JP7139822B2 - Noise estimation device, noise estimation program, noise estimation method, and sound collection device - Google Patents

Description

The present invention relates to a noise estimation device, a noise estimation program, a noise estimation method, and a sound collection device, for example, a device that suppresses noise components superimposed on an input signal using estimation results of noise components included in the input signal. can be applied to

Since noise exists everywhere in the natural environment, speech observed in the real world generally contains noise from various sources. Various noise suppression methods have been developed to enhance only speech from an observed noisy input signal. Most of these have methods for estimating the noise to be suppressed and methods for computing filters to suppress the noise. Some conventional speech processing devices that suppress noise from an input signal estimate noise power in the frequency domain.

Conventionally, as an example of the simplest noise estimation method, there is a method of averaging an input spectrum in an interval in which speech does not exist. However, such a conventional noise estimation method requires pre-estimation of a section in which no speech exists. For this reason, a technique called voice activity detection (VAD) for estimating an interval in which voice exists has been actively developed, but perfect VAD has not yet been achieved. In the noise estimation process, if the estimation of the speech period is erroneous, the estimated noise includes the target speech, which causes the problem of distorting the emphasized speech and the residual noise. In addition, the noise estimation method described above estimates noise only in the noise interval, and therefore has the disadvantage that it cannot follow changes in noise if there is a long speech interval.

Against this background, there are techniques described in Non-Patent Document 1, Non-Patent Document 2, and Patent Document 1 as noise estimation methods for continuing noise estimation even in speech segments. Both documents relate to noise suppression methods (also called speech enhancement methods).

The conventional noise estimation method described in Non-Patent Document 1 is based on the discovery that the temporal peaks of the input power indicate the presence of the target speech, while the valleys can be used to estimate the smoothed noise power. Specifically, the minimum value of the input power from the present to the past for a predetermined time is taken as the first estimated noise power. However, the first estimated noise power has a bias and has the property of being smaller than the true noise power. This bias is estimated from the expected value of the first noise power estimate, and the resulting bias estimate is used to correct the first noise power estimate to produce a second noise power estimate (final estimate) get

The conventional noise estimation method described in Patent Document 1 multiplies the input power by an appropriate weighting factor, stores the obtained weighted input power for a predetermined time period, and estimates the average value of the stored weighted input power. noise power. Appropriate weighting factors are calculated based on the a posteriori SNR (Signal-to-Noise Ratio), which is the current input power divided by the previous estimated noise power. Specifically, when the posterior SNR is less than or equal to a predetermined value G1, the weighting factor is set to 1, when the posterior SNR is greater than or equal to G1, the weighting factor is set so as to be inversely proportional to the posterior SNR, and when the posterior SNR is greater than or equal to the predetermined value G2, the weighting factor is set to 0. Also, when the weighting factor is 0, the weighted input power is not stored.

The conventional noise estimation method described in Non-Patent Document 2 is based on the hypothesis that the distributions of the complex spectra of the target speech and the noise both follow a complex normal distribution with a mean of zero. be the estimated noise power. Based on this hypothesis, the distribution of the complex spectrum of the input speech is a zero-mean complex normal distribution whose variance is the sum of the variance of the speech complex spectrum and the noise complex spectrum. Here, we introduce a hidden variable regarding whether the current input is degraded speech or noise, and apply an on-line EM (Expectation Maximization) algorithm with a forgetting factor to obtain the maximum likelihood estimate of the complex spectrum of the noise. Calculate

Japanese Unexamined Patent Application Publication No. 2002-204175

R. Martin, "Spectral Subtraction Based on Minimum Statistics," in Proceedings of the 7th European Signal Processing Conference, 1994, pp. 1182-1185 M. Souden, M. De1croix, K.; Kinsoshita, T. Yoshioka, and T. Nakatani, "Noise Power Spectral Density Tracking: A Maximum Likelihood Perspective", IEEE Signal Processing Letters, Vol. 19, No. 8, 2012, pp. 495-498

However, conventional noise estimation methods have the following problems.

The method of Non-Patent Document 1 has a problem that when noise suddenly increases, the estimated noise power increases rapidly with a delay. Specifically, the estimated noise power remains small for a predetermined time after the noise increases. Then, the estimated noise power sharply increases a predetermined time after the noise increases.

