JP5187666B2 - Noise suppression device and program - Google Patents

Description

  The present invention relates to a technique for suppressing noise from a mixed sound of a target sound and noise.

  Conventionally, a technique for suppressing noise from a mixed sound of target sound and noise has been proposed. For example, Non-Patent Document 1 discloses a spectral subtraction (SS) method that subtracts a noise spectrum from a spectrum of an acoustic signal. Non-Patent Document 2 discloses an MMSE-STSA (minimum mean square error short time spectral amplitude) method for multiplying the spectrum of an acoustic signal by a spectral gain selected so that the target sound (sound) is emphasized. ing.

  However, in the method of suppressing noise from an acoustic signal in the frequency domain as in Non-Patent Document 1 and Non-Patent Document 2, components that are dispersedly scattered on the time axis and the frequency axis after noise suppression are disturbing. There is a problem that the listener perceives it as a musical noise. Therefore, Non-Patent Document 3 discloses a technique for removing musical noise after noise suppression.

S. F. Boll, "Suppression of acoustic noise in using spectral subtraction", IEEE Trans., Acoustics, Speech and Signal Processing, vol.ASSP-27, no. 2, p.113-120, Apr. 1979 Y. Ephraim and D. Malah, "Speech enhancement using a minimum mean-square error short-time spectral amplitude estimator", IEEE ASSP, vol.ASSP-32, no.6, p.1109-1121, Dec. 1984 Tomomi Abe, Mitsuharu Matsumoto, Shuji Hashimoto, "Musical noise removal by time-frequency M conversion", Proceedings of the Acoustical Society of Japan, 3-6-9, p.727-p.730, March 2008

  However, the degree of musical noise generated after noise suppression differs depending on the acoustic characteristics of the acoustic signal. Therefore, even if the musical noise generated in a specific section of the acoustic signal is reduced by the technique of Non-Patent Document 3, it is not always possible to sufficiently reduce the musical noise in another section having different acoustic characteristics. In view of the above circumstances, an object of the present invention is to effectively reduce musical noise after noise suppression even when the acoustic characteristics of an acoustic signal change.

  In order to solve the above problems, a noise suppression device according to the present invention includes a plurality of noise suppression units that perform different noise suppression processes on an acoustic signal, and a kurtosis in the frequency distribution of the intensity of the acoustic signal. Index calculation means for calculating the kurtosis index value indicating the degree of change before and after the suppression process for each noise suppression process by each noise suppression means, and each noise suppression means according to each kurtosis index value calculated by the index calculation means Selecting means for selecting any one of a plurality of noise suppression processes. For example, the selection unit selects a noise suppression process having a small change in kurtosis indicated by the kurtosis index value among the plurality of noise suppression processes.

  In the above configuration, the selection means selects according to the kurtosis index value indicating the degree to which the kurtosis in the frequency distribution of the intensity of the acoustic signal has changed before and after the noise suppression process (that is, the degree of occurrence of musical noise). Since the noise suppression processing is changed, there is an advantage that the musical noise after noise suppression can be effectively reduced even when the acoustic characteristics of the acoustic signal change.

  A noise suppression apparatus according to a preferred aspect of the present invention includes a signal classification unit that classifies an acoustic signal into a noise section and a target sound section on a time axis, and a weight value for each noise suppression process is assigned to the noise section and the target sound section. And a weighting means for weighting each kurtosis index value by changing, and the selection means selects a noise suppression process according to each kurtosis index value after weighting. The above aspect has an advantage that noise suppression processing suitable for each of the noise section and the target sound section can be stably selected even when the kurtosis index values are close to each other.

  A subtractive noise suppression process (for example, a spectral subtraction method) that subtracts a noise spectrum from the spectrum of an acoustic signal is suitable as a noise suppression process that suppresses musical noise in a section where speech is dominant. On the other hand, multiplicative noise suppression processing (for example, the MMSE-STSA method or MAP method) that multiplies the spectrum of an acoustic signal by a spectral gain that enhances the target sound, for example, eliminates musical noise in a section where noise (especially stationary noise) is dominant. It is suitable as a noise suppression process to suppress. Therefore, according to the configuration in which the plurality of noise suppression processes that are candidates for selection by the selection unit include the subtraction type noise suppression process and the multiplication type noise suppression process, the voice dominant section and the noise dominant section of the acoustic signal There is an advantage that generation of musical noise can be suppressed.

  In addition, the noise suppression device according to each aspect described above is realized by hardware (electronic circuit) such as a DSP (Digital Signal Processor) dedicated to noise suppression, or a general-purpose such as a CPU (Central Processing Unit). This is also realized by cooperation between the arithmetic processing unit and the program. A program according to the present invention is a kurtosis index value indicating a degree of change in kurtosis in a frequency distribution of a plurality of noise suppression processes individually performed on an acoustic signal and before and after the noise suppression process. For each noise suppression process, and a selection process for selecting one of a plurality of noise suppression processes according to each kurtosis change index calculated in the index calculation process. According to the above program, operations and effects similar to those of the noise suppression device according to each aspect of the present invention are exhibited. The program of the present invention is provided to a user in a form stored in a computer-readable recording medium and installed in the computer, or provided from a server device in a form of distribution via a communication network and installed in the computer. Is done.

