Noise estimation device, noise estimation program, noise estimation method, sound collection device, sound collection program, and sound collection method

Abstract

To estimate noise power without being affected by abrupt change in an input voice without adjusting an estimation parameter.SOLUTION: The present invention relates to a noise estimation device for estimating a noise component included in an input signal. The noise estimation device includes: stationary probability calculation means for calculating steady probability of input power; and environment shift determination means for calculating environment shift determination results that determines whether a ratio of a value based on estimated noise power or a value based on the estimated noise power to average input power is within a predetermined stationary determination range. When determining that environment is shifted, the stationary probability calculation means makes stationary probability a predetermined value. When determining that environment is not shifted, the stationary probability calculation means calculates stationary probability by a stationary probability calculation function for calculating the steady probability based on a posterior SNR obtained by dividing input power by the past estimated noise power.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a noise estimation device, a noise estimation program, a noise estimation method, a sound collection device, a sound collection program, and a sound collection method, and can be applied, for example, to a method for suppressing a noise component superimposed on an input signal by using an estimation result of a noise component contained in the input signal. [Background technology]

[0002] Since noise is ubiquitous in the natural environment, speech observed in the real world generally contains noise from various sources. In order to emphasize only speech from an input signal observed containing noise, various noise suppression methods have been developed. Most of these include a method of estimating the noise to be suppressed and a method of calculating a filter to suppress the noise. In a conventional speech processing device for suppressing noise from an input signal, there is a device that estimates the power of noise in the frequency domain.

[0003] Conventionally, the simplest noise estimation method is a method of averaging the input spectrum in sections where no speech exists. However, such conventional noise estimation methods require estimating sections where no speech exists in advance. For this reason, a technology called Voice Activity Detection (VAD) that estimates sections where speech exists has been actively developed, but perfect VAD has not yet been achieved. If the speech section is estimated incorrectly in the noise estimation process, the estimated noise will contain the target speech, causing a problem of distorting the emphasized speech and residual noise. In addition, the above-mentioned noise estimation method has the disadvantage that it cannot track changes in noise if there is a long speech section because it estimates noise only in the noise section.

[0004] In light of this background, conventional noise estimation methods that continue estimating noise even during speech segments include the techniques described in Non-Patent Document 1, Non-Patent Document 2, and Patent Document 1. All of these documents relate to noise suppression methods (also called speech enhancement methods).

[0005] The conventional noise estimation method described in Non-Patent Document 1 is based on the discovery that the peaks of the input power in the time direction indicate the presence of the target voice, while the valleys can be used to estimate the smoothed noise power. Specifically, the minimum value of the input power from the present to a certain time in the past is taken as the first estimated noise power. However, the first estimated noise power has a bias and tends to be smaller than the true noise power. This bias is estimated from the expected value of the first estimated noise power, and the first estimated noise power is corrected using the obtained bias estimate to obtain the second estimated noise power (final estimate).

[0006] The conventional noise estimation method described in Patent Document 1 multiplies the input power by an appropriate weighting factor, stores the weighted input power obtained for a predetermined time, and sets the average value of the stored weighted input power as the estimated noise power. The appropriate weighting factor is calculated based on the posterior SNR (SNR is Signal-to-Noise Ratio) obtained by dividing the current input power by the immediately preceding estimated noise power. Specifically, when the posterior SNR is equal to or less than a predetermined value G1, the weighting factor is set to be inversely proportional to the posterior SNR when the posterior SNR is equal to or more than G1, and the weighting factor is set to be 0 when the posterior SNR is equal to or more than a predetermined value G2. When the weighting factor is 0, the weighted input power is not stored.

[0007] The conventional noise estimation method described in Non-Patent Document 2 uses the maximum likelihood estimate of the variance of the noise complex spectrum as the estimated noise power, based on the hypothesis that the distributions of the complex spectra of the target speech and noise both follow a complex normal distribution with a mean of zero. Based on this hypothesis, the distribution of the complex spectrum of the input speech is a complex normal distribution with a mean of zero, whose variance is the sum of the variance of the complex spectrum of the speech and the variance of the complex spectrum of the noise. A hidden variable indicating whether the current input is degraded speech or noise is introduced here, and an online EM (Expectation Maximization) algorithm with a forgetting factor is applied to calculate the maximum likelihood estimate of the noise complex spectrum. [Prior art documents] [Patent documents]

[0008] [Patent Document 1] JP 2002-204175 A [Non-patent literature]

[0009] [Non-Patent Document 1] R. Martin, “Spectral Subtraction Based on Minimum Statistics”, in Proceedings of 7th European Signal Processing Conference, 1994, pp. 1182-1185. [Non-Patent Document 2] M.Souden, M.Delcroix, K.Kinsoshita, T.Yoshioka, andT.Nakatani, “Noise Power Spectral Density Tracking: A Maximum Likelihood Perspective”, IEEE Signal Processing Letters, Vol.19, No.8, 2012, pp.495-498 Summary of the Invention [Problem to be solved by the invention]

[0010] However, the conventional noise estimation method has the following problems.

[0011] The method of Non-Patent Document 1 has a problem that when the noise suddenly increases, the estimated noise power increases suddenly after a delay. Specifically, the estimated noise power remains small for a predetermined time after the noise increases. Then, after the predetermined time has passed since the noise increased, the estimated noise power increases instantaneously.

[0012] The method of Patent Document 1 has a problem that when the input voice temporarily becomes small due to packet loss in voice transmission or a voice switch for echo countermeasures, the estimated noise power remains small for a while even if the input voice returns to normal. That is, the small input voice continues to be used for estimating the noise power because the posterior SNR is smaller than G1, and the estimated noise power becomes small. If the input voice becomes large in this state, the posterior SNR becomes larger than G2, so it is no longer used for estimating the noise power, and the estimated noise power is no longer updated.

[0013] The method of Non-Patent Document 2 has a problem that the online EM algorithm used in this noise estimation method has a trade-off between the tracking speed and the stability of maximum likelihood estimation, in that increasing the forgetting factor increases stability and slows tracking, and decreasing the forgetting factor increases tracking speed and decreases stability. The forgetting factor needs to be set according to the noise level of the observation environment, making this method less practical.

[0014] As described above, conventional noise estimation methods have problems such as the estimated noise power momentarily increasing at an inappropriate timing when the input speech suddenly increases, the estimated noise power remaining small when the input speech temporarily decreases, and the need to adjust the estimation parameters to suit the environment.

[0015] Therefore, there is a demand for a noise estimation device, a noise estimation program, a noise estimation method, a sound collection device, a sound collection program, and a sound collection method that can estimate noise power without adjusting estimation parameters and without being affected by sudden changes in input sound. [Means for solving the problem]

[0016] The first invention of the present invention is a noise estimation device for estimating a noise component contained in an input signal in which speech and noise are mixed to obtain an estimated noise power, the noise estimation device comprising: (1) a stationary probability calculation means for calculating a stationary probability of the input power from the input power of the input signal and the past estimated noise power; (2) a buffer update means for updating an input buffer for holding the input power for a predetermined first number of samples and a probability buffer for holding the stationary probability for the first number of samples only when the stationary probability is greater than a predetermined stationary probability threshold; and (3) a weighted averaging means for calculating the estimated noise power by taking a weighted average of the input power held in the input buffer using the stationary probability held in the probability buffer as a weighting coefficient. (4) an input averaging means for calculating an average input power by retaining and averaging a predetermined second number of samples of the input power; and (5) an environment transition determination means for calculating an environment transition determination result by determining whether a ratio of the estimated noise power or a value based on the estimated noise power to the average input power falls within a predetermined steady-state determination range, and (6) the steady-state probability calculation means is characterized in that, when the environment transition determination means determines that the environment has transitioned, the steady-state probability is set to a predetermined value, and when the environment transition determination means determines that the environment has not transitioned, the steady-state probability calculation means calculates the steady-state probability using a steady-state probability calculation function that calculates the steady-state probability based on a posterior SNR obtained by dividing the input power by the past estimated noise power.

