise suppression device 100 (100A, 100B, 100C) including the noise suppression unit 36 is illustrated. However, variables (for example, flooring coefficient η, suppression coefficient β, The present invention can also be realized as a device that variably sets the number of iterations Q) (hereinafter referred to as “noise suppression coefficient setting device”). The noise suppression device 100 is realized by adding the noise suppression unit 36 and the noise estimation unit 34 to the noise suppression coefficient setting device. For example, the noise suppression coefficient setting device corresponding to the first embodiment includes the characteristic value calculation unit 42 and the first coefficient setting unit 44 in each of the above-described embodiments. Further, the noise suppression coefficient setting unit 64 of the fourth embodiment corresponds to the noise suppression coefficient setting apparatus according to each aspect of the present invention. In the noise suppression coefficient setting device, the second coefficient setting unit 46 of the second embodiment and the iteration number setting unit 48 of the third embodiment can be appropriately added.

100A, 100B, 100C, 100D .... Noise suppression device, 12 ... Signal supply device, 14 ... Sound emitting device, 22 ... Calculation processing device, 24 ... Storage device, 32,52 ... Frequency analysis unit, 34 , 62 ... Noise estimation unit, 36, 66 ... Noise suppression unit, 38, 58 ... Waveform synthesis unit, 42 ... Characteristic value calculation unit, 44 ... First coefficient setting unit, 46 ... Second coefficient setting , 48... Repeat count setting section, 50... Sound collection section, 51... Sound collection equipment, 54... Signal processing section, U [q] (U [1] to U [Q]). Part, 56... Output processing part.

Claims (6)

A characteristic value calculating means for calculating a noise characteristic value according to the shape of the intensity distribution of the input acoustic signal;
When the kurtosis of the intensity distribution of the acoustic signal does not change before and after the noise suppression process, a relational expression that defines the relationship between the noise characteristic value of the acoustic signal and the flooring coefficient applied to the flooring process of the noise suppression process is A noise suppression coefficient setting device comprising: first coefficient setting means for variably setting a flooring coefficient in accordance with the noise characteristic value calculated by the characteristic value calculation means so as to satisfy. The noise suppression coefficient setting device according to claim 1, wherein the first coefficient setting means sets the flooring coefficient so that a noise suppression rate by the noise suppression processing becomes a positive number. The noise suppression coefficient according to claim 1 or 2, further comprising second coefficient setting means for variably setting a suppression coefficient for controlling a noise suppression strength according to a noise characteristic value calculated by the characteristic value calculation means. Setting device. The noise according to the noise characteristic value calculated by the characteristic value calculation means and the flooring coefficient set by the first coefficient setting means so that the accumulated value of the noise suppression rate by each noise suppression process exceeds the target value. The noise suppression coefficient setting device according to any one of claims 1 to 3, further comprising: an iteration count setting unit that variably sets an iteration count of the suppression process. A noise suppression coefficient setting device according to any one of claims 1 to 4,
A noise suppression device comprising: noise suppression means for executing noise suppression processing to which an coefficient set by the noise suppression coefficient setting device is applied to an input acoustic signal. Q-stage unit processing means for sequentially processing acoustic signals of D (D is a natural number of 2 or more) channels generated by sound collection devices arranged apart from each other;
Output processing means for emphasizing a target sound component in a specific sound source direction by delay addition of the D-channel acoustic signal after processing by the unit processing means in the final stage of the Q stages,
Each of the Q-stage unit processing means includes:
Noise estimation means for generating an estimated noise component by independent component analysis for the D-channel acoustic signal supplied to the unit processing means;
The noise suppression coefficient setting device according to any one of claims 1 to 4, wherein a flooring coefficient corresponding to a noise characteristic value of the estimated noise component is variably set for each channel;
A noise suppression device comprising: noise suppression means for executing a noise suppression process to which the flooring coefficient set for each channel by the noise suppression coefficient setting device is applied to an acoustic signal of the channel and outputting the result.

Images