1 is a block diagram of a noise suppression device according to a first embodiment of the present invention. It is a conceptual diagram for demonstrating the division of an acoustic signal. It is a block diagram of a 1st noise suppression part. It is a block diagram of a 2nd noise suppression part. It is a conceptual diagram which shows a mode that the frequency distribution of the intensity | strength of an acoustic signal changes by noise suppression. It is a block diagram of an index calculation unit. It is a graph of kurtosis ratio in a noise area. It is a graph of the kurtosis ratio in the target sound section. It is a chart of each coefficient applied to noise suppression processing. It is a block diagram of the noise suppression apparatus which concerns on 2nd Embodiment of this invention.

<A: First Embodiment>
FIG. 1 is a block diagram of a noise suppression apparatus according to the first embodiment of the present invention. The noise suppression apparatus 100 is supplied with a time domain acoustic signal VIN representing a waveform of a mixed sound of the target sound and noise. The supply source (not shown) of the acoustic signal VIN is, for example, a sound collecting device that generates an acoustic signal VIN corresponding to surrounding sounds, or a playback device that acquires and outputs the acoustic signal VIN from a recording medium. The noise suppression device 100 generates the acoustic signal VOUT by suppressing the noise of the acoustic signal VIN. The acoustic signal VOUT is supplied to a sound emitting device (not shown) such as a speaker or headphones and reproduced as a sound wave.

  The noise suppression device 100 is realized by a computer system including an arithmetic processing device 12 and a storage device 14. The storage device 14 stores a program and various data for generating the acoustic signal VOUT from the acoustic signal VIN. A known recording medium such as a semiconductor recording medium or a magnetic recording medium is arbitrarily employed as the storage device 14.

  The arithmetic processing unit 12 functions as a plurality of elements (frequency analysis unit 20, suppression processing unit 30, index calculation unit 42, selection unit 44, waveform synthesis unit 46) by executing a program stored in the storage device 14. . In addition, a configuration in which an electronic circuit (DSP) dedicated to processing of the acoustic signal VIN realizes each element of the arithmetic processing device 12 or a configuration in which each element of the arithmetic processing device 12 is mounted on a plurality of integrated circuits in a distributed manner. Adopted.

As shown in FIG. 2, the frequency analysis unit 20 in FIG. 1 calculates a frequency spectrum Xm (n) for each of a plurality of frames FR obtained by dividing the acoustic signal VIN on the time axis. The symbol n means the n-th frequency fn among N frequencies (frequency bins) f1 to fN discretely set on the frequency axis (n = 1 to N), and the symbol m represents the frame FR. Means a number. A known technique (for example, short-time Fourier transform) is arbitrarily adopted for calculating the frequency spectrum Xm (n). The frequency spectrum Xm (n) of the mth frame FR corresponds to the addition of the frequency spectrum Sm (n) of the target sound and the frequency spectrum Nm (n) of the noise (Formula (1)).
Xm (n) = Sm (n) + Nm (n) (1)

  The suppression processing unit 30 in FIG. 1 executes different types of noise suppression processing in parallel on the frequency spectrum Xm (n) (acoustic signal VIN) generated by the frequency analysis unit 20. As shown in FIG. 1, the suppression processing unit 30 includes a first noise suppression unit 31 and a second noise suppression unit 32. The first noise suppression unit 31 generates a frequency spectrum Ym (n) _1 for each frame FR by performing a first noise suppression process on the frequency spectrum Xm (n) of each frame FR. The second noise suppression unit 32 generates a frequency spectrum Ym (n) _2 for each frame FR by executing a second noise suppression process on the frequency spectrum Xm (n).

  The first noise suppression process is a process of subtracting the frequency spectrum of noise estimated from the acoustic signal VIN (hereinafter referred to as “estimated noise spectrum”) ψm (n) from the frequency spectrum Xm (n) of the acoustic signal VIN. On the other hand, the second noise suppression process is a process of multiplying the frequency spectrum Xm (n) of the acoustic signal VIN by the spectrum gain Gm (n) selected so that the target sound of the acoustic signal VIN is emphasized. That is, the first noise suppression process corresponds to a subtraction type noise suppression process (spectral subtraction method), and the second noise suppression process corresponds to a multiplication type noise suppression process.

FIG. 3 is a block diagram of the first noise suppression unit 31. As shown in FIG. 3, the first noise suppression unit 31 includes a noise estimation unit 312 and a subtraction unit 314. The noise estimation unit 312 estimates an estimated noise spectrum (power spectrum) ψm (n) for each frame FR. A known technique is arbitrarily employed to generate the estimated noise spectrum ψm (n) (noise estimation). For example, the noise estimation unit 312 calculates the following equation (2) using the frequency spectrum Xm (n) of each frame FR in the noise section of the acoustic signal VIN in which the target sound does not exist as the noise frequency spectrum Nm (n). To generate an estimated noise spectrum ψm (n). The symbol E in Equation (2) means an average (addition) over a plurality of frames FR.
ψm (n) = E {| Nm (n) | ² } (2)

The subtraction unit 314 in FIG. 3 calculates the frequency spectrum Ym (n) _1 by subtracting the estimated noise spectrum ψm (n) from the frequency spectrum Xm (n). The frequency spectrum Ym (n) _1 is expressed by Equation (3) using the amplitude spectrum Pm (n) ^1/2 and the phase θx (n) of the frequency spectrum Xm (n).
Ym (n) _1 = (Pm (n)) ^1/2 · e ^{jθx (n)} (3)
The power spectrum Pm (n) of the equation (3) is calculated by the following equations (4a) and (4b).