[0017] A noise estimation program according to a second aspect of the present invention includes a computer installed in a noise estimation device that estimates a noise component contained in an input signal in which speech and noise are mixed to obtain an estimated noise power, the computer comprising: (1) a stationary probability calculation means that calculates a stationary probability of the input power from the input power of the input signal and the past estimated noise power; (2) a buffer update means that updates an input buffer that holds the input power for a predetermined first number of samples and a probability buffer that holds the stationary probability for the first number of samples only when the stationary probability is greater than a predetermined stationary probability threshold; and (3) a weighted average of the input power held in the input buffer using the stationary probability held in the probability buffer as a weighting coefficient to calculate the estimated noise power. (4) an input averaging means for calculating an average input power by retaining and averaging a predetermined second number of samples of the input power; (5) an environmental transition determination means for calculating an environmental transition determination result by determining whether a ratio of the estimated noise power or a value based on the estimated noise power to the average input power falls within a predetermined steady-state determination range; and (6) the steady-state probability calculation means calculates the steady-state probability by a steady-state probability calculation function that sets the steady-state probability to a predetermined value when the environmental transition determination means determines that the environment has transitioned, and calculates the steady-state probability based on a posterior SNR obtained by dividing the input power by the past estimated noise power when the environmental transition determination means determines that the environment has not transitioned.

[0018] The third aspect of the present invention is a noise estimation method performed by a noise estimation device that estimates a noise component contained in an input signal in which speech and noise are mixed to obtain an estimated noise power, the noise estimation device comprising: (1) a stationary probability calculation means, a buffer update means, a weighted average means, an input averaging means, and an environment transition determination means; (2) the stationary probability calculation means calculates a stationary probability of the input power from the input power of the input signal and the past estimated noise power; (3) the buffer update means updates an input buffer that holds the input power for a predetermined first number of samples and a probability buffer that holds the stationary probability for the first number of samples only when the stationary probability is greater than a predetermined stationary probability threshold; and (4) the weighted average means updates the stationary probability held in the probability buffer as a weighting coefficient and stores the stationary probability in the input buffer. (5) the input averaging means calculates an average input power by retaining and averaging a predetermined second number of samples of the input power; (6) the environment transition determination means calculates an environment transition determination result by determining whether a ratio of the estimated noise power or a value based on the estimated noise power to the average input power falls within a predetermined steady-state determination range; and (7) the steady-state probability calculation means sets the steady-state probability to a predetermined value when the environment transition determination means determines that the environment has transitioned, and calculates the steady-state probability by a steady-state probability calculation function that calculates the steady-state probability based on a posterior SNR obtained by dividing the input power by the past estimated noise power when the environment transition determination means determines that the environment has not transitioned.

[0019] A fourth aspect of the present invention is a sound collection device that collects sound from an input signal in which sound and noise are mixed, comprising: (1) a noise estimation unit that estimates a noise component contained in the input signal; and (2) a sound collection unit that extracts the sound signal from the input signal by using an estimation result of the noise estimation unit and collects the sound, wherein (3) the noise estimation device of the first aspect of the present invention is applied as the noise estimation unit.

[0020] A noise estimation program according to a fifth aspect of the present invention causes a computer mounted on a sound collection device that collects sound from an input signal containing a mixture of sound and noise to function as (1) a noise estimation unit that estimates noise components contained in the input signal, and (2) a sound collection unit that extracts and collects the sound signal from the input signal using an estimation result from the noise estimation unit, and (3) the noise estimation unit functions as the noise estimation device according to the first aspect of the present invention.

[0021] The sixth aspect of the present invention is a sound collection method performed by a sound collection device that collects sound from an input signal containing a mixture of sound and noise, characterized in that (1) it is provided with a noise estimation unit and a sound collection unit, (2) the noise estimation unit estimates a noise component contained in the input signal by the noise estimation method of the third aspect of the present invention, and (3) the sound collection unit extracts the sound signal from the input signal and collects the sound using the estimation result of the noise estimation unit. Effect of the Invention

[0022] According to the present invention, it is possible to estimate noise power without adjusting estimation parameters and without being affected by sudden changes in the input speech. [Brief description of the drawings]

[0023] [Figure 1] FIG. 2 is a block diagram showing a functional configuration of a noise estimation means according to the first embodiment. [Diagram 2] 1 is a block diagram showing a functional configuration of a sound collecting device according to a first embodiment. [Diagram 3] 1 is a block diagram showing a hardware configuration of a sound collecting device according to a first embodiment. [Figure 4] 1 is a block diagram showing a functional configuration of a noise estimating unit (noise estimating device) according to a first embodiment. [Diagram 5] 4 is a diagram (graph) showing an example of a stationary probability calculation function according to the first embodiment. FIG. [Figure 6] 1 is a diagram (graph) showing an example in which a stationary probability calculation function according to the first embodiment is designed on a logarithmic scale. [Figure 7] FIG. 13 is a block diagram showing a functional configuration of a noise suppressor (noise estimation device) according to a third embodiment. [Figure 8] FIG. 11 is a block diagram showing a functional configuration of a noise estimation means according to a third embodiment. [Figure 9] FIG. 13 is a diagram (graph) showing an example of a stationary probability calculation function according to the third embodiment. [Figure 10] FIG. 13 is a diagram (graph) showing an example in which a stationary probability calculation function according to the third embodiment is designed on a logarithmic scale. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0024] (A) First embodiment Hereinafter, a first embodiment of a noise estimation device, a noise estimation program, a noise estimation method, a sound collection device, a sound collection program, and a sound collection method according to the present invention will be described in detail with reference to the drawings. In this embodiment, an example in which the noise estimation device, the noise estimation program, and the noise estimation method of the present invention are applied to a noise estimation means will be described.

[0025] (A-1) Configuration of the First Embodiment Fig. 2 is a block diagram showing the functional configuration of the sound collection device 1 of this embodiment. Note that the reference characters in parentheses in Fig. 2 are the same as those used in the second to fourth embodiments described later.

[0026] The sound collection device 1 is a device that performs a sound collection process to collect a target sound from an acoustic signal captured by a microphone M.

[0027] In this embodiment, the microphone M is mounted on a receiver of a telephone terminal (not shown). In this case, the acoustic signal captured by the microphone M includes, for example, a voice as a target sound (e.g., a voice of a near-end talker) and noise as a non-target sound (e.g., background noise, etc.). In this embodiment, the sound collection device 1 is mounted on a telephone terminal (not shown) and performs a process of removing non-target sounds (e.g., noise such as background noise) from the acoustic signal captured by the microphone M and collecting the target sound (e.g., a voice of a near-end talker).

[0028] Next, the internal configuration of the sound collection device 1 will be described.

[0029] In this embodiment, the sound collection device 1 includes a signal input unit 10, a noise suppression processing unit 20, a signal output unit 30, and a noise estimation unit .

[0030] The sound collection device 1 may be configured entirely of hardware (for example, a dedicated chip, etc.), or may be configured partially or entirely as software (programs). The sound collection device 1 may be configured, for example, by installing a program (a sound collection program including the noise estimation program of the embodiment) in a computer having a processor and memory.

[0031] The signal input unit 10 has a function of converting an analog acoustic signal supplied from the microphone M into a digital signal and supplying it to the computer 200. In the following, the acoustic signal captured by the microphone M and digitally converted by the signal input unit 10 will be referred to as an input signal x.

[0032] The noise estimation unit 40 has a function of estimating noise (non-target sound) included in the input signal x. Hereinafter, the noise estimated by the noise estimation unit 40 (estimated noise signal) will be referred to as “estimated noise.”

[0033] The noise suppression processing unit 20 has a function of using the estimated noise estimated by the noise estimation unit 40 to output a signal in which the noise component contained in the input signal x is suppressed (hereinafter, referred to as a “noise-suppressed signal”).

[0034] The signal output unit 30 has a function of outputting the sound collection result of the sound collection device 1 (in this embodiment, the noise suppressed signal output by the noise suppression processing unit 20).

[0035] Fig. 3 is a block diagram showing an example of the hardware configuration of the sound collection device 1. Note that the reference characters in parentheses in Fig. 3 are reference characters used in the second embodiment described later.