In the method of Patent Document 1, when a phenomenon occurs in which the input voice is temporarily reduced due to packet loss in voice transmission, a voice switch for echo countermeasures, or the like, the estimated noise power remains small for a while even after the input voice returns to normal. There is a problem of That is, a small input speech has a posterior SNR smaller than G1, so it continues to be used for noise power estimation, and the estimated noise power becomes smaller. In this state, when the input speech becomes louder, the posterior SNR becomes larger than G2, so it is not used for noise power estimation, and the estimated noise power is not updated.

In the method of Non-Patent Document 2, the online EM algorithm used in this noise estimation method has a higher forgetting factor to increase stability and slow down tracking, and a small forgetting factor to speed up tracking and improve stability. There is a problem that there is a trade-off between the speed of tracking and the stability of maximum likelihood estimation. In the technology described in Non-Patent Document 2, the forgetting factor needs to be set according to the noise level of the observation environment, so it is not practical.

As described above, in the conventional noise estimation method, when the input speech suddenly increases, the estimated noise power suddenly increases at an inappropriate timing, and when the input speech decreases temporarily, the estimated noise power remains small. Also, there was a problem that it was necessary to adjust the estimated parameters according to the environment.

In view of the above problems, noise estimation apparatus, noise estimation program, noise estimation apparatus, noise estimation program, A noise estimation method and sound collection device are desired.

A first aspect of the present invention provides a noise estimation apparatus for obtaining an estimated noise power by estimating a noise component contained in an input signal in which speech and noise are mixed, wherein: (1) the input power of the input signal and the past estimation; (2) an input buffer for buffering past input power data; and a buffer for past stationary probabilities calculated by the stationary probability calculation means. (3) buffer updating means for updating the content of the input buffer and the content of the probability buffer based on the input power and the stationary probability calculated by the stationary probability calculating means; weighted averaging means for calculating the estimated noise power by taking a weighted average of the input powers held in the input buffer using the stationary probability held in the probability buffer as a weighting factor.

A noise estimation program according to a second aspect of the present invention provides a computer installed in a noise estimation device for obtaining an estimated noise power by estimating a noise component contained in an input signal in which speech and noise are mixed: (1) the input (2) an input buffer for buffering data of the input power in the past; a probability buffer for buffering the steady-state probability calculated by the calculation means, and updating the content of the input buffer and the content of the probability buffer based on the input power and the steady-state probability calculated by the steady-state probability calculation means. (3) weighted average means for calculating the estimated noise power by taking a weighted average of the input powers held in the input buffer using the stationary probability held in the probability buffer as a weighting factor; It is characterized by functioning as

A third aspect of the present invention provides a noise estimation method performed by a noise estimation apparatus for obtaining an estimated noise power by estimating a noise component contained in an input signal in which speech and noise are mixed, comprising: (1) stationary probability calculating means; (2) the stationary probability calculation means calculates the stationary probability of the input power from the input power of the input signal and the past estimated noise power; (3) the buffer; The updating means includes an input buffer for buffering data of the past input power and a probability buffer for buffering the steady-state probability calculated by the steady-state probability calculating means in the past, and the input power and the steady-state probability calculating means. (4) the weighted averaging means updates the contents of the input buffer and the probability buffer based on the steady-state probability calculated in (4), wherein the weighted average means updates the steady-state probability held in the probability buffer as a weighting factor; The estimated noise power is calculated by taking a weighted average of the input powers held in a buffer.

A fourth aspect of the present invention is a sound collection device for collecting speech from an input signal in which speech and noise are mixed, comprising: (1) a noise estimation unit for estimating a noise component included in the input signal; a sound collecting unit that extracts and collects the voice from the input signal using the estimation result of the noise estimating unit, and the noise estimating apparatus according to the first aspect of the present invention is applied as the noise estimating unit. It is characterized by

According to the present invention, the number of items that need to be adjusted is small, and the noise power can be estimated without being influenced by sudden changes or changes in the input speech.