That is, when the intensity (power) | Xm (n) | ^{2 of} the frequency spectrum Xm (n) exceeds a predetermined value (multiplied value of the estimated noise spectrum ψm (n) and the coefficient αm), the power spectrum Pm (n) is As shown in the equation (4a), the difference value between the intensity | Xm (n) | ² and the predetermined value (αm · ψm (n)) is set. On the other hand, when the intensity | Xm (n) | ² is lower than a predetermined value (αm · ψm (n)), the power spectrum Pm (n) has an estimated noise spectrum ψm (n) and a coefficient as shown in the equation (4b). It is set to a multiplication value (βm · ψm (n)) with βm. The coefficient αm (subtraction coefficient) and the coefficient βm (flooring coefficient) are appropriately selected according to the required degree of noise suppression.

  Next, FIG. 4 is a block diagram of the second noise suppression unit 32. As shown in FIG. 4, the second noise suppression unit 32 includes a noise estimation unit 322, a gain calculation unit 324, and a multiplication unit 326. The noise estimation unit 322 generates an estimated noise spectrum ψm (n) by the same method as the noise estimation unit 312. A configuration in which the second noise suppression unit 32 diverts the estimated noise spectrum ψm (n) generated by the noise estimation unit 312 (a configuration in which the noise estimation unit 322 is omitted) is also employed.

The gain calculation unit 324 in FIG. 4 calculates the spectral gain Gm (n) for enhancing the target sound for each frame FR. The spectral gain Gm (n) is set to a value closer to zero as the noise prevails at the frequency fn (set to a larger value as the target sound prevails at the frequency fn). For the calculation of the spectral gain Gm (n), the MMSE-STSA method of Non-Patent Document 2 is suitable as exemplified below. Specifically, the gain calculation unit 324 calculates the spectral gain Gm (n) by calculating the following formula (5). The symbol Γ in Equation (5) means a gamma function. Further, the symbol I0 means a zero-order modified Bessel function, and the symbol I1 means a first-order modified Bessel function.

The posterior signal-to-noise ratio (posteriori SNR) γm (n) and the prior signal-to-noise ratio (priori SNR) ξm (n) in Equation (5) are generated by the frequency spectrum Xm (n) generated by the frequency analysis unit 20 and the noise estimation unit 322. The estimated noise spectrum ψm (n) is calculated by the following equations (6a) and (6b). The function value F [x] of Expression (6b) is set to the variable x when the variable x is a positive number, and is set to zero when the variable x is zero or a negative number. In addition, the coefficient η in Expression (6b) is a predetermined positive number less than 1.
γm (n) = | Xm (n) | ² / ψm (n) (6a)
ξm (n) = η · | Sm-1 (n) | ² / ψm-1 (n) + (1-η) · F [γm (n) -1] (6b)

  4 calculates the frequency spectrum Ym (n) _2 by multiplying the frequency spectrum Xm (n) and the spectrum gain Gm (n) (Ym (n) _2 = Gm (n) · Xm (n). )). Since the spectrum gain Gm (n) is set so as to enhance the target sound, the noise of the acoustic signal VIN is suppressed in the frequency spectrum Ym (n) _2. The specific configuration of the suppression processing unit 30 has been described above.

  By the way, in the time series of the frequency spectrum Ym (n) _1 and the time series of the frequency spectrum Ym (n) _2, musical noise may be scattered on the time axis and the frequency axis. The index calculation unit 42 in FIG. 1 calculates a kurtosis index value σm (σm_1, σm_2), which is a quantitative measure of the degree of occurrence of musical noise caused by noise suppression processing, for each frame FR. The kurtosis index value σm will be described in detail below.

  Part (A) of FIG. 5 is a frequency distribution of the intensity (a probability density function with intensity as a random variable) in a predetermined section of the acoustic signal VIN before noise suppression. As shown in part (A) of FIG. 5, the intensity of the acoustic signal VIN is non-linearly distributed so that the frequency decreases as the intensity increases from zero.

Part (B) in FIG. 5 is a frequency distribution of intensity after noise suppression. As understood from the comparison between part (A) and part (B) in FIG. 5, the frequency of the intensity close to zero in the acoustic signal VIN (before noise suppression) tends to increase due to noise suppression. . That is, the slope of the frequency distribution in the range where the intensity is near zero changes to a steep shape after noise suppression. When kurtosis is introduced as a measure of the shape of the frequency distribution (steepness of inclination), the kurtosis KBm after the noise suppression processing is executed is the kurtosis KAm before the noise suppression processing (acoustic signal VIN). Compared to a higher value (KBm> KAm). The kurtosis κ is a higher-order statistic calculated from the n-th moment by the following formula (7).

  An acoustic signal having a lot of musical noise after noise suppression tends to have a high intensity frequency near zero. Therefore, it can be evaluated that the more the frequency at which the intensity becomes zero in the frequency distribution increases before and after noise suppression, the more musical noise is generated due to the noise suppression processing. That is, the greater the change in kurtosis κ before and after noise suppression (KAm → KBm), the more musical noise is generated after noise suppression.

  From the above tendency, the index calculation unit 42 in FIG. 1 calculates the kurtosis index value σm (σm_1, σm_2) corresponding to the change in the kurtosis κ before and after the processing by the suppression processing unit 30. The kurtosis index value σm_1 is a measure of the degree to which the kurtosis κ has changed before and after the first noise suppression processing by the first noise suppression unit 31, and is an index of the degree of occurrence of musical noise in the frequency spectrum Ym (n) _1. Used as The kurtosis index value σm_2 is a measure of the degree to which the kurtosis κ has changed before and after the second noise suppression processing by the second noise suppression unit 32, and is an index of the degree of occurrence of musical noise in the frequency spectrum Ym (n) _2. Used as The kurtosis κ in the frequency distribution of M intensities x1 to xM is calculated by the following method, for example.