[0036] Fig. 3 shows a configuration in which the sound collection device 1 is configured using software (computer). Note that in Fig. 3, the reference characters in parentheses are reference characters used in the second to fourth embodiments described later.

[0037] The sound collection device 1 shown in FIG. 3 has, as hardware components, at least a signal input unit 10 and a computer 400 in which a program (a sound collection program including the noise estimation program of the embodiment) is installed.

[0038] The signal input unit 10 can be configured using, for example, a D / A converter. If the computer 400 itself is equipped with a D / A converter, there is no need to provide a separate signal input unit 10.

[0039] The computer 400 performs a process of performing a predetermined process on the acoustic signal (digital acoustic signal) supplied from the signal input unit 10 and outputting the processed signal. In this embodiment, it is assumed that the computer 400 has installed therein at least a program (a sound collection program in this embodiment) corresponding to the noise suppression processing unit 20, the signal output unit 30, and the noise estimation unit 40. Note that the sound collection program in this embodiment includes a noise estimation program corresponding to the noise estimation unit 40.

[0040] The computer 400 may be a computer dedicated to the sound collection program, or may be configured to be shared with programs for other functions (for example, a function for outputting a far-end signal (received signal) received by a telephone terminal from a speaker (not shown)).

[0041] The computer 400 shown in FIG. 3 includes a processor 401, a primary storage unit 402, and a secondary storage unit 403. The primary storage unit 402 is a storage unit that functions as a working memory (work memory) for the processor 401, and is implemented, for example, by a high-speed memory such as a DRAM. The secondary storage unit 403 is a storage unit that records various data such as an OS (Operating System) and program data (including data of the sound collection program according to the embodiment), and is implemented, for example, by a non-volatile memory such as a FLASH (registered trademark) memory or HDD. In the computer 400 of this embodiment, when the processor 401 starts up, the OS and programs (including the sound collection program according to the embodiment) recorded in the secondary storage unit 403 are read, deployed on the primary storage unit 402, and executed.

[0042] Note that the specific configuration of the computer 400 is not limited to the configuration in Fig. 3, and various configurations can be applied. For example, if the primary storage unit 402 is a non-volatile memory (e.g., a flash memory, etc.), the secondary memory may be excluded.

[0043] Next, the internal configuration of the noise estimation unit 40 will be described with reference to FIG.

[0044] Fig. 4 is a block diagram showing the functional configuration of the noise estimation unit 40. Note that the reference characters in parentheses in Fig. 4 are reference characters used in the second embodiment described later.

[0045] The noise estimation unit 40 includes a band division means 41, K power calculation means 42 (42-1 to 42-K), and K noise estimation means 43 (43-1 to 43-K).

[0046] Fig. 1 is an explanatory diagram showing the internal configuration of each of the noise estimation means 43 (43-1 to 43-K). In this embodiment, the inside of each of the noise estimation means 43-1 to 43-K is assumed to have a configuration that can be shown in Fig. 1. Note that the reference characters in parentheses in Fig. 1 are reference characters used in the second embodiment described later.

[0047] As shown in FIG. 1, the noise estimation means 43 includes a stationary probability calculation means 101 , a buffer update means 102 , a weighted average means 103 , an input average means 104 , and an environment transition determination means 105 .

[0048] The detailed functions (operations) of each element constituting the noise estimation unit 40 (including each element constituting the noise estimation means 43) will be described later.

[0049] (A-2) Operation of the First Embodiment Next, the operation of the sound collection device 1 (noise estimation unit 40) of the first embodiment having the above-mentioned configuration will be described with reference to FIG.

[0050] The band division means 41 performs frequency analysis on the input signal x to calculate a frequency spectrum (hereinafter also referred to as an "input spectrum"), divides the obtained input spectrum into K parts, and provides the divided input spectra (hereinafter referred to as "frequency band signals") to power calculation means 42-1 to 42-K (hereinafter collectively referred to as power calculation means 42). For example, a fast Fourier transform (FFT), a wavelet transform, a filter bank, or the like can be applied to the frequency analysis, but an FFT is preferable.

[0051] The input signal x is, for example, an acoustic signal captured by a microphone of a telephone terminal. The acoustic signal captured by the microphone of the telephone terminal includes, for example, a speaker's voice as a target sound and noise (for example, background noise, etc.) as a non-target sound.

[0052] The power calculation means 42-1 to 42-K calculate input power based on the input frequency band signal, and provide it to the noise estimation means 43-1 to 43-K. Each power calculation means 42 can apply various calculation methods as a power calculation method. Each power calculation means 42 may calculate, for example, the square of the absolute value or the absolute value as the input power. Hereinafter, the input power output by each power calculation means 42 (42-1 to 42-K) is represented as PX (PX_1 to PX_K).

[0053] Each noise estimation means 43 (43-1 to 43-K) estimates the power of the noise component contained in the input power PX (PX_1 to PX_K) supplied from each power calculation means 42 (21 to 2K) and outputs the result (hereinafter also referred to as "estimated noise power"). Hereinafter, the estimated noise power output from each noise estimation means 43 (43-1 to 43-K) is represented as PN (PN_1 to PN_K).

[0054] The method (e.g., signal format or data format) and output destination of each noise estimation means 43 (43-1 to 43-K) when outputting the estimated noise power PN (PN_1 to PN_K) are not limited. For example, the noise estimation means 43 (43-1 to 43-K) may supply the estimated noise power PN (PN_1 to PN_K) to a voice processing method (not shown) that emphasizes the target sound of the input power PX (suppresses noise). Furthermore, the sound collection device 1 may be configured as a part of the voice processing method described above.

[0055] Next, the operation of each of the noise estimation means 43 (43-1 to 43-K) will be described with reference to FIG.

[0056] In the following, the operations of the noise estimation means 43-1 to 43-K are similar except for the band of the signal to be processed, so below we will explain the operation when input power PX (any of PN_1 to PN_K) of any band is supplied to any noise estimation means 43 (any of the noise estimation means 43-1 to 43-K).

[0057] The noise estimation means 43 provides the given input power PX to the stationary probability calculation means 101 , the buffer update means 102 , and the input averaging means 104 .

[0058] The stationary probability calculation means 101 calculates the stationary probability SP based on the input power PX, the estimated noise power PND for a predetermined D samples in the past, and an environment transition determination result E, which will be described later, and provides the obtained stationary probability SP to the buffer update means 102. A specific value of D will also be described later.

[0059] The details of the environment transition determination result E will be described later in the explanation of the operation of the environment transition determination means 105, but in this embodiment example, if the environment has transitioned significantly (a state in which the background noise or the input signal itself has become significantly quieter or louder continues for a long period of time), it is determined that the environment has transitioned and True is set, and if not, it is determined that the environment has not transitioned and False is set.

[0060] The predetermined number of delay samples D is preferably, for example, D=1, but may be D>1. Also, the predetermined number of delay samples D is preferably determined in advance, but may be changed during the process.

[0061] In this embodiment, when the environmental transition determination result E is True, the stationary probability SP is set to a predetermined value of 1.0. Any value can be set as the predetermined value of SP. In addition, in this embodiment, when the environmental transition determination result E is False, the stationary probability SP is calculated based on the posterior SNR G. Furthermore, in this embodiment, the posterior SNR is calculated by equation (1). The stationary probability SP is calculated by a stationary probability calculation function described later.

number

[0062] In this embodiment, the stationary probability calculation function is a function of the stationary probability SP for the posterior SNR G, and is set in advance in the stationary probability calculation means 101. In addition, in this embodiment, the domain (G) is a real number equal to or greater than 0.0, and the range (SP) is a real number equal to or greater than 0.0 and equal to or less than 1.0. Furthermore, in this embodiment, the shape of the stationary probability calculation function is a mountain shape, that is, for a predetermined G value G0, the stationary probability calculation function increases monotonically in a wide sense in the range 0.0≦G≦G0, and decreases monotonically in a wide sense in the range G0≦G.

[0063] In this embodiment, G0 is not particularly limited and can take any value within the range of the domain of G. However, in this embodiment, since the stationary probability calculation function is a function that outputs a higher probability as it becomes more stationary, it is most preferable to set G0=1.0 (i.e., PX=PND).

[0064] Examples of stationary probability calculation functions are shown in Figs.

