JP2003316381A - Method and program for restricting noise - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To obtain a noise restricting effect with high precision by reducing a spectrum residual. <P>SOLUTION: Noise spectrums of each frame from a first frame to a T-th frame are obtained in a step S1. Then an average spectrum N and a standard deviation V are obtained in the noise in a step S3. In a step S4, a noise restriction amount D(ω, t) is calculated concerning each frequency ω by defining a present frame as the t-th frame (t>T). In this case, the noise restriction amount D(ω, t) is calculated based on a present input spectrum X(t), and the average value N and the standard deviation V of the noise. In a step S5, a voice spectrum is estimated through the use of S(t)=X(t)-D(t). The noise restriction amount D(ω, t) is calculated based on not only the present input spectrum X(t) and the noise average value N but the noise standard deviation (the square root of a variance) V. Thus, the restriction amount D(ω, t) is made to correspond further to an actual environmental noise, so that the spectrum residual is sufficiently reduced. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise suppression method and a noise suppression program for highly accurately suppressing noise from a voice uttered in a noise environment.

[0002]

2. Description of the Related Art In recent years, as the performance of voice recognition technology has improved,
Practical application of voice recognition engine in real environment is becoming active. Especially, in a situation where input devices such as a car navigation system and a mobile device are limited, expectations for voice recognition are great.

In the voice recognition process, the voice recognition result is obtained by comparing the input voice taken from the microphone with the vocabulary to be recognized. In a real environment, since there are various noise sources, environmental noise is mixed in the voice signal captured by the microphone. In speech recognition processing, noise resistance greatly affects recognition accuracy.

Spectral subtraction (hereinafter also referred to as SS) is widely used as a noise suppression method for a voice signal (spectrum) uttered in such a noise environment.

The most basic algorithm for noise suppression by SS is shown below. Basically, the SS predicts the noise level in advance according to the average level of the observed noise.
Then, the noise level is suppressed by subtracting the predicted noise level from the input signal.

That is, first, prior to noise suppression, a time series {N (t)}, N of a noise spectrum containing no speech at all.
(t) = (N (ω ₁ , t), N (ω ₂ , t), ..., N (ω _d , t))
From (where ω is the frequency and t is the time, that is, the frame number), the average vector N _{ave of} noise is calculated by the following equation (1).

[0007] _{N ave = Σ t N (t} ) / T ... (1) where, sigma _t denotes the sum of a suitable time interval, T is the length of the time interval (number of frames).

Next, the time series {X (t)} of the voice spectrum in which noise is mixed is expressed as X (t) = (X (ω ₁ , t), X (ω ₂ , t),
, X (ω _d , t)), and the estimated amount S (t) of the speech spectrum containing no noise is S (t) = (S (ω ₁ , t), S (ω ₂ , t),
, S (ω _d , t)) using the suppression coefficient α as shown in the following equation (2), the input signal X (ω,
The estimated value S (ω, t) of the speech power spectrum in which the noise component is suppressed is calculated from t).

S (ω, t) = X (ω, t) −αN _ave (ω) (when X (ω, t) −αN _ave (ω)> 0) = 0 (X (ω, t) − In the case of α N _ave (ω) ≤ 0) (2) Regarding SS, reference 1 (Jean-Claude Janqua,
ROBUSTNESS IN AUTOMATIC SPEECH by Jean-Paul Haton
RECOGNITION ”Kluwer Academic Publishers).

[0010]

By the way, the above-mentioned S
In the S algorithm, processing is performed assuming that the environmental noise is constant. That is, the suppression coefficient α in the above equation (2) is processed as a fixed value. However, since the spectrum of noise is not constant in general,
It is possible to obtain a higher noise suppression effect by making the suppression coefficient α variable according to the environment. The following formula (3) shows the reason for this.

That is, the above-mentioned estimated value S (ω, t) of the voice spectrum by SS can be shown in more detail by the following equation (3).