3 is a block diagram showing the functional configuration of noise estimation means according to the first and second embodiments; FIG. 2 is a block diagram showing the functional configuration of the sound collecting device according to the first and second embodiments; FIG. 3 is a block diagram showing the hardware configuration of the sound collecting device according to the first and second embodiments; FIG. 3 is a block diagram showing the functional configuration of a noise estimator (noise estimator) according to the first and second embodiments; FIG. FIG. 4 is an explanatory diagram (graph) showing an example of a stationary probability calculation function according to the first and second embodiments; FIG. 4 is an explanatory diagram (graph) showing an example of a case where the stationary probability calculation function according to the first and second embodiments is designed on a logarithmic scale;

(A) First Embodiment Hereinafter, a first embodiment of a noise estimation device, a noise estimation program, a noise estimation method, and a sound collection device 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 apparatus, noise estimation program, and noise estimation method of the present invention are applied to noise estimation means will be described.

(A-1) Configuration of First Embodiment FIG. 2 is a block diagram showing the functional configuration of the sound pickup device 1 of this embodiment. Note that the symbols in parentheses in FIG. 2 are symbols used only in the second embodiment described later.

The sound collection device 1 is a device that performs sound collection processing for collecting a target sound from an acoustic signal captured by the microphone M. FIG.

In the example of this embodiment, it is assumed that the microphone M is mounted on the receiver of a telephone terminal (not shown). In this case, the acoustic signal captured by the microphone M includes, for example, speech as the target sound (for example, the speech of the near-end speaker) and noise as the non-target sound (for example, background noise). In the example of this embodiment, the sound collecting device 1 is mounted on a telephone terminal (not shown), and removes non-target sounds (for example, noise such as background noise) from the acoustic signal captured by the microphone M to obtain the target sound. (for example, the voice of the near-end speaker).

Next, the internal configuration of the sound collecting device 1 will be described.

In this embodiment, the sound collecting device 1 includes a signal input section 10, a noise suppression section 20, a signal output section 30, and a noise estimation section 40. FIG.

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

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 the digital signal to the computer 200 . Hereinafter, the acoustic signal captured by the microphone M and digitally converted by the signal input section 10 will be referred to as an input signal x.

The noise estimation unit 40 has a function of estimating noise (non-target sound) included in the input signal x. Hereinafter, the noise (estimated noise signal) estimated by the noise estimation unit 40 is referred to as "estimated noise".

The noise suppressor 20 uses the estimated noise estimated by the noise estimator 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"). there is

The signal output unit 30 has a function of outputting the sound pickup result of the sound pickup device 1 (in this embodiment, the noise-suppressed signal output by the noise suppression unit 20).

FIG. 3 is a block diagram showing an example of the hardware configuration of the sound collecting device 1. As shown in FIG. Note that the symbols in parentheses in FIG. 3 are symbols used in the second embodiment described later.

FIG. 3 shows a configuration when the sound collecting device 1 is configured using software (computer).

The sound collecting device 1 shown in FIG. 3 has, as hardware components, at least a signal input unit 10 and a computer 200 in which a program (a sound collecting program including a noise estimation program of the embodiment) is installed. .

The signal input unit 10 can be configured using, for example, a D/A converter. Note that if the computer 200 itself is equipped with a D/A converter, the signal input unit 10 need not be provided separately.

The computer 200 performs predetermined processing on the acoustic signal (digital acoustic signal) supplied from the signal input unit 10 and outputs the processed signal. In this embodiment, it is assumed that the computer 200 is installed with programs corresponding to at least the noise suppression unit 20, the signal output unit 30, and the noise estimation unit 40 (sound pickup program of this embodiment). Note that the sound collection program of this embodiment includes a noise estimation program corresponding to the noise estimation unit 40 .

Note that the computer 200 may be a computer dedicated to the sound collection program, or may be shared with a program for other functions (for example, a function of outputting a far-end signal (receiving signal) received by the telephone terminal from a speaker (not shown)). may be configured.

A computer 200 shown in FIG. 3 has a processor 201 , a primary storage unit 202 and a secondary storage unit 203 . The primary storage unit 202 is storage means that functions as a working memory (work memory) for the processor 201, and for example, a high-speed memory such as a DRAM is applied. The secondary storage unit 203 is storage means for recording various data such as an OS (Operating System) and program data (including sound pickup program data according to the embodiment). sensitive memory is applied. In the computer 200 of this embodiment, when the processor 201 is activated, the OS and programs (including the sound pickup program according to the embodiment) recorded in the secondary storage unit 203 are read, and expanded on the primary storage unit 202. Run.