[0065] FIG. 5 shows an example of a stationary probability calculation function smoothly defined for G.

[0066] 5, G(G0) at which SP=1.0 is not particularly limited, but may be set to, for example, G=1.0. It is preferable that SP asymptotically converges to 0.0 as G→+∞.

[0067] Figure 4 shows a stationary probability calculation function designed on a logarithmic scale.

[0068] As shown in Figure 4, by designing a stationary probability calculation function on a logarithmic scale for both axes, PX<PNDのときとPX> The relationship at the time of PND becomes fair, and the stationary probability calculation function can be easily defined over the entire domain.

[0069] When the stationary probability SP is greater than a predetermined stationary probability threshold SP0, the buffer update means 102 updates the input buffer BX and the probability buffer BP based on the input power PX and the stationary probability SP, and provides the obtained input buffer BX and probability buffer BP to the weighted averaging means 103.

[0070] Moreover, in the buffer update means 102, the stationary probability threshold SP0 must be set to a sufficiently small value (for example, about 0.1 or less). The reason for this will be described later in the explanation of the weighted average means 103. For example, a value of about 0.01 is suitable for the stationary probability threshold SP0.

[0071] The input buffer BX is a buffer that holds T1 samples of the input power PX from the past to the present. The first number of samples T1 may be any value, but is set to a length equivalent to 100 milliseconds to several seconds depending on the application. If the first number of samples T1 is set to a short value (e.g., equivalent to 100 milliseconds), the tracking of changes in the noise of the estimation target becomes faster, but there is a risk that, for example, a voice that is spoken slowly may be erroneously estimated as noise. On the other hand, if the first number of samples T1 is set to a long value (e.g., equivalent to 8 seconds), the stationary components in the input signal can be accurately estimated, but if the noise environment changes suddenly (e.g., when going from a quiet room to a noisy outdoor area, when air conditioning is turned on / off all at once, etc.), it is not possible to immediately track them. Therefore, it is preferable that the first number of samples T1 is set to a length equivalent to 200 milliseconds to 1 second.

[0072] The probability buffer BP is a buffer that holds T1 samples of stationary probability SP from the past to the present. The buffer length of the probability buffer BP is the same as the buffer length of the input buffer BX.

[0073] The buffer update means 102 deletes the oldest input power PX from the input buffer BX and stores the newly given input power PX in the input buffer BX. Similarly, the buffer update means 102 deletes the oldest stationary probability SP from the probability buffer BP and stores the newly given stationary probability SP in the probability buffer BP. Note that if the stationary probability SP is greater than the stationary probability threshold SP0, the input buffer BX and the probability buffer BP are not updated.

[0074] The weighted average means 103 calculates an estimated noise power PN based on the input buffer BX and the probability buffer BP, and provides the obtained estimated noise power PN to the outputs of the stationary probability calculation means 101 , the environmental transition determination means 105 and the noise estimation means 100 .

[0075] The weighted average means 103 calculates the weighted average of the input power PX held in the input buffer BX using the stationary probability SP held in the probability buffer BP as a weighting coefficient to calculate the estimated noise power PN. That is, if the i-th (i=1 to T1) values ​​of the input buffer BX and the probability buffer BP are the input power PX_i and the stationary probability SP_i, respectively, the estimated noise power PN is calculated by equation (2).

number

[0076] The method of calculating the estimated noise power PN by weighted averaging has the following advantages. That is, since a large weight is applied to the input power PX with a high stationary probability SP, it has a large influence on the calculation of the estimated noise power PN. On the other hand, since a small weight is applied to the input power PX with a low stationary probability SP, it has a small influence on the calculation of the estimated noise power PN and is almost ignored. If the input power PX with a small stationary probability SP is completely ignored in the noise estimation unit 40 (for example, in Patent Document 1, it is ignored if the posterior SNR is greater than a predetermined value), a problem occurs in which the input signal cannot be tracked when it returns after being temporarily reduced due to packet loss or voice switching. Therefore, the noise estimation unit 40 can continue stable estimation by considering the input power PX with a small stationary probability SP with a small influence.

[0077] Also, it has been stated that the stationary probability threshold SP0 must be a small value in the buffer update means 102. If SP0 is set to a large value, the merit of the present invention of continuing the above-mentioned stable estimation is lost.

[0078] While the environment transition determination means 105 determines that the environment is transitioning, a sufficiently small value (for example, about 0.1 or less) is continuously set for the stationary probability SP_k, so that the estimated noise power PN calculated according to equation (2) becomes a simple average of the input power PX, and it can quickly follow the environment after the transition.

[0079] The input averaging means 104 holds and averages the input power PX for T2 samples from the past to the present, and gives the obtained average input power APX to the environment transition determination means 105. The value of the second number of samples T2 will be described later.

[0080] The environment transition determination means 105 determines whether the ratio of the average input power APX to the estimated noise power PN falls within a predetermined stationary determination range R, sets the environment transition determination result E, and gives the obtained environment transition determination result E to the stationary probability calculation means 101.

[0081] There are two conceivable processing methods (definition method and its usage method) for the stationary determination range R.

[0082] The first processing method for the stationary determination range R is a method of defining the stationary determination range as a vector of two positive values (R1 < R2) and calculating the environment transition determination result E using equation (3).

[0083] The second processing method for the stationary determination range R is a method of defining the stationary determination range as a scalar of one positive value and calculating the environment transition determination result E using equation (4). In this case, R is on the decibel scale.

Equation

[0084] In the environment transition determination means 105, either of the above two processing methods can be used without problems, but from the viewpoint of simplicity of setting, it is preferable to use the second processing method. In that case, the stationary determination range R is preferably set to 6 decibels, for example.

[0085] The role of the environment transition determination means 105 is to make the estimated noise power PN quickly follow the input power PX with some delay when the characteristics of the input signal X change significantly (for example, when there is complete silence or when an abnormally loud volume is input due to an incorrect connection of an audio device). When the characteristics of the input signal X change significantly and the input power PX changes significantly, the stationary probability SP continues to take a small value, and a deadlock state occurs in which the input buffer BX and the probability buffer BP are not updated. To avoid this problem, the environment transition determination means 105 calculates the long-term average of the input power PX. Since the proportion of non-stationary components (called target components) that are not the target of noise suppression in the time direction is often small, the influence of the target components can be removed by taking the long-term average. Therefore, the environment transition determination means 105 can detect the transition of the environment by comparing this long-term average with the estimated noise power PN, which is an estimated value of the average of the noise power over a relatively short period of time, as the average input power APX. Therefore, the second sample number T2 for obtaining the average input power APX in the environmental transition determination means 105 is preferably, for example, 5 to 20 seconds. Note that, although there is a method unrelated to the present invention in which the average input power APX is set as the estimated noise power PN, this is not a preferred method since it significantly impairs the ability to follow small changes in the environment.

[0086] (A-3) Effects of the First Embodiment According to the first embodiment, the following effects can be obtained.

[0087] In the first embodiment of the sound collection device 1 (noise estimation means 43A), the parameters required for noise estimation are only the first sample number T1, the second sample number T2, and the delay sample number D, so it is possible to provide a noise estimation method that does not require adjustment of the estimation parameters.

[0088] In addition, in the first embodiment of the sound collection device 1 (noise estimation means 43A), even if the noise suddenly increases, a weighted average with a stationary probability as a weighting coefficient is used, so that the noise is tracked little by little using small noise power that appears with a low frequency among the noise power that changes probabilistically, and therefore the noise does not increase suddenly.

[0089] Furthermore, in the first embodiment of the sound collection device 1 (noise estimation means 43A), if the input sound suddenly becomes quieter, it is considered to be unsteady, so there is no excessive tracking of the temporarily quieter input sound, and the problem of the estimated noise power remaining small does not occur.

[0090] Furthermore, since the sound collection device 1 (noise estimation means 43A) of the first embodiment aims to estimate only stationary noise components contained in the input sound, there is no need to adjust the estimation parameters in accordance with the environment.

[0091] In addition, in the first embodiment of the sound collection device 1 (noise estimation means 43A), when the environment changes and the estimated noise power becomes a value that deviates from the average input power, the stationary probability is set to a predetermined value (e.g., 1.0) for the first sample number, thereby quickly tracking the new environment.