S (ω, t) = X (ω, t) -αN _ave (ω) = S ₀ (ω, t) + N _R (ω, t) -αN _ave (ω) = S ₀ (ω, t) + (R (ω, t) −aN _ave (ω)) + (N (ω) − (1 + b) N _ave (ω)) (3) Note that X (ω, t) = S ₀ (ω, t) + N _R (ω, t) ＝ S ₀ (ω, t) + N
(ω, t) + R (ω, t) N _R (ω, t) = X (ω, t) −S ₀ (ω, t) = R (ω, t) + N
(ω, t) R (ω, t) = 2 ((S ₀ N) ^1/2 ) cos θ (ω, t) α = 1 + a + b, and X (ω, t) ... Input signal spectrum S ₀ (ω, t) ... True voice spectrum N _R (ω, t) ... Non-voice component of input signal (consisting of noise and correlation between noise and voice) N (ω, t) ... Noise spectrum R (ω, t) ) ... It is the correlation between the voice signal and the noise signal.

Incidentally, for example, Japanese Patent Laid-Open No. 2001-228892
As described in Japanese Patent Publication (hereinafter referred to as Document 2), generally, when a constant is used as the suppression coefficient α, the suppression coefficient α is (a + b) as shown in Expression (3).
It is known that a value larger than 1 should be set by the amount of.

_NR (ω, t) -αN _ave (ω) in the above equation (3)
Indicates the spectral residual which is the residual of the noise component. The spectral subtraction technique enables high-speed processing with a simple configuration by processing on the assumption that environmental noise is constant. However,
In reality, environmental noise fluctuates. Therefore, if the suppression coefficient α is a constant, the spectrum residual N _R (ω, t) −αN
_{Since ave} (ω) varies with time, sufficient noise suppression accuracy cannot be obtained.

Therefore, in Reference 1, the suppression coefficient α is set in order to make the term of the spectrum residual closer to 0.
The method of making variable is considered. That is, in Document 1, noise to speech S / N ratio (Signal to Noise Rati
The suppression coefficient α is determined according to o).
However, this method still does not provide sufficient noise suppression accuracy.

[0016] Also, in Document 2, a method of varying the suppression coefficient α is disclosed, but even in the method of Document 2, sufficient noise suppression accuracy is not obtained.

It is an object of the present invention to provide a noise suppression method and a noise suppression program that can improve noise suppression accuracy by making the noise suppression coefficient variable according to the variance of noise.

[0018]

According to a first aspect of the present invention, there is provided a noise suppression method, wherein a noise mean vector and standard deviation are determined from a spectral time series of a noise-only input signal, and a spectral residual in spectral subtraction. A noise suppression amount determining procedure for determining an amount of noise suppression based on an input signal mixed with noise and an average vector and standard deviation of the noise so as to reduce the difference, and the noise suppression amount from the input signal mixed with noise. And a step of suppressing the noise of the input signal by subtracting the noise. The noise suppression method according to claim 2 of the present invention comprises: , A procedure for dividing the noise into a cluster, a procedure for obtaining a noise average vector and a standard deviation for each cluster from the noise-only input signal, and noise A cluster that most closely approximates the input spectrum using a predetermined distance measure from the input spectrum of the input signal, a procedure for obtaining the average vector and standard deviation of the selected cluster, and reducing the spectral residual in the spectral subtraction. , A noise suppression amount determining procedure for determining an amount of noise suppression based on an input signal containing noise, and an average vector and standard deviation of the selected clusters, and subtracting the amount of noise suppression from the input signal containing noise. Therefore, a procedure for suppressing the noise of the input signal is provided.

In the first aspect of the present invention, first, prior to suppression of noise with respect to the input signal, the average vector and standard deviation of noise are obtained from the spectral time series of the input signal containing only noise. The amount of noise suppression in the spectral subtraction is determined based on the input signal in which noise is mixed and the average vector and standard deviation of the noise.
The noise of the input signal is suppressed by subtracting the determined noise suppression amount from the noise-containing input signal. The amount of noise suppression is determined using the standard deviation of noise, and the spectral residual in spectral subtraction is reduced.