Note that the specific configuration of the computer 200 is not limited to the configuration in FIG. 3, and various configurations can be applied. For example, if the primary storage unit 202 is a non-volatile memory (for example, FLASH memory), the secondary memory may be excluded.

Next, the internal configuration of the noise estimator 40 will be explained using FIG.

FIG. 4 is a block diagram showing the functional configuration of the noise estimation section 40. As shown in FIG. Note that the symbols in parentheses in FIG. 4 are symbols used in the second embodiment described later.

The noise estimator 40 has a band dividing means 41, K power calculating means 42 (42-1 to 42-K), and K noise estimating means 43 (43-1 to 43-K). ing.

FIG. 1 is an explanatory diagram showing the internal configuration of each noise estimation means 43 (43-1 to 43-K). In this embodiment, the interiors of the noise estimation means 43-1 to 43-K are all configured as shown in FIG. The symbols in parentheses in FIG. 1 are symbols used in the second embodiment described later.

As shown in FIG. 1, each noise estimation means 43 has stationary probability calculation means 101 , buffer update means 102 , weighted average means 103 and storage means 104 .

Detailed functions (operations) of the elements constituting the noise estimation section 40 (including the elements constituting the noise estimation means 43) will be described later.

(A-2) Operation of the First Embodiment Next, the operation of the sound pickup device 1 (noise estimation unit 40) having the configuration as described above will be described.

In the following, the operation of the noise estimator 40 of the sound collecting device 1, which is a feature of the present invention, will be mainly described.

The band dividing means 41 frequency-analyzes the input signal x to calculate a frequency spectrum (hereinafter also referred to as "input spectrum"). Then, the band dividing means 41 divides the obtained input spectrum into K (divides into K frequency bands), divides the divided input spectrum (hereinafter referred to as "frequency band signal") into each power calculation It is given to means 42 (42-1 to 42-K).

Although the method of frequency analysis processing performed by the band dividing means 41 is not limited, for example, Fast Fourier Transform (FFT), wavelet transform, filter bank, etc. can be applied. It should be noted that FFT is preferably applied to the band dividing means 41 of this embodiment.

The power calculation means 42 (42-1 to 42-K) calculate the input power based on the input frequency band signal and give it to the corresponding noise estimation means 43 (43-1 to 43-K). The power calculators 42-1 to 42-K supply the calculated input powers to the noise estimators 43-1 to 43-K.

In each power calculation means 42, various calculation methods can be applied as power calculation methods. For example, the power calculation means 42 acquires the absolute value of the power of each frequency constituting the input frequency band signal, and calculates the "sum of squares of the acquired absolute values" or "sum of the acquired absolute values" as the input power. You may make it calculate as. Hereinafter, the power output by each power calculating means 42 (42-1 to 42-K) (hereinafter also referred to as "input power") is represented by PX (PX_1 to PX_K).

Each noise estimation means 43 (43-1 to 43-K) calculates the power of the noise component contained in the input power PX (PX_1 to PX_K) supplied from the power calculation means 42 (42-1 to 42-K). Estimate and output the result (hereinafter also referred to as "estimated noise power"). In the following, the estimated noise power output by each noise estimation means 43 (43-1 to 43-K) is represented as PN (PN_1 to PN_K). Since the estimated noise powers PN_1 to PN_K represent the power of each component of the estimated noise, it can be said that the aggregate of the estimated noise powers PN_1 to PN_K represents the signal of the estimated noise. In this embodiment, the PNs (PN_1 to PN_K) output by the noise estimation means 43 (43-1 to 43-K) are supplied to the noise suppressor 20. FIG.

Next, the operation inside each noise estimation means 43 (43-1 to 43-K) will be described.

Since the noise estimating means 43-1 to 43-K are different in the frequency band of the input power PX to be processed and the processing itself is the same, hereinafter, the noise estimating means 43-1 to 43-K will be collectively referred to as noise estimating means. 43 will be described.

In the noise estimating means 43 , the input power PX is supplied to the stationary probability calculating means 101 and the buffer updating means 102 .

First, the operation of the stationary probability calculation means 101 will be described.