数式(8)の係数Ｃは、ガンマ関数Γ(k)を利用して以下のように定義される。

The frequency distribution of M intensities x1 to xM is approximated by a function Ga (x; k, θ) of the following formula (8).

The coefficient C in Expression (8) is defined as follows using the gamma function Γ (k).

By substituting the distribution function P (x) in the defining equation of the second moment μ2 with the function Ga (x; k, θ) of the equation (8), the following equation (9) is derived.

Similar to the derivation of the second-order moment μ2, the distribution function P (x) in the definition of the fourth-order moment μ4 is replaced with the function Ga (x; 10) is derived.

Substituting the second-order moment μ2 in Equation (9) and the fourth-order moment μ4 in Equation (10) into Equation (7) yields the following Equation (11) that defines kurtosis κ.

  The kurtosis κ is calculated by calculating Expression (11) for M intensities x1 to xM. However, the method for calculating the kurtosis κ is not limited to the above examples. For example, a configuration that approximates the frequency distribution of the intensities x1 to xM with a predetermined function (for example, Equation (8)) is not essential.

FIG. 6 is a block diagram of the index calculation unit 42. The first kurtosis calculation unit 421 in FIG. 6 calculates the kurtosis (kurtosis before noise suppression) KAm in the frequency distribution of the intensity of the acoustic signal VIN from the time series of the frequency spectrum Xm (n). That is, the first kurtosis calculation unit 421 calculates the kurtosis KAm by executing the calculation of Expression (11) for M intensities x1 to xM extracted from the time series of the frequency spectrum Xm (n). As shown in FIG. 2, the intensities x1 to xM used for calculating the kurtosis KAm are the intensities | Xm in the frequency spectrum Xm (n) of each of the τ frames FR with the mth frame FR as the last. It corresponds to (n) | ² (M = τ × N).

The second kurtosis calculation unit 422 of FIG. 6 calculates the kurtosis KBm_1 after execution of the first noise suppression processing from the time series of the frequency spectrum Ym (n) _1 generated by the first noise suppression unit 31. Specifically, the second kurtosis calculation unit 422 generates M intensities | Ym (n) _1 constituting the frequency spectrum Ym (n) _1 of τ frames FR whose last is the mth frame FR. | The kurtosis KBm_1 is calculated by executing the calculation of Equation (11) with ² as the intensity x1 to xM. The third kurtosis calculation unit 423 in FIG. 6 calculates the kurtosis KBm_2 after execution of the second noise suppression processing from the time series of the frequency spectrum Ym (n) _2 by the same method as the second kurtosis calculation unit 422. To do.

The first index calculation unit 425 in FIG. 6 calculates the kurtosis index value σm_1 from the kurtosis KAm calculated by the first kurtosis calculation unit 421 and the kurtosis KBm_1 calculated by the second kurtosis calculation unit 422. The kurtosis index value σm_1 is defined by a function Fa using a relative ratio of the kurtosis KBm_1 to the kurtosis KAm (hereinafter referred to as “kurtosis ratio”) Rm_1 as shown in the following formula (12) (Rm_1) = KBm_1 / KAm). The function Fa is a monotonically increasing function that defines the relationship between the kurtosis index value σm_1 and the kurtosis ratio Rm_1.
σm_1 = Fa (Rm_1)
= Fa (KBm_1 / KAm) ...... (12)
As described above with reference to FIG. 5, the smaller the kurtosis ratio Rm_1 (change from the kurtosis KAm to the kurtosis KBm_1), the more the musical noise generated in the first noise suppression process is reduced. Therefore, it can be evaluated that the smaller the kurtosis index value σm_1 is, the smaller the musical noise after the execution of the first noise suppression process is. That is, the kurtosis index value σm_1 corresponds to an index value (scale) indicating the degree of occurrence of musical noise caused by the first noise suppression process.

6 calculates the kurtosis index value σm_2 from the kurtosis KAm calculated by the first kurtosis calculation unit 421 and the kurtosis KBm_2 calculated by the third kurtosis calculation unit 423. The kurtosis index value σm_2 is defined by a function Fa having a kurtosis ratio Rm_2 of the kurtosis KBm_2 with respect to the kurtosis KAm as a variable (Rm_2 = KBm_2 / KAm) as shown in the following equation (13). Therefore, it can be evaluated that the smaller the kurtosis index value σm_2 (the smaller the change from the kurtosis KAm indicated by the kurtosis ratio Rm_1 to the kurtosis KBm_1), the smaller the musical noise generated in the second noise suppression process. That is, the kurtosis index value σm_2 corresponds to an index value indicating the degree of occurrence of musical noise caused by the second noise suppression process.
σm_2 ＝ Fa (Rm_2)
= Fa (KBm_2 / KAm) (13)

  The selection unit 44 in FIG. 1 performs either the first noise suppression process or the second noise suppression process according to the kurtosis index value σm_1 and the kurtosis index value σm_2 calculated by the index calculation unit 42 (first noise suppression unit 31). And any one of the second noise suppression units 32) are selected for each frame FR. Specifically, the noise suppression process corresponding to the smaller one of the kurtosis index value σm_1 and the kurtosis index value σm_2 is selected. That is, when the kurtosis index value σm_1 is smaller than the kurtosis index value σm_2, the first noise suppression processing (first noise suppression unit 31) is selected, and the kurtosis index value σm_2 is smaller than the kurtosis index value σm_1. In this case, the second noise suppression process (second noise suppression unit 32) is selected. The selection unit 44 outputs the frequency spectrum Ym (n) (Ym (n) _1, Ym (n) _2) generated by the noise suppression processing selected by itself to the waveform synthesis unit 46 sequentially for each frame FR. That is, the selection unit 44 outputs the frequency spectrum Ym (n) _1 in the frame FR selected for the first noise suppression processing, and outputs the frequency spectrum Ym (n) _2 in the frame FR selected for the second noise suppression processing. .