[0092] As described above, in the first embodiment of the sound collection device 1 (noise estimation means 43A), the noise power is estimated using a weighted average with the stationary probability as the weighting coefficient. Therefore, even if the noise changes suddenly, the estimated noise power does not change suddenly, and furthermore, the estimated noise power does not remain small or large. Therefore, even if the environment changes significantly, a deadlock state does not occur, and a noise estimation method capable of stable estimation can be provided.

[0093] (B) Second embodiment Hereinafter, a second embodiment of the noise estimation device, the noise estimation program, the noise estimation method, the sound collection device, the sound collection program, and the sound collection method according to the present invention will be described in detail with reference to the drawings. In this embodiment, an example in which the noise estimation device, the noise estimation program, and the noise estimation method of the present invention are applied to a noise estimation means will be described.

[0094] (B-1) Configuration of the second embodiment The configuration of the sound collection device 1A of the second embodiment can also be shown using the above-mentioned FIGS.

[0095] The following describes the configuration of a sound collection device 1A of the second embodiment, focusing on differences from the first embodiment.

[0096] In the sound collection device 1A of the second embodiment, the noise estimation unit 40 is replaced with a noise estimation unit 40A. Also, in the noise estimation unit 40A of the second embodiment, the noise estimation means 43 (43-1 to 43-K) are replaced with noise estimation means 43A (43A-1 to 43A-K). Furthermore, in the noise estimation units 40A (43A-1 to 43A-K) of the second embodiment, the weighted average means 103 is replaced with a weighted average means 103A.

[0097] In the noise estimation unit 40 of the first embodiment, the estimation bias of the estimated noise power PN is not taken into consideration. If the noise estimation unit 40 of the first embodiment calculates the estimated noise power PN as an average of the input power PX in the noise section, the estimation bias does not occur. However, in the noise estimation unit 40 of the first embodiment, the estimated noise power PN is calculated as a weighted average by the stationary probability SP of the input power PX. As a result, in the noise estimation unit 40 of the first embodiment, there is a risk of causing an estimation bias, especially for the input power PX in the noise section.

[0098] Therefore, in the noise estimation means 43A (weighted average means 103A) of the second embodiment, the estimation bias is investigated in advance, and the estimated noise power PN is estimated while correcting the estimation bias.

[0099] (B-2) Operation of the Second Embodiment Next, the operation of the sound collection device 1A of the second embodiment having the above-mentioned configuration will be described with respect to differences from the first embodiment.

[0100] As described above, in the second embodiment, the operation of the weighted average means 103A constituting the noise estimation means 43A is different from that of the first embodiment. As described above, the noise estimation means 43A (weighted average means 103A) is different from that of the first embodiment in that it investigates an estimation bias in advance and estimates the estimated noise power PN while correcting the estimation bias.

[0101] In the weighted average means 103A, the estimated bias B can be calculated by simulation. In this embodiment, in the stationary probability calculation function in the stationary probability calculation means 101, a domain (posterior SNR G) in which the output of the stationary probability calculation function (stationary probability SP) is equal to or greater than the stationary probability threshold SP0 is prepared uniformly or uniformly randomly, the stationary probability calculation function is calculated for all elements of the set Gtest of values ​​of the prepared domain, and a weighted average of all elements of the obtained set SPtest of values ​​of the region and the corresponding elements of Gtest is calculated in the same manner as in equation (2). In the second embodiment, the weighted average value obtained at this time becomes the estimated bias B.

[0102] The weighted average means 103A calculates the estimated noise power PN using equation (5) based on the input buffer BX, the probability buffer BP, and the estimated bias B, and provides the obtained estimated noise power PN to the outputs of the stationary probability calculation means 101, the environmental transition determination means 105, and the noise estimation means 100.

number

[0103] (B-3) Effects of the Second Embodiment According to the second embodiment, in addition to the effects of the first embodiment, the following effects can be achieved.

[0104] In the sound collection device 1A (noise estimation means 43A) of the second embodiment, the estimation bias of the estimated noise power is corrected, so that it is possible to provide a noise estimation method capable of estimating the noise power more stably.

[0105] (C) Third embodiment Hereinafter, a third embodiment of the noise estimation device, the noise estimation program, the noise estimation method, the sound collection device, the sound collection program, and the sound collection method according to the present invention will be described in detail with reference to the drawings. In this embodiment, an example in which the noise estimation device, the noise estimation program, and the noise estimation method of the present invention are applied to a noise estimation means will be described.

[0106] (C-1) Configuration of the third embodiment The configuration of the sound collection device 1B of the third embodiment can also be shown using the above-mentioned FIGS.

[0107] The configuration of a sound collection device 1B of the third embodiment will be described below, focusing on the differences from the first embodiment.

[0108] In the sound collection device 1B of the third embodiment, the noise estimation section 40 is replaced with a noise estimation section 40B.

[0109] FIG. 7 is a block diagram showing the functional configuration of a noise estimation unit 40B according to the third embodiment.

[0110] In Fig. 7, the same or corresponding parts as those in Fig. 4 are denoted by the same or corresponding reference numerals. Note that the reference numerals in parentheses in Fig. 7 are reference numerals used only in the fourth embodiment described later.

[0111] As shown in FIG. 7, the noise estimation unit 40B includes a band division means 41, a power calculation means 42, and a noise estimation means 43B.

[0112] The noise estimation unit 40B differs from the first embodiment in that the power calculation means 42 and the noise estimation means 43B are not divided into elements for each band (frequency) but are configured to process all bands together.

[0113] The power calculation means 42 is configured to process all bands together using one element, and the actual signal processing is the same as in the first embodiment.

[0114] The noise estimation means 43B is not only configured to process all bands together using one element, but also differs from the first embodiment in the substantial signal processing.

[0115] Fig. 8 is a block diagram showing the functional configuration of a noise estimation means 43B according to the third embodiment. Note that in Fig. 8, the reference characters in parentheses are used only in the fourth embodiment described later.

[0116] In FIG. 8, the same or corresponding parts as those in FIG. 1 are denoted by the same or corresponding reference numerals.

[0117] The noise estimation unit 40B of the third embodiment differs from the first embodiment in that the stationary probability calculation means 101, the buffer update means 102, the weighted average means 103, the input averaging means 104, and the environmental transition determination means 105 are replaced with stationary probability calculation means 101B, the buffer update means 102B, the weighted average means 103B, the input averaging means 104B, and the environmental transition determination means 105B, respectively. The noise estimation unit 40B of the third embodiment also differs from the first embodiment in that an initial state control means 106 and an estimated noise averaging means 107 are added.

[0118] (C-2) Operation of the Third Embodiment Next, the operation of the sound collection device 1B of the third embodiment having the above-mentioned configuration will be described, focusing on the differences from the first embodiment.

[0119] As described above, in the third embodiment, the noise estimation means 43B is substantially different, so the operation of the noise estimation means 43B will be mainly described.

[0120] The band division means 41 performs frequency analysis on the input signal x to calculate an input spectrum, divides the obtained input spectrum into K parts, and provides the divided frequency band signals to the power calculation means .

[0121] The power calculation means 42 calculates input power for each of K bands based on the input frequency band signal, and provides the input power PX (PX_1 to PX_K) to the noise estimation means 43. The power calculation means 42 can apply various calculation methods as a method for calculating power. For example, the power calculation means 42 may calculate the square of the absolute value or the absolute value as the input power.

[0122] The noise estimation means 43 estimates the power of the noise component contained in the input power PX (PX_1 to PX_K) supplied from the power calculation means 42, and outputs the result (estimated noise power PN (PN_1 to PN_K)).

[0123] The method (e.g., signal format or data format) and output destination of the noise estimation means 43B when outputting the estimated noise power PN (PN_1 to PN_K) are not limited. For example, the noise estimation means 43 may supply the estimated noise power PN (PN_1 to PN_K) to a voice processing method (not shown) that emphasizes the target sound of the input power PX (suppresses noise). Also, the sound collection device 1 may be configured as a part of the voice processing method described above.