The stationary probability calculation means 101 calculates a stationary probability (hereinafter referred to as "SP") regarding the input power PX based on the input power PX and an estimated noise power (hereinafter referred to as "PND") for D samples to be described later. ) is calculated. Then, the stationary probability calculating means 101 gives the obtained stationary probability SP to the buffer updating means 102 . A specific numerical value of D will be described later.

The stationary probability SP is calculated based on the posterior SNR (hereinafter, this posterior SNR value is also referred to as "G"). The posterior SNR(G) can be calculated by equation (1). The steady-state probability SP can be calculated by a steady-state probability calculation function, which will be described later.

The stationary probability calculation function of this embodiment is a function of the stationary probability SP with respect to the posterior SNR(G). It is assumed that a previously designed stationary probability calculation function is set in the stationary probability calculation means 101 .

Next, the design of the steady-state probability calculation function set in the steady-state probability calculation means 101 will be described.

In this embodiment, the domain (G) of the stationary probability calculation function is a real number of 0.0 or more, and the range (SP) is a real number of 0.0 or more and 1.0 or less. The shape of the stationary probability calculation function in this embodiment (the shape of the curve when the stationary probability calculation function is represented by a graph; the shape of the curve in the graph when the horizontal axis is the value of G and the vertical axis is the stationary probability SP) , G is a predetermined value (hereinafter referred to as “G0”) (that is, when G=G0). That is, the steady-state probability calculation function in this embodiment is designed to have characteristics such that the steady-state probability SP increases monotonically in the broad sense in the interval of 0.0≦G≦G0, and the value of the steady-state probability SP monotonically decreases in the interval of G0≦G. shall have been made.

G0 is not particularly limited, and can take any value within the domain of G. However, since the steady-state probability calculation function is a function that outputs a higher probability as the state becomes more steady, it is most preferable to set G0=1.0 (that is, PX=PND).

5 and 6 are explanatory diagrams (graphs) showing examples of stationary probability calculation functions.

The stationary probability calculation function shown in FIG. 5 is a function smoothly defined for G. In FIG.

In the stationary probability calculation function used in this embodiment, G(G0) at which SP=1.0 is not particularly limited. In the example of this embodiment, G (G0) at which SP=1.0 is set to 1.0 (that is, G0=1.0). Further, the stationary probability calculation function used in this embodiment is such that when G is increased toward +∞ (G→+∞), the stationary probability SP asymptotically converges toward 0.0 (SP→0.0 ) (smooth convergence).

FIG. 6 is an explanatory diagram (graph) showing a function when the stationary probability calculation function is designed on a logarithmic scale.

FIG. 6(a) is a graph showing the vertical and horizontal axes of the stationary probability calculation function in logarithmic scale. FIG. 6(b) shows a graph when the stationary probability calculation function designed on a logarithmic scale as in FIG. 6(a) is displayed on a linear scale. As shown in FIGS. 6(a) and 6(b), by designing the stationary probability calculation function on both the vertical and horizontal axes on a logarithmic scale, the relationship between when PX<PND and when PX>PND is , is fair (symmetric), and the stationary probability calculation function can be easily defined over the entire domain of G.

Next, the operation (function) of the buffer updating means 102 will be described.

The buffer update means 102 has a buffer (hereinafter referred to as "input buffer BX") that holds the supplied input power PX (sample; data) and a buffer that holds the supplied stationary probability SP (sample; data) ( hereinafter referred to as a "probability buffer BP"), and supplies the contents of the input buffer BX and the probability buffer BP to the weighted average means 103. That is, the buffer updating means 102 updates the input buffer BX and the probability buffer BP based on the input power PX and the stationary probability SP, and gives the contents of the obtained input buffer BX and the probability buffer BP to the weighted average means 103 .

The input buffer BX is a buffer that holds T samples (data) of the input power PX from the past to the present (T is an arbitrary integer equal to or greater than 1).