In claim 2 of the present invention, the spectral time series of the noise-only input signal is divided into a plurality of clusters of a predetermined number of clusters. Then, the average vector and standard deviation of noise are obtained for each cluster. Among the clusters, the cluster closest to the input spectrum of the noise-containing input signal is selected using a predetermined distance measure, and the average vector and standard deviation of the selected clusters are obtained. The amount of noise suppression in spectral subtraction is determined based on the input signal containing noise and the average vector and standard deviation of the selected clusters.
The noise of the input signal is suppressed by subtracting the noise suppression amount from the input signal containing the noise. The noise is divided into clusters, the cluster closest to the input spectrum is selected, the average vector and standard deviation are obtained, and the noise suppression amount is more adapted to the actual noise environment, and the spectral residual is more Will be reduced.

The present invention relating to the method is also realized as a program for causing a computer to execute the processing corresponding to the present invention.

[0022]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a flowchart showing a noise suppression method according to an embodiment of the present invention.

The present embodiment is applied to noise suppression using spectral subtraction adopted for voice recognition and the like. The above equation (3) showing the estimated value S (ω, t) of the speech spectrum in the spectral subtraction is
It shows the possibility of reducing the spectrum residual by setting the suppression coefficient α appropriately.

Here, to clarify the argument, the correlation term R (ω, t) -aN _ave (ω) can be almost ignored,
Considering the case where b is 0, the estimated value S (ω, t) of the speech spectrum can be expressed by the following equation (4).

S (ω, t) = S ₀ (ω, t) + (N (ω, t) -N _ave (ω)) (4) The variation term N (ω, t) in the equation (4) −N _ave (ω) means a deviation from the average value of the noise component in the current frame t. Therefore, it can be seen that the variation term has a strong correlation with the variance, which is one of the statistical properties of noise. That is, in the case of noise with small variance, the absolute value of the variation term is small on average, and conversely, in the case of noise with large variance, the absolute value of the variation term is averagely large.

When this is applied to the equation (3), it is better to set the suppression coefficient α to a value close to 1 for noise with small variance, and conversely, to increase the suppression amount by increasing α when the variance is large. It is speculated that it is preferable. At this time, if α is increased for noise with large variance, N (ω, t)
When N is larger than N _ave (ω), the variation term becomes smaller, but in the opposite case, the variation term becomes larger in the negative direction.
On the contrary, there is a risk of adverse effects. However, as described above, in the case of SS that uses a constant as α, it is known that it is better to make the value larger than 1 to increase the noise suppression amount. Therefore, although there is a disadvantage when the variation term becomes negative, it can be expected that it is better to increase α.

For these reasons, in the present embodiment, α is made variable according to the variance of noise, and the residual (variation term) is changed.
Is designed to be small. In the present embodiment, the amount of noise suppression is variable according to the standard deviation which is the square root of the variance.

FIG. 1 shows the algorithm of the entire noise suppression process.

Now, let the time series of the input spectrum be {X (t)
}. Note that t is a frame number, and t = 1,2,
3, ... Further, ω represents the frequency. Here, it is assumed that voice is not mixed and only noise is input from the first frame of the input signal to at least the T-th frame (T is a predetermined constant). FIG. 2 is a spectrum diagram showing an input spectrum time series that satisfies such a condition, where the horizontal axis represents time in frame units and the vertical axis represents frequency.

FIG. 2 shows the signal level in each frequency band for each frame by shading. The dark part shows that the signal level is high, and the light part shows that the signal level is low.

From time 0 to T frame in FIG. 2, frames in which the input signal is guaranteed to be only noise are shown, and the frames after T frame are the frames to be subjected to noise suppression. Note that the time zone having a relatively dark portion in the center of FIG. 2 shows the spectrum of the input voice when “Hachinohe” is actually uttered.

First, the frame number t is initialized to 1, and the noise spectrum of the first frame is acquired in step S1 of FIG. The noise spectrum of each frame from the first frame to the T-th frame is acquired by determining whether t has reached T while incrementing t in step S212 (step S2).

Next, in step S3, the average spectrum N = (N (ω ₁ ), N (ω ₂ ), ..., N (ω _d )) of noise and the standard deviation V = (V (ω ₁ ), V (ω ₂ ), ..., V (ω _d )) is obtained by the following equations (5) and (6).