[0124] The noise estimation means 43B estimates the power of the noise components contained in the input powers PX (PX_1 to PX_K) supplied from the power calculation means 42, and obtains the estimated noise powers PN (PN_1 to PN_K) as a result.

[0125] Next, the operation of the noise estimation means 43 will be described with reference to FIG.

[0126] The noise estimation means 43 provides the input powers PX (PX_1 to PX_K) to the stationary probability calculation means 101B, the buffer update means 102B, and the input averaging means 104B.

[0127] The initial state control means 106 updates an initial state counter IC based on an environment transition determination result E, which will be described later, and provides the updated initial state counter IC to the stationary probability calculation means 101B.

[0128] The method for setting the environment transition determination result E can be the same as in the first embodiment, so a detailed description will be omitted.

[0129] The initial state control means 106 sets the initial state counter IC to a first sample number T1 described below when the noise estimation means 43 starts operating and when the given environment transition judgment result E becomes True, and does not update the initial state counter IC otherwise.

[0130] The initial state counter IC is prepared by the initial state control means 106 when the noise estimation means 43 starts operating, but thereafter is provided by the stationary probability calculation means 101B.

[0131] The steady-state probability calculation means 101B calculates a steady-state probability SP_k for each band k (k=1 to K) based on the input power PX_k, the estimated noise power PND_k of a predetermined D samples in the past, and an initial state counter IC described below, and provides the obtained steady-state probability SP_k to the buffer update means 102B, and further updates the initial state counter IC and provides it to the initial state control means 106.

[0132] The method for setting D can be the same as in the first embodiment, and therefore a detailed description thereof will be omitted.

[0133] In this embodiment, when the initial state counter IC is greater than 0, the stationary probability calculation means 101B sets the stationary probability SP_k to a predetermined value of 1.0 and decreases (decrements) the initial state counter IC by 1. Note that any value can be set as the predetermined value set in the above SP_k.

[0134] In addition, in this embodiment, when the initial state counter IC is 0, the stationary probability SP_k is calculated based on the posterior SNR G_k. In addition, in this embodiment, the posterior SNR is calculated by the formula (6). In addition, in this embodiment, the stationary probability SP (SP_k) is calculated by a stationary probability calculation function to be described later. In this embodiment, the initial state counter IC is not updated.

number

[0135] In this embodiment, the stationary probability calculation function is a function of stationary probability SP_k for posterior SNR G_k, and is determined in advance. In addition, in this embodiment, the domain (G_k) is a real number of 0.0 or more, and the range (SP_k) is a real number of 0.0 or more and 1.0 or less. Furthermore, in this embodiment, the shape of the stationary probability calculation function is a mountain shape, that is, for a predetermined value G0 of G_k, the stationary probability calculation function increases monotonically in a wide sense in the range of 0.0≦G_k≦G0, and the stationary probability calculation function decreases monotonically in a wide sense in the range of G0≦G_k.

[0136] In this embodiment, G0 is not particularly limited and can take any value within the range of the domain of G_k. However, in this embodiment, since the stationary probability calculation function is a function that outputs a higher probability as the stationary state increases, it is most preferable to set G0=1.0 (i.e., PX_k=PND_k).

[0137] An example of the stationary probability calculation function in this embodiment can also be shown using the above-mentioned FIGS.

[0138] FIG. 9 shows an example of a stationary probability calculation function smoothly defined for G_k.

[0139] In the stationary probability calculation function shown in Fig. 9, G_k(G0) at which SP_k=1.0 is not particularly limited, but is set to, for example, G0=1.0. In the stationary probability calculation function shown in Fig. 9, it is preferable that SP_k asymptotically converges to 0.0 as G_k→+∞.

[0140] FIG. 10 is an example of a stationary probability calculation function designed on a logarithmic scale.

[0141] As shown in Figure 10, by designing a stationary probability calculation function on a logarithmic scale for both axes, PX_k<PND_kのときとPX_k> The relationship for PND_k becomes fair, and the stationary probability calculation function can be easily defined over the entire domain.

[0142] The buffer updating means 102B updates the input buffer BX_k and the probability buffer BP_k based on the input power PX_k and the stationary probability SP_k only when the stationary probability SP_k is greater than a predetermined stationary probability threshold SP0 for each band k (k=1 to K), and provides the obtained input buffer BX_k and probability buffer BP_k to the weighted average means 103C. Note that in the buffer updating means 102B, the above comparison and update operations are performed independently for each band k (k=1 to K). For example, in the buffer updating means 102B, at a certain time, two buffers may be updated in one band but not updated in another band.

[0143] In addition, in the buffer update means 102B, the stationary probability threshold SP0 must be set to a sufficiently small value (for example, about 0.1 or less). The reason for this will be described later in the explanation of the weighted average means 103C. For example, the stationary probability threshold SP0 is preferably 0.01.

[0144] The input buffer BX_k is a buffer that holds T1 samples of the input power PX_k from the past to the present. The first number of samples T1 may be any value, but is set to a length equivalent to 100 milliseconds to several seconds depending on the application. If the first number of samples T1 is set to a short value (e.g., equivalent to 100 milliseconds), the tracking of changes in the noise of the estimation target becomes faster, but there is a risk that, for example, a voice that is spoken slowly may be erroneously estimated as noise. On the other hand, if the first number of samples T1 is set to a long value (e.g., equivalent to 8 seconds), the stationary components in the input signal can be accurately estimated, but if the noise environment changes suddenly (e.g., when going from a quiet room to a noisy outdoor area, when air conditioning is turned on / off all at once, etc.), it is not possible to immediately track them. Therefore, it is preferable that the first number of samples T1 is set to a length equivalent to 200 milliseconds to 1 second.

[0145] The probability buffer BP_k is a buffer that holds T1 samples of the stationary probability SP_k from the past to the present. The buffer length of the probability buffer BP_k is the same as the buffer length of the input buffer BX_k.

[0146] The buffer update means 102B deletes the oldest input power PX_k from the input buffer BX_k for each band k (k=1 to K) according to equation (7) and stores the newly given input power PX_k in the input buffer BX. Similarly, the buffer update means 102B deletes the oldest stationary probability SP_k from the probability buffer BP_k for each band k (k=1 to K) according to equation (8) and stores the newly given stationary probability SP_k in the probability buffer BP_k. Note that if the stationary probability SP_k is greater than the stationary probability threshold SP0, the input buffer BX_k and the probability buffer BP_k are not updated.

number

[0147] The weighted averaging means 103C calculates an estimated noise power PN_k for each band k (k=1 to K) based on the input buffer BX_k and the probability buffer BP_k, and provides the obtained estimated noise power PN_k to the output of the stationary probability calculation means 101B, the estimated noise averaging means 107, and the noise estimation means 43C.

[0148] In addition, the weighted average means 103C calculates the weighted average of the input power PX_k stored in the input buffer BX_k using the stationary probability SP_k stored in the probability buffer BP_k as a weighting coefficient according to equation (9), thereby calculating the estimated noise power PN_k.

number

[0149] The method of calculating the estimated noise power PN_k by weighted averaging has the following advantages. That is, since a large weight is applied to the input power PX_k with a high stationary probability SP_k, it has a large influence on the calculation of the estimated noise power PN_k. On the other hand, since a small weight is applied to the input power PX_k with a low stationary probability SP, it has a small influence on the calculation of the estimated noise power PN_k and is almost ignored. Here, if the input power PX_k with a small stationary probability SP_k is completely ignored (for example, in Patent Document 1, it is ignored if the posterior SNR is greater than a predetermined value), a problem occurs in which it is not possible to follow when the input signal returns after being temporarily reduced due to packet loss or voice switching. Therefore, by considering the input power PX_k with a small stationary probability SP_k with a small influence, it becomes possible to continue stable estimation.

[0150] Also, in the buffer update means 102B, the steady probability threshold SP0 must be set to a value smaller than a predetermined value. If SP0 is set to a value larger than a predetermined value, the merit of the present invention of continuing the above-mentioned stable estimation is lost.