  The waveform synthesizing unit 46 synthesizes the time-domain acoustic signal VOUT from the frequency spectrum Ym (n) (Ym (n) _1, Ym (n) _2) selected by the selection unit 44 for each frame FR. That is, the waveform synthesizer 46 calculates the acoustic signal VOUT by overlapping the time-domain signals calculated by the inverse Fourier transform for the frequency spectrum Ym (n) with respect to the plurality of frames FR on the time axis.

  In the above embodiment, the noise suppression process with the smaller kurtosis index value σm of the first noise suppression process and the second noise suppression process is selectively used for generating the acoustic signal VOUT. Therefore, as compared with a configuration in which only the first noise suppression processing is performed and a configuration in which only the second noise suppression processing is performed, as described in detail below, even when the acoustic characteristics of the acoustic signal VIN change, the musical noise Can be effectively reduced.

  FIG. 7 is a graph showing the relationship between the SN ratio and the kurtosis ratio Rm (Rm_1, Rm_2) in the noise section of the acoustic signal VIN (the section in which noise is dominant with respect to the target sound). FIG. 8 is a graph showing the relationship between the SN ratio and the kurtosis ratio Rm in the target sound section (section in which the target sound is dominant) of the acoustic signal VIN. In addition, each coefficient ((alpha) m. (Beta) m, (eta)) applied to a noise suppression process was set like FIG. 9 according to S / N ratio.

  As shown in FIG. 7, in the noise section, the kurtosis ratio Rm_2 before and after the second noise suppression process is lower than the kurtosis ratio Rm_1 before and after the first noise suppression process. That is, musical noise after noise suppression is reduced in the second noise suppression processing (MMSE-STSA method) than in the first noise suppression processing (SS method). Therefore, the selection unit 44 selects the second noise suppression processing (frequency spectrum Ym (n) _2) by the second noise suppression unit 32 in each frame FR within the noise section.

  On the other hand, as shown in FIG. 8, in the target sound section, the kurtosis ratio Rm_1 before and after the first noise suppression process is lower than the kurtosis ratio Rm_2 before and after the second noise suppression process. That is, musical noise after noise suppression is reduced in the first noise suppression process than in the second noise suppression process. Therefore, the selection unit 44 selects the first noise suppression processing (frequency spectrum Ym (n) _1) by the first noise suppression unit 31 in each frame FR within the target sound section.

  As described above, since the noise suppression processing applied to the generation of the acoustic signal VOUT is changed between the noise section and the target sound section (that is, changed according to the acoustic characteristics of the acoustic signal VIN), the acoustic signal VIN is changed. The musical noise in the noise section is reduced as compared with the configuration in which the first noise suppression processing is performed over the entire section of the target sound, and the target sound section is compared with the configuration in which the second noise suppression processing is performed over the entire section of the acoustic signal VIN. There is an advantage that the musical noise at is reduced.

  A specific usage situation of the noise suppression apparatus 100 will be exemplified. First, a case is assumed in which a speaker passes near a sound collection point (for example, a sound collection device) in a space where stationary noise such as an operation sound of an air conditioning facility exists. Since only noise is picked up when a speaker is present at a position sufficiently far from the sound pickup point, an acoustic signal VOUT is generated by the second noise suppression processing. When the speaker is close to the sound collection point, the first noise suppression processing and the second noise suppression processing are switched at any time according to the presence or absence of speech from the speaker. That is, the first noise suppression process is selected in a section where the utterance sound by the speaker is dominant, and the second noise suppression process is selected in a section where noise is dominant (for example, a section where speech is paused). Then, when the speaker moves away from the sound collection point (that is, when the noise becomes dominant), the generation of the acoustic signal VOUT by the second noise suppression processing is executed.

  Next, a state where stationary noise such as an operation sound of an air conditioner exists and a state where low stationary noise such as a mixed sound of many uttered sounds is switched at any time (for example, many noises) Suppose an exhibition where there is a speaker. In a state where stationary noise exists, the kurtosis Rm_2 is lower than the kurtosis Rm_1 as shown in FIG. 7, and therefore the generation of the acoustic signal VOUT by the second noise suppression process is selected. On the other hand, in a state where noise with low stationarity exists (that is, a state closer to FIG. 8 than FIG. 7), the kurtosis Rm_1 is lower than the kurtosis Rm_2, so that the generation of the acoustic signal VOUT by the first noise suppression processing is selected Is done. That is, the first noise suppression process or the second noise suppression process is selected according to the acoustic characteristics of the noise.

<B: Second Embodiment>
Next, a second embodiment of the present invention will be described. In addition, about the element in which an effect | action and a function are equivalent to 1st Embodiment in each following form, the same code | symbol as the above is attached | subjected and each detailed description is abbreviate | omitted suitably.

  FIG. 10 is a block diagram of a noise suppression device 100A according to the second embodiment. As illustrated in FIG. 10, the noise suppression device 100A has a configuration in which a signal sorting unit 52 and a weighting unit 54 are added to the noise suppression device 100 of the first embodiment. The signal classification unit 52 divides the acoustic signal VIN into a noise section and a target sound section on the time axis. A known technique is arbitrarily employed for selecting the noise section and the target sound section.