The buffer length T may be an arbitrary value, but may be a length corresponding to 100 milliseconds to several seconds (the number of samples corresponding to 100 milliseconds to several seconds) depending on the application of the noise estimation means 43. . In the noise estimation means 43, if the buffer length T of the buffer updating means 102 is too short (for example, if the buffer length T is set to 100 milliseconds), the change in the noise to be estimated will be tracked faster. Depending on the contents of the input power PX (input signal x), noise estimation accuracy may deteriorate. For example, in the noise estimation means 43, if the buffer length T of the buffer updating means 102 is too short, and if the input power PX (input signal x) contains a speech component that is spoken slowly, the speech component will be detected as noise. It may be estimated incorrectly. Also, in the noise estimation means 43, if the buffer length T of the buffer update means 102 is set too long (for example, if the buffer length T is set to 8 seconds), the stationary component in the input signal can be accurately estimated. When the noise environment changes abruptly (for example, when you leave a quiet room and go out into the noisy outdoors, or when you turn on/off the air conditioner all at once), you cannot immediately follow it. Therefore, in the noise estimating means 43, the buffer length T of the buffer updating means 102 is preferably a length corresponding to a range that is neither too short nor too long (for example, 200 milliseconds to 1 second). For example, in consideration of the characteristics of the noise estimating means 43 as described above, a suitable buffer length T may be obtained in advance through experiments or the like and set in the noise estimating means 43 .

The probability buffer BP is a buffer that holds T samples (data) of the stationary probability SP from the past to the present. That is, the buffer length T of the probability buffer BP is the same as the buffer length T of the input buffer BX.

The buffer update means 102 updates the buffers (BX and BP) in FIFO (First In, First Out). That is, the oldest input power PX is deleted from the input buffer BX, and the newly applied input power PX is stored in the input buffer BX. Similarly, the oldest stationary probability SP is deleted from the probability buffer BP and the newly given stationary probability SP is stored in the probability buffer BP.

The weighted average means 103 calculates the estimated noise power PN based on the samples (data) held in the input buffer BX and the probability buffer BP, outputs the obtained estimated noise power PN, and supplies it to the storage means 104. memorize it.

The weighted average means 103 calculates the estimated noise power PN by calculating 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 factor. That is, assuming that the i-th (i is an integer from 1 to T) 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 weighted average means 103 performs the following (2 ) can be used to calculate the estimated noise power PN.

The method of calculating the estimated noise power PN by weighted averaging in the noise estimator 43 has the following merits. For example, in the noise estimating means 43, the input power PX with a high stationary probability SP is given a large weight, so it has a great influence on the calculation of the estimated noise power PN. In the noise estimating means 43, since the input power PX with the low stationary probability SP is given a small weight, its influence on the calculation of the estimated noise power PN is small and is almost ignored. Here, if the noise estimation means 43 completely ignores the input power PX with a small stationary probability SP (for example, Patent Document 1 ignores it if the posterior SNR is larger than a predetermined value), packet loss and voice A problem arises when the input signal is temporarily reduced by the switch and then returned, and cannot be tracked. Therefore, in the noise estimation means 43 of this embodiment, stable estimation can be continued by considering the input power PX with a small stationary probability SP with a small influence.

The storage means 104 holds the samples (data) of the estimated noise power PN for D periods, and provides the stationary probability calculation means 101 with D samples (D past estimated noise power PND). The delay sample number D is preferably set to D=1, but may be set to D>1.

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

The noise estimating unit 40 installed in the sound collecting device 1 of the first embodiment estimates the noise power using a weighted average in which the stationary probability calculated based on Yamagata's stationary probability calculation function is used as a weighting factor. For example, the noise estimating unit 40 mounted on the sound collecting device 1 of the first embodiment uses a weighted average with the stationary probability as a weighting factor even when the noise included in the input signal x suddenly increases. Therefore, of the noise power that stochastically changes, small noise power that appears infrequently is used to gradually follow, so that the estimated noise power does not suddenly increase. That is, in the noise estimation unit 40 mounted in the sound collecting device 1 of the first embodiment, even if the noise changes abruptly, the estimated noise power does not change abruptly. It is possible to provide a noise estimation method capable of stably estimating without stopping.

Further, in the noise estimation unit 40 installed in the sound collecting device 1 of the first embodiment, since the stationary probability calculation function has a chevron shape, when the input voice contained in the input signal x suddenly decreases is regarded as non-stationary, there is no excessive tracking of the input speech that has temporarily decreased, and the problem of the estimated noise power remaining small does not occur.

Furthermore, in the noise estimation unit 40 installed in the sound collection device 1 of the first embodiment, the only parameters that need to be set are the buffer length T and the number of delay samples D, so noise estimation processing that does not require adjustment of estimation parameters is realized. can.