[0151] When the environment transition determination means 105B determines that the environment has transitioned, the initial state control means 106 and the stationary probability calculation means 101B cause 1.0 to be continuously set for the stationary probability SP_k during the first sample number T1, so that the estimated noise power PN_k calculated according to equation (9) becomes a simple average of the input power PX_k (input buffer BX_k), and can quickly follow the environment after the transition.

[0152] The input averaging means 104B averages the K bands of the input power PX_k according to equation (10), holds the band average value for T2 samples from the past to the present and averages them to calculate the average input power APX, and gives the obtained average input power APX to the environment transition determination means 105B. The value of the second sample number T2 will be described later.

Equation

[0153] The estimated noise averaging means 107 averages the K bands of the estimated noise power PN_k and gives the obtained average noise power APN to the environment transition determination means 105B.

[0154] The environment transition determination means 105B determines whether the ratio of the average input power APX to the average noise power APN falls within a predetermined stationary determination range R, sets the environment transition determination result E, and gives the obtained environment transition determination result E to the stationary probability calculation means 101B.

[0155] In this embodiment, there are two possible methods for the processing method (definition method and its usage method) of the stationary determination range R.

[0156] The first processing method of the stationary determination range R is a method of defining the stationary determination range as a vector of two positive values (R1 < R2) and calculating the environment transition determination result E using equation (11).

[0157] The second method of processing the steady-state judgment range R is to define the steady-state judgment range as a positive scalar with a value of 1, and to calculate the environmental transition judgment result E using equation (12). In this case, R is in decibel scale.

number

[0158] In the environmental transition determination means 105B, either of the above two processing methods may be used, but the second method is preferable because of the ease of setting. In that case, it is preferable to set the steady-state determination range R to, for example, 6 decibels.

[0159] The role of the environment transition determination means 105B is to make the estimated noise power PN_k quickly follow the input power PX_k with some delay when a state in which the characteristics of the input signal x have changed significantly continues (for example, when there is complete silence, or when an abnormally loud volume is input due to an incorrect connection of an audio device, or when the normal state is restored from such a state).

[0160] If the characteristics of the input signal x change significantly, causing the input power PX_k to change significantly, the stationary probability SP_k will continue to take a small value, resulting in a deadlock state in which the input buffer BX_k and the probability buffer BP_k are not updated. To avoid this problem, the environmental transition determination means 105B calculates the long-term average of the input power PX_k. Since the proportion of non-stationary components (called target components) that are not subject to noise suppression in the time direction is often small, the environmental transition determination means 105B can remove the influence of the target components by taking the long-term average of the band average value of the input power PX_k.

[0161] In the environmental transition determination means 105B, this long-term average is taken as the average input power APX, and by comparing it with the band average value of the estimated noise power PN_k, which is an estimate of the average of the noise power in a relatively short period, it is possible to detect a transition of the environment. Therefore, in the environmental transition determination means 105B, the second sample number T2 for obtaining the average input power APX must be sufficiently long, and a suitable value is, for example, 5 to 20 seconds.

[0162] (C-3) Effects of the Third Embodiment According to the third embodiment, the following effects can be obtained.

[0163] In the sound collection device 1B (noise estimation means 43B) of the third embodiment, even if the noise suddenly increases, a weighted average with the stationary probability as a weighting coefficient is used, so that the noise is tracked little by little using small noise power that appears with low frequency among the noise power that changes probabilistically, and therefore the noise does not increase suddenly.

[0164] In addition, in the sound collection device 1B (noise estimation means 43B) of the third embodiment, if the input sound suddenly becomes quieter, it is considered to be unsteady, so there is no excessive tracking of the temporarily quieter input sound, and the problem of the estimated noise power remaining small does not occur.

[0165] Furthermore, in the sound collection device 1B (noise estimation means 43B) of the third embodiment, since the object is to estimate only the stationary noise components contained in the input sound, there is no need to adjust the estimation parameters in accordance with the environment.

[0166] Furthermore, in the sound collection device 1B (noise estimation means 43B) of the third embodiment, when the environment changes and the estimated noise power becomes a value that deviates from the average input power, the stationary probability is set to 1.0 for the first sample number, thereby quickly tracking the new environment.

[0167] From the above, the sound collection device 1B (noise estimation means 43B) of the third embodiment can estimate noise power without adjusting estimation parameters and without being influenced by sudden changes or variations in the input sound.

[0168] (D) Fourth embodiment Hereinafter, a fourth embodiment of the noise estimation device, the noise estimation program, the noise estimation method, the sound collection device, the sound collection program, and the sound collection method according to the present invention will be described in detail with reference to the drawings. In this embodiment, an example in which the noise estimation device, the noise estimation program, and the noise estimation method according to the present invention are applied to a noise estimation means will be described.

[0169] (D-1) Configuration of the Fourth Embodiment The configuration of the sound collection device 1C of the fourth embodiment can also be shown using the above-mentioned Fig. 2 and Fig. 3. In addition, the functional configuration of the noise estimation means 43C constituting the sound collection device 1C according to the fourth embodiment can also be shown using Fig. 8.

[0170] The following describes the configuration of a sound collection device 1C of the fourth embodiment, focusing on differences from the third embodiment.

[0171] In the sound collection device 1C of the fourth embodiment, the noise estimation unit 40B is replaced with a noise estimation unit 40C. Also, in the noise estimation unit 40C of the fourth embodiment, the noise estimation means 43B is replaced with a noise estimation means 43C. Furthermore, in the noise estimation unit 40C of the fourth embodiment, the weighted average means 103B is replaced with a weighted average means 103C.

[0172] In the noise estimation means 43B of the third embodiment, the estimation bias of the estimated noise power PN_k is not taken into consideration. If the noise estimation means 43B of the third embodiment calculates the estimated noise power PN_k as an average of the input power PX_k in the noise section, the estimation bias does not occur. However, in the noise estimation means 43B of the third embodiment, the estimated noise power PN_k is calculated as a weighted average by the stationary probability SP of the input power PX_k, so there is a risk of causing an estimation bias for the input power PX_k in the noise section.

[0173] Therefore, in the noise estimation means 43C of the fourth embodiment, the estimation bias is investigated in advance, and the estimated noise power PN_k is estimated while correcting the estimation bias.

[0174] (D-2) Operation of the Fourth Embodiment Next, the operation of the sound collection device 1C of the fourth embodiment having the above-mentioned configuration will be described with respect to differences from the third embodiment.

[0175] As described above, the noise estimation means 43C of the fourth embodiment is different from the third embodiment in that the estimated bias is investigated in advance and the estimated noise power PN_k is estimated while correcting the estimated bias.

[0176] In the fourth embodiment, the estimated bias B may be calculated by simulation. For example, first, in the stationary probability calculation function used in the stationary probability calculation means 101, a domain (posterior SNRG) in which the output (stationary probability SP) of the stationary probability calculation function is equal to or greater than the stationary probability threshold SP0 is prepared uniformly or uniformly randomly, the stationary probability calculation function is calculated for all elements of the set Gtest of values ​​in the prepared domain, and a weighted average of all elements of the obtained set SPtest of values ​​and the corresponding elements of Gtest is calculated in the same manner as in equation (9). Then, the weighted average value obtained by equation (13) may be set as the estimated bias B.

[0177] Specifically, the weighted averaging means 103C calculates the estimated noise power PN_k using equation (13) based on the input buffer BX_k, the probability buffer BP_k, and the estimated bias B, and provides the obtained estimated noise power PN_k to the outputs of the stationary probability calculation means 101, the estimated noise averaging means 107, and the noise estimation means 100.

number

[0178] (D-3) Effects of the Fourth Embodiment According to the fourth embodiment, in addition to the effects of the third embodiment, the following effects can be achieved.

[0179] In the sound collection device 1C (noise estimation section 40C) of the fourth embodiment, the estimation bias of the estimated noise power is corrected, so that it is possible to provide a noise estimation method capable of estimating the noise power more stably.

[0180] (E) Other embodiments The present invention is not limited to the above-described embodiments, and may include modified embodiments such as those exemplified below.

[0181] (E-1) In the above embodiment, for simplicity, the description will be given assuming that one microphone is installed, but a configuration in which multiple microphones are arranged may be used, or a configuration in which a microphone array using multiple microphones M is arranged may be used. In that case, the noise estimation unit (noise estimation means) will perform noise estimation processing on each input signal from the multiple microphones.