  The weighting unit 54 weights the kurtosis index value σm_1 and the kurtosis index value σm_2 calculated by the index calculation unit 42. That is, the weighting unit 54 multiplies the kurtosis index value σm_1 by the weight value w1 and multiplies the kurtosis index value σm_2 by the weight value w2. The weight value w1 and the weight value w2 are variably set according to the result of the classification by the signal classification unit 52. For example, the weight value w1 is set to a value larger than the weight value w2 in each frame FR in the target sound section, and the weight value w2 is set to a value larger than the weight value w1 in each frame FR in the noise section.

  In the state where the kurtosis index value σm_1 and the kurtosis index value σm_2 are close to each other, there is a possibility that both magnitudes are frequently reversed in a short time. Therefore, in the configuration of the first embodiment (a configuration in which the weighting unit 54 is omitted), when the first noise suppression processing is selected with a few frames FR in the noise section in which the second noise suppression processing is easily selected, There is a case where the second noise suppression process is selected in a few frames FR within the target sound section in which the first noise suppression process is easily selected. As described above, in the section in which the noise suppression processing is instantaneously changed, the reproduced sound of the acoustic signal VOUT may be audibly unnatural sound.

  In the second embodiment, since the weight value w1 of the kurtosis index value σm_1 and the weight value w2 of the kurtosis index value σm_2 are changed between the noise section and the target sound section, the kurtosis index value σm_1 and the kurtosis index Even in the state where the value σm_2 is close, the first noise suppression process is selected for each frame FR in the target sound section, and the second noise suppression process is selected for each frame FR in the noise section. That is, an instantaneous change in the noise suppression process is prevented. Therefore, compared with the first embodiment, it is possible to generate an acoustically natural reproduced sound.

<C: Modification>
Various modifications can be made to each of the forms exemplified above. An example of a specific modification is as follows. Two or more aspects may be arbitrarily selected from the following examples and combined.

(1) Modification 1
In each of the above embodiments, the spectral subtraction method is adopted as the first noise suppression processing and the MMSE-STSA method is exemplified as the second noise suppression processing. However, the types of noise suppression processing that are selection candidates by the selection unit 44 are as follows. It is not limited to the above illustration.

  For example, disclosed in Hack-Yoon KIM, et al., “Speech Enhancement Based on Short-Time spectral Amplitude Estimation with Two-Channel Beamformer”, IEICE Trans. Fundamentals, Vol. E79-A, No. 12, December 1996 Noise suppression processing (hereinafter referred to as “noise suppression processing A”), Ryo Mukai, “Residual noise removal after sound source separation by non-stationary spectral subtraction”, Acoustical Society of Japan 2001 Autumn Meeting, 2-6-14, p The noise suppression processing disclosed in 617-618 (hereinafter referred to as “noise suppression processing B”) is employed as the first noise suppression processing or the second noise suppression processing.

  In the noise suppression processing A, the target sound and noise are spatially obtained by adding (forming a beam with respect to the target sound direction) and subtracting (forming a blind spot with respect to the target sound direction) of the plurality of acoustic signals VIN generated by the plurality of sound collecting devices. Are separated (Griffith-Jim type adaptive beamformer), and the frequency spectrum Ym (n) after noise suppression is generated by subtracting the frequency spectrum of the noise from the frequency spectrum of the target sound. On the other hand, in the noise suppression processing B, a separated signal in which the target sound is emphasized is generated by independent component analysis of a plurality of acoustic signals VIN, and a frequency spectrum of noise (residual noise) estimated from the separated signal is obtained from the frequency spectrum of the separated signal. The frequency spectrum Ym (n) after noise suppression is generated by subtraction. When a plurality of acoustic signals VIN are used for noise suppression processing, the kurtosis KAm before noise suppression is calculated from any one of the plurality of acoustic signals VIN.

Although the subtraction type noise suppression process has been exemplified above, the content of the multiplication type noise suppression process is also changed as appropriate. For example, T. Lotter and P. Vary, "Speech enhancement by MAP spectral amplitude estimation using a Super-Gaussian speech model", EURASIP Journal on Applied Signal Processing, vol. 2005, no, 7, p.1110-1126, July 2005 The multiplication type noise suppression process using the MAP (maximum a posteriori estimation) estimation disclosed in the above is used for the estimation of the spectrum gain Gm (n) as the first noise suppression process or the second noise suppression process. Specifically, the spectrum gain Gm (n) is calculated by the calculation of the following formula (14). The coefficient φ and the coefficient τ in Expression (14) are constants (for example, τ = 2.5, φ = 1) that determine the shape of the noise probability distribution (probability density function).

  Further, the noise suppression processing as a selection candidate by the selection unit 44 is not limited to the subtraction type or the multiplication type. For example, a process of generating a frequency spectrum Ym (n) of an acoustic signal in which the target sound is emphasized (that is, an acoustic signal in which noise is suppressed) by independent component analysis for a plurality of acoustic signals VIN, or a beam is formed in the direction of the target sound. The processing (beamformer) that generates the frequency spectrum Ym (n) after noise suppression by forming a dead angle on sound collection in the direction of noise (or beam former) is also adopted as the first noise suppression processing or the second noise suppression processing. Is done.