As described above, the noise estimating unit 40 installed in the sound collecting device 1 of the first embodiment estimates the noise power without adjusting the estimation parameters and without being affected by abrupt changes or changes in the input speech. be able to.

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

(B-1) Configuration of Second Embodiment The configuration of the sound collecting device 1A (including the noise estimation unit 40A) of the second embodiment can also be shown using FIGS. 1 to 4 described above. 1 to 4 are used only in the second embodiment.

The difference between the second embodiment and the first embodiment will be described below.

As shown in FIG. 2, the sound collecting device 1A of the second embodiment differs from the first embodiment in that the noise estimator 40 is replaced with a noise estimator 40A.

Further, as shown in FIG. 3, in the noise estimator 40A of the second embodiment, the noise estimator 43 (43-1 to 43-K) is the noise estimator 43A (43A-1 to 43A-K). It differs from the first embodiment in that it is replaced.

Furthermore, as shown in FIG. 1, in the noise estimating means 43A (43A-1 to 43A-K) of the second embodiment, the buffer updating means 102 is replaced with the buffer updating means 102A. is different from

Next, differences from the first embodiment will be described with respect to the buffer updating means 102A in the second embodiment.

The buffer updating means 102 of the first embodiment always updates the input buffer BX and the probability buffer BP. However, in the operation of the buffer updating means 102 in the first embodiment, a large input (the power of the input signal x is large) or a small input (the power of the input signal x is small) is longer than the period corresponding to the buffer length T. continues for a long time, even though the absolute values of the stationary probabilities SP stored in the probability buffer BP are small, all the stationary probabilities SP are small, so each has a large influence. Therefore, in the operation of the buffer updating means 102 in the first embodiment, it cannot be denied that there is a possibility that a large input or a small input will be rapidly followed as a result.

Therefore, the buffer update means 102A of the second embodiment differs from the first embodiment in that it does not update the input buffer BX and the probability buffer BP when the stationary probability SP is extremely small (smaller than a predetermined value). different. As a result, the noise estimator 40A (noise estimator 43A) of the second embodiment can estimate the noise power more stably than the first embodiment.

(B-2) Operation of Second Embodiment Next, the operation of the sound pickup device 1A (noise estimation unit 40A) having the configuration as described above will be described.

The second embodiment will be described below, focusing on differences from the first embodiment. As described above, the main difference between the first embodiment and the second embodiment is that the buffer updating means 102 is replaced with the buffer updating means 102A. Therefore, the operation of the buffer updating means 102A will be mainly described below.

The buffer updating means 102A updates the input buffer BX and the probability buffer BP based on the input power PX and the stationary probability SP, and provides the weighted average means 103 with the contents of the obtained input buffer BX and probability buffer BP.

Buffer updating means 102A updates each buffer (input Update the buffer BX and the probability buffer BP). The update operation of each buffer (input buffer BX and probability buffer BP) in the buffer update means 102A is the same as that of the buffer update means 102 of the first embodiment.

If the value of SP0 is too large, the advantage of weighted averaging described in the explanation of the weighted averaging means 103 of the first embodiment is lost, so SP0 should preferably be a small value (not too large). For example, SP0 is preferably set to a value in the range of 0.001 to 0.05.

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

In the noise estimator 40A (buffer update means 102A) installed in the sound collection device 1A of the second embodiment, when the steady-state probability SP is very small (smaller than a predetermined value), the input buffer BX and the probability buffer This differs from the first embodiment in that the BP is not updated. As a result, the noise estimation unit 40A installed in the sound collecting device 1A of the second embodiment can stably estimate the noise power even when the power of the input signal x continues to be high or low for a long period of time. It has the effect of

(C) Other Embodiments The present invention is not limited to the above-described embodiments, and modified embodiments such as those illustrated below can be exemplified.

(C-1) In each of the above embodiments, one microphone is installed for the sake of simplicity. A configuration in which a microphone array is arranged may also be used. In that case, the noise estimation units 40 and 40A perform noise estimation processing on each input signal from a plurality of microphones.

(C-2) In each of the above embodiments, an example in which the noise estimation units 40 and 40A are installed in the sound collecting devices 1 and 1A has been described, but the devices in which the noise estimation units 40 and 40A are installed are not limited. be. Also, the noise estimation units 40 and 40A may be configured as independent devices.