[0182] (E-2) In each of the above embodiments, an example in which the noise estimation unit is mounted in a sound collection device has been described, but the device in which the noise estimation unit is mounted is not limited. In addition, the noise estimation unit or the noise estimation means may be configured as a separate device (noise estimation device).

[0183] Therefore, in each of the above embodiments, the output destination and output method (e.g., data format and output interface) of the estimated noise power PN (PN_1 to PN_K) acquired by the noise estimation unit (noise estimation means) are not limited, and the noise estimation unit (noise estimation means) may output the estimated noise power PN (PN_1 to PN_K) to an output destination according to the application of the noise estimation unit in an output method according to the application of the noise estimation unit. For example, the noise estimation unit (noise estimation means) may output the estimated noise power PN using an interface (e.g., a signal line on a circuit or a serial interface, etc.) provided in a computer, or may output the estimated noise power PN using a wired or wireless communication interface (e.g., a wired / wireless LAN interface, various serial interfaces, etc.).

[0184] (E-3) In the sound collection device, signal input unit, noise suppression unit, and noise estimation unit of each of the above embodiments, when the processing of the noise suppression unit is performed in the frequency domain, it is necessary to have a band splitting means inside the noise suppression unit, but the configuration of the band splitting means in the noise estimation unit or the output of the band splitting means may be shared. For example, the band splitting means may be included inside the signal input unit, and the output of the band splitting means may be supplied to the noise suppression unit and the noise estimation unit as the output of the signal input unit. [Explanation of symbols]

[0185] 1...sound collection device, M...microphone, 10...signal input section, 20...noise suppression processing section, 30...signal output section, 40...noise estimation section, 41...band division means, 42, 42-1 to 42-K...power calculation means, 43, 43-1 to 43-K...noise estimation means, 101...stationary probability calculation means, 102...buffer update means, 103...weighted averaging means, 104...input averaging means, 105...environment transition determination means.

Claims

1. A noise estimation device for estimating a noise component contained in an input signal containing a mixture of speech and noise to obtain an estimated noise power, a stationary probability calculation means for calculating a stationary probability of the input power from the input power of the input signal and the estimated noise power in the past; a buffer update means for updating an input buffer for holding the input power for a predetermined first number of samples and a probability buffer for holding the stationary probability for the first number of samples only when the stationary probability is greater than a predetermined stationary probability threshold; a weighted averaging means for calculating the estimated noise power by taking a weighted average of the input power stored in the input buffer using the stationary probability stored in the probability buffer as a weighting factor; an input averaging means for calculating an average input power by retaining and averaging the input power for a second predetermined number of samples; an environmental transition determination means for determining whether a ratio of the estimated noise power or a value based on the estimated noise power to the average input power falls within a predetermined steady-state determination range and calculating an environmental transition determination result; The stationary probability calculation means sets the stationary probability to a predetermined value when the environment transition determination means determines that the environment has transitioned, and calculates the stationary probability by a stationary probability calculation function that calculates the stationary probability based on a posterior SNR obtained by dividing the input power by the past estimated noise power when the environment transition determination means determines that the environment has not transitioned. A noise estimation device comprising:

2. 2. The noise estimation device according to claim 1, wherein the stationary probability calculation function is a function that, with respect to a peak posterior SNR at which the stationary probability calculation function is maximized, increases monotonically in a broad sense for posterior SNRs equal to or lower than the peak posterior SNR, and decreases monotonically in a broad sense for posterior SNRs equal to or higher than the peak posterior SNR.

3. the input averaging means calculates the average input power by retaining and averaging a predetermined second number of samples of an average value of the input power over the entire band; The estimated noise power is further averaged over a whole band by an average noise power averaging unit, The environmental transition determination means determines whether a ratio of the average noise power to the average input power falls within a predetermined steady-state determination range, The method further includes an initial state control means for controlling the stationary probability calculation means to operate in an initial state during the first execution and during the first sample number from when the environment transition determination means determines that the environment has transitioned.

3. The noise estimation device according to claim 1 or 2.

4. 3. The noise estimation device according to claim 2, wherein said environment transition determining means, after determining that the environment has transitioned, continues the determined state that the environment has transitioned for the first number of samples.

5. 5. The noise estimation device according to claim 1, wherein the weighted average means uses a previously calculated estimation bias to correct the weighted average of the input power with the stationary probability as a weighting coefficient to obtain the estimated noise power.

6. A computer installed in a noise estimation device that estimates a noise component contained in an input signal in which speech and noise are mixed to obtain an estimated noise power, a stationary probability calculation means for calculating a stationary probability of the input power from the input power of the input signal and the estimated noise power in the past; a buffer update means for updating an input buffer for holding the input power for a predetermined first number of samples and a probability buffer for holding the stationary probability for the first number of samples only when the stationary probability is greater than a predetermined stationary probability threshold; a weighted averaging means for calculating the estimated noise power by taking a weighted average of the input power stored in the input buffer using the stationary probability stored in the probability buffer as a weighting factor; an input averaging means for calculating an average input power by retaining and averaging the input power for a second predetermined number of samples; functioning as an environmental transition determination means for calculating an environmental transition determination result by determining whether a ratio of the estimated noise power or a value based on the estimated noise power to the average input power falls within a predetermined steady-state determination range; The stationary probability calculation means sets the stationary probability to a predetermined value when the environment transition determination means determines that the environment has transitioned, and calculates the stationary probability by a stationary probability calculation function that calculates the stationary probability based on a posterior SNR obtained by dividing the input power by the past estimated noise power when the environment transition determination means determines that the environment has not transitioned. A noise estimation program comprising:

7. A noise estimation method performed by a noise estimation device for estimating a noise component contained in an input signal containing a mixture of speech and noise to obtain an estimated noise power, comprising: the noise estimation device comprises a stationary probability calculation means, a buffer update means, a weighted average means, an input average means, and an environment transition determination means; the stationary probability calculation means calculates a stationary probability of the input power from the input power of the input signal and the estimated noise power in the past; the buffer update means updates an input buffer that holds the input power for a predetermined first number of samples and a probability buffer that holds the stationary probability for the first number of samples only when the stationary probability is greater than a predetermined stationary probability threshold; the weighted average means calculates the estimated noise power by taking a weighted average of the input power stored in the input buffer using the stationary probability stored in the probability buffer as a weighting coefficient; the input averaging means holds and averages a second predetermined number of samples of the input power to calculate an average input power; the environmental transition determination means calculates an environmental transition determination result by determining whether a ratio of the estimated noise power or a value based on the estimated noise power to the average input power falls within a predetermined steady-state determination range; The stationary probability calculation means sets the stationary probability to a predetermined value when the environment transition determination means determines that the environment has transitioned, and calculates the stationary probability by a stationary probability calculation function that calculates the stationary probability based on a posterior SNR obtained by dividing the input power by the past estimated noise power when the environment transition determination means determines that the environment has not transitioned.

2. A noise estimation method comprising:

8. 1. A sound collecting device for collecting a sound from an input signal containing a mixture of the sound and noise, a noise estimator that estimates a noise component contained in the input signal; a sound collection unit that extracts the speech signal from the input signal by using the estimation result of the noise estimation unit, The noise estimation unit is configured to use a noise estimation device according to any one of claims 1 to 5. A sound recording device characterized by:

9. A computer installed in a sound collection device that collects a sound from an input signal in which the sound and noise are mixed, a noise estimator that estimates a noise component contained in the input signal; A sound collecting unit that extracts the speech signal from the input signal and collects the speech signal by using an estimation result of the noise estimating unit, The noise estimation unit functions as a noise estimation device according to any one of claims 1 to 5. A sound recording program characterized by:

10. A sound collection method performed by a sound collection device that collects sound from an input signal in which sound and noise are mixed, comprising: A noise estimation unit and a sound collection unit are provided, The noise estimation unit estimates a noise component included in the input signal by a noise estimation method according to claim 7, The sound collection unit extracts the voice signal from the input signal by using the estimation result of the noise estimation unit. A sound collection method comprising:

Images