  It should be noted that the principle of noise suppression is not necessarily different between a plurality of noise suppression processes that are selection candidates by the selection unit 44. For example, both the first noise suppression process and the second noise suppression process are the same type of subtraction type noise suppression process (formula (4a)), and a coefficient (for example, the coefficient αm of the formula (4a) A configuration is also adopted in which the coefficient βm) of the equation (4b) is different between the first noise suppression process and the second noise suppression process. In addition, both the first noise suppression process and the second noise suppression process are multiplication-type noise suppression processes, and a coefficient that affects noise suppression (for example, the coefficient η in Equation (6b)) is used as the first noise suppression process and the second noise suppression process. A configuration different from the noise suppression processing is also adopted. Since the coefficients (αm, βm, η) that can effectively reduce the musical noise after noise suppression change according to the acoustic characteristics of the acoustic signal VIN, they are of the same type (that is, the principle of noise suppression as described above). Even in a configuration in which a plurality of noise suppression processes are selected as candidates for selection by the selection unit 44, the intended effect of effectively reducing musical noise regardless of changes in the acoustic characteristics of the acoustic signal VIN Is realized. As can be understood from the above description, it is sufficient for the plurality of noise suppression processes to be selected by the selection unit 44 to be processes having different degrees of musical noise generated after execution, and the difference in the principle of noise suppression is as follows. It is unquestionable.

(2) Modification 2
In each of the above embodiments, the kurtosis KBm_1 is calculated from the frequency spectrum Ym (n) _1 actually generated by the first noise suppression processing. As described below, the estimated noise spectrum ψm (n) and the noise are calculated. A configuration for estimating the kurtosis KBm_1 from the frequency spectrum Xm (n) before suppression is also employed.

Now, when the subtractor 314 in FIG. 3 subtracts A times (A · ψm (n)) of the estimated noise spectrum ψm (n) from the frequency spectrum Xm (n) (the coefficient αm in equation (4a) is set to a predetermined value A). Assuming that the function Gb (x; k, θ) approximating the frequency distribution of the intensity of the acoustic signal VOUT after noise suppression is replaced with the intensity (x + A) in the equation (8) It is expressed by the following formula (15).

Similar to equation (10), equation (16) is derived by substituting function Gb (x; k, θ) of equation (15) into distribution function P (x) in the definition equation of fourth-order moment μ4. The

(X + A) ^{k-1 in} Expression (16) is Taylor-expanded as in Expression (17) below.

When the high-order term of the equation (17) is ignored for convenience and substituted into the equation (16), the following equation (18) that approximates the fourth-order moment μ4 is derived.

Similarly for the second moment, after substituting the function Gb (x; k, θ) of Equation (15) into the distribution function P (x) (see Equation (9)) of the definition equation, By ignoring the higher order terms, the following equation (19) is derived.

θ＝１／ｋ ……(21) Substituting the fourth-order moment μ4 of Equation (18) and the second-order moment μ2 of Equation (19) into Equation (7), the following Equation (20) that defines the kurtosis KBm_1 after noise suppression: Is derived. In order to derive the formula (20), the relationship of the following formula (21) derived by normalizing the average k · θ of the gamma function Γ (k) was used. The second kurtosis calculation unit 422 calculates the kurtosis KBm_1 (estimated value) by executing Expression (20). The predetermined value A in the equation (20) is appropriately selected so that a desired effect (noise suppression degree) is realized by the first noise suppression processing.

θ = 1 / k (21)

  As described above, it is not necessary to perform the first noise suppression process for calculating the kurtosis index value σm_1. Therefore, the first noise suppression unit 31 actually executes the first noise suppression process only for the frame FR for which the selection unit 44 has selected the first noise suppression process. According to the above configuration, since the first noise suppression process is omitted for the frame FR in which the selection unit 44 selects the second noise suppression process, the processing load of the first noise suppression unit 31 (the arithmetic processing device 12) is omitted. There is an advantage that is reduced.

(3) Modification 3
In each of the above embodiments, the noise suppression processing is selected for each frame FR using the kurtosis index value σm (σm_1, σm_2) of each frame FR. However, the cycle for calculating the kurtosis index value σm is set in the present invention. Is optional. For example, the acoustic signal VIN is divided into sections (hereinafter referred to as “unit sections”) composed of a plurality of frames FR, and the calculation of the kurtosis index value σm and the selection of the noise suppression processing are performed for each unit section. Adopted. That is, the index calculation unit 42 calculates the kurtosis index value σm (σm_1, σm_2) for the first frame FR of each unit section. The selection unit 44 selects the noise suppression process by comparing the kurtosis index value σm. The noise suppression processing selected by the selection unit 44 is maintained until a new kurtosis index value σm is calculated in the first frame FR of the next unit section. According to the above configuration, there is an advantage that the processing load by the index calculation unit 42 and the selection unit 44 is reduced. A configuration in which the kurtosis index value σm calculated for each frame FR is averaged over a plurality of frames FR (that is, a configuration in which temporal fluctuation of the kurtosis index value σm is smoothed) is also preferable.

(4) Modification 4
The method for calculating the kurtosis index value σm (σm_1, σm_2) is appropriately changed. For example, the degree of occurrence of musical noise after noise suppression tends to show a particularly significant correlation with the logarithmic value of the kurtosis ratio Rm (Rm_1, Rm_2). Therefore, a configuration for calculating the kurtosis index value σm from the logarithmic value of the kurtosis ratio Rm is also suitable. A configuration is also employed in which the kurtosis index value σm is calculated according to the kurtosis change amount (difference value) before and after the noise suppression processing. That is, the kurtosis index value σm_1 is set according to the difference value (KBm_1−KAm) between the kurtosis KBm_1 after execution of the first noise suppression process and the kurtosis KAm before execution, and the kurtosis index value σm_2 is It is set according to the difference value between the kurtosis KBm_2 after execution of the second noise suppression process and the kurtosis KAm before execution.