Therefore, in each of the above embodiments, the output destination and output method (for example, data format and output interface) of the estimated noise power PN (PN_1 to PN_K) acquired by the noise estimation units 40 and 40A are not limited. The noise estimation units 40 and 40A may be output to output destinations according to the purpose of the noise estimation units 40 and 40A in accordance with the output method. For example, the noise estimating units 40 and 40A may output using an interface provided in the computer 200 (for example, a signal line on a circuit, a serial interface, etc.), or a wired or wireless communication interface (for example, a wired /wireless LAN interface, various serial interfaces, etc.).

(C-3) In the sound collecting device 1, the signal input unit 10, the noise suppression unit 20, and the noise estimation unit 40 or 40A, when the processing of the noise suppression unit 20 is performed in the frequency domain, the band division means is replaced by the noise suppression unit. 20, the configuration of the band division means 41 in the noise estimation section 40 or the output of the band division means 41 may be shared. For example, the signal input section 10 may include band division means, and the output of the band division means may be supplied to the noise suppression section 20 and the noise estimation section 40 as the output of the signal input section 10 .

Reference Signs List 1 sound collection device M microphone 10 signal input unit 20 noise suppression processing unit 30 signal output unit 40 noise estimation unit 41 band division means 42, 42-1 to 42-K Power calculating means 43, 43-1 to 43-K Noise estimating means 101 Stationary probability calculating means 102 Buffer updating means 103 Weighted average means 104 Storage means 200 Computer 201 Processor , 202... primary storage unit, 203... secondary storage unit.

Claims (6)

A noise estimator that obtains an estimated noise power by estimating noise components contained in an input signal in which speech and noise are mixed,
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;
an input buffer for buffering data of the past input power; and a probability buffer for buffering the steady-state probability calculated by the steady-state probability calculation means in the past. buffer update means for updating the content of the input buffer and the content of the probability buffer based on the stationary probability;
and weighted averaging means for calculating the estimated noise power by taking the weighted average of the input powers held in the input buffer using the stationary probability held in the probability buffer as a weighting factor. Noise estimator. 2. The noise estimation apparatus according to claim 1, wherein said stationary probability calculation means calculates said stationary probability using a stationary probability calculation function regarding a posteriori SNR based on said input power and said past estimated noise power. . 3. The buffer updating means according to claim 1, wherein said buffer updating means updates said input buffer and said probability buffer only when said stationary probability given from said stationary probability calculating means is greater than a predetermined value. noise estimator. A computer installed in a noise estimator that obtains an estimated noise power by estimating noise components contained in an input signal in which speech and noise are mixed,
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;
an input buffer for buffering data of the past input power; and a probability buffer for buffering the steady-state probability calculated by the steady-state probability calculation means in the past. buffer update means for updating the content of the input buffer and the content of the probability buffer based on the stationary probability;
It functions as weighted averaging means for calculating the estimated noise power by taking a weighted average of the input powers held in the input buffer using the stationary probability held in the probability buffer as a weighting factor. A noise estimation program for . In a noise estimation method performed by a noise estimation device for obtaining estimated noise power by estimating noise components contained in an input signal in which speech and noise are mixed,
having stationary probability calculating means, buffer updating means, and weighted averaging means;
The stationary probability calculation means calculates the stationary probability of the input power from the input power of the input signal and the past estimated noise power,
The buffer update means includes an input buffer for buffering past input power data and a probability buffer for buffering the steady-state probability calculated by the steady-state probability calculation means in the past. updating the content of the input buffer and the content of the probability buffer based on the stationary probability calculated by the calculating means;
The weighted averaging means calculates the estimated noise power by taking a weighted average of the input powers held in the input buffer using the stationary probability held in the probability buffer as a weighting factor. Noise estimation method. A sound pickup device that picks up the sound from an input signal in which speech and noise are mixed,
a noise estimation unit that estimates a noise component included in the input signal;
a sound collecting unit that extracts and collects the voice from the input signal using the estimation result of the noise estimating unit;
A sound collecting device, wherein the noise estimating device according to any one of claims 1 to 3 is applied as the noise estimating unit.

Images