  In each of the above forms, the function Fa is used to calculate the kurtosis index value σm. However, the relative kurtosis ratio (Rm_1, Rm_2) and the change in kurtosis (KBm_1-KAm, KBm_2-KAm) A configuration in which the index value σm (σm_1, σm_2) is used is also suitable.

  Furthermore, the relationship between the magnitude of the kurtosis index value σm and the degree of musical noise (the degree of change in kurtosis κ) is changed. For example, when the function Fa in Expression (12) or Expression (13) is a monotone decreasing function, the smaller the kurtosis index value σm, the more musical noise after noise suppression (that is, the greater the change in kurtosis). Therefore, the selection unit 44 selects a noise suppression process corresponding to the larger one of the kurtosis index value σm_1 and the kurtosis index value σm_2. That is, the process by the selection unit 44 is a process of selecting one of a plurality of noise suppression processes according to each kurtosis index value σm (more preferably, the kurtosis index value σm of the plurality of noise suppression processes indicates This is included as a process for selecting a noise suppression process with a small change in kurtosis.

(5) Modification 5
In each of the above embodiments, the kurtosis index value σm is selected according to the degree of difference between the kurtosis KAm and the kurtosis KBm (KBm_1, KBm_2). However, since the kurtosis index value σm before noise suppression is common to the kurtosis index value σm_1 and the kurtosis index value σm_2, a configuration for calculating the kurtosis index value σm from the kurtosis index value mm after noise suppression (kurtosis index A configuration in which the kurtosis KAm is not used for calculating the value σm is also employed. That is, the index calculation unit 42 calculates the kurtosis index value σm_1 from the kurtosis KBm_1 and calculates the kurtosis index value σm_2 from the kurtosis KBm_2. The above configuration has an advantage that the configuration and processing of the index calculation unit 42 are simplified (specifically, the first kurtosis calculation unit 421 is omitted).

(6) Modification 6
A configuration is also preferable in which the coefficient αm of the equation (4a) in the first noise suppression processing of the first embodiment is variably controlled according to the kurtosis index value σm_1. For example, since the musical noise resulting from the first noise suppression processing increases as the coefficient αm increases, a configuration in which the coefficient αm is decreased as the kurtosis index value σm_1 increases. Similarly, the coefficient βm of Expression (4b) is also variably controlled according to the kurtosis index value σm_1.

(7) Modification 7
The number of types of noise suppression processing that are candidates for selection by the selection unit 44 is arbitrary. For example, in a configuration in which the suppression processing unit 30 performs each of three or more types of noise suppression processing in parallel with the acoustic signal VIN, the index calculation unit 42 performs the kurtosis index value σm (three or more) for each noise suppression process. And the selection unit 44 selects one of three or more types of noise suppression processing. Further, when three or more types of noise suppression processes are selected as selection candidates, a configuration in which the selection unit 44 selects two or more noise suppression processes is also employed. The two or more frequency spectra Ym (n) generated by the two or more noise suppression processes selected by the selection unit 44 are output to the waveform synthesis unit 46 after being mixed, for example.

DESCRIPTION OF SYMBOLS 100 ... Noise suppression apparatus, 12 ... Arithmetic processing apparatus, 14 ... Memory | storage device, 20 ... Frequency analysis part, 30 ... Suppression processing part, 31 ... 1st noise suppression part, 312 ... Noise estimation part, 314: Subtraction unit, 32: Second noise suppression unit, 322: Noise estimation unit, 324: Gain calculation unit, 326: Multiplication unit, 42: Indicator calculation unit, 421: First kurtosis calculation , 422... Second kurtosis calculation unit, 423... Third kurtosis calculation unit, 425... First index calculation unit, 426... Second index calculation unit, 44. Synthesis department.

Claims (5)

A plurality of noise suppression means for performing different noise suppression processing on the acoustic signal;
Index calculation means for calculating the kurtosis index value indicating the degree to which the kurtosis in the frequency distribution of the intensity of the acoustic signal has changed before and after the noise suppression process for each noise suppression process by the noise suppression means;
A noise suppression apparatus comprising: selection means for selecting one of a plurality of noise suppression processes by each noise suppression means according to each kurtosis index value calculated by the index calculation means. The noise suppression apparatus according to claim 1, wherein the selection unit selects a noise suppression process in which a change in kurtosis indicated by the kurtosis index value is small among the plurality of noise suppression processes. A signal dividing means for dividing the acoustic signal into a noise section and a target sound section on a time axis;
Weighting means for weighting each kurtosis index value by changing a weighting value for each noise suppression process between a noise interval and a target sound interval;
The noise suppression apparatus according to claim 1, wherein the selection unit selects a noise suppression process according to each weighted kurtosis index value. The plurality of noise suppression processes include a subtraction-type noise suppression process for subtracting a noise spectrum from the spectrum of the acoustic signal, and a multiplication-type noise suppression process for multiplying the spectrum of the acoustic signal by a spectrum gain that enhances the target sound. The noise suppression device according to any one of claims 1 to 3. A plurality of noise suppression processes individually performed on the acoustic signal;
An index calculation process for calculating, for each noise suppression process, a kurtosis index value indicating the degree to which the kurtosis in the frequency distribution of the intensity of the acoustic signal has changed before and after the noise suppression process;
A program that causes a computer to execute a selection process that selects any one of the plurality of noise suppression processes in accordance with each kurtosis change index calculated in the index calculation process.

Images