[0034] _{N (ω) = Σ t X} (ω, t) / T ... (5) V (ω) = {Σ t X (ω, t) 2 / T - N (ω) 2} 1/2 ... (6) The first to T-th frames are frames for calculating noise statistics, and are frames that are not the target of noise suppression. The speech spectrum {S (t)} t = T estimated by the noise suppression processing for the (T + 1) th frame and thereafter.
+1, T + 2, ... Are output.

The current frame is the t-th frame (t> T)
Then, the noise suppression amount D (ω, t) is calculated for each frequency ω (step S4). In the present embodiment, the noise suppression amount D (ω, t) is calculated based on the current input spectrum X (t), the average value N of noise, and the standard deviation V.

Next, in step S5, the following (7)
The speech spectrum S (t) is estimated from the equation.

S (t) = X (t) -D (t) (7) Next, at step S6, t is incremented by
The noise suppression process of steps S3 to S5 is repeated for the input voice until the end of the input is confirmed in step S7.

In the embodiment configured as above, the noise suppression amount D (ω, t) is calculated as the current input spectrum X.
It is calculated based on not only (t) and the average value N of noise, but also the standard deviation (square root of variance) V of noise. As a result, the noise suppression amount D (ω, t) further corresponds to the actual environmental noise, and the spectrum residual can be sufficiently reduced.

As described above, in this embodiment, since the noise suppression amount is set according to the noise variance,
It has an effect that the noise suppression effect can be sufficiently improved.

FIG. 3 is a flow chart showing another embodiment of the present invention. In FIG. 3, the same steps as those in FIG.

The present embodiment is capable of improving the accuracy of noise suppression by performing noise spectrum clustering.

Also in the present embodiment, the time series of the input spectrum is {X (t)}, and no voice is mixed from the first frame of the input signal to at least the Tth frame,
It is assumed that only noise is input.

In steps S1, S2 and S212 of FIG. 3, the noise spectrum of each frame from the first frame to the Tth frame is acquired. Next, step S1 in FIG.
In S1 and S12, noise spectrum clustering is performed on the noise from the first frame to the Tth frame. The number of noise clusters to be estimated is M _e , and the standard deviation is obtained for each cluster. The average value N (ω) of noise and the standard deviation V (ω) of each cluster are calculated in the same manner as the above equations (5) and (6).

FIG. 4 is a flow chart for explaining a concrete process of clustering in the processes S11 to S13 in FIG.

An example of the cluster calculation method when the maximum number of noise clusters is M (≧ 2) is shown.

First, the input (noise) spectra from the first frame to the Tth frame are stored in an appropriate buffer (steps S21, S211, S212). After the spectrum of the T-th frame is stored, the process proceeds from step S21 to step S22 to cluster T noise spectra.

As a clustering method, for example,
The k-means method described in detail in "Cognitive Engineering", Corona Publishing, edited by the Television Society is used. In this method, first, predetermined M clusters C = {N _1, N _2, ..., N _M }, N _m =
(N _m (ω ₁ ), N _m (ω ₂ ), ..., N (ω _d )), m = 1
2, ..., M are created.

Next, a predetermined distance measure is examined by using L (for example, Euclidean distance) to determine whether or not the number of clusterings determined appropriately is appropriate. When the distance between any two clusters is smaller than a predetermined threshold value, it is determined that these clusters should originally be one cluster.

That is, assuming that a predetermined distance measure is L (for example, Euclidean distance), a pair of two different centroids (mean vectors of noise) in the cluster C having the smallest distance L among all combinations. (Step S
twenty four). Here, the minimum distance combination is N _k and N _l .

If the distance L (N _k , N _l ) is less than or equal to a predetermined value, the center of gravity of these two combinations is calculated by the following equation N _h = (n _k N _k + n _l N _l ) / ( n _k + n _l ) n _k and n _l are respectively calculated by the number of noise spectra belonging to the centroid N _k , and these two centroids are merged (step S26) to obtain clusters C to N _k ,
N _l is deleted and N _h is newly added (step S27).

At this time, the number of centroids of C (elements of C)
Is reduced to M-1. Also, until now N _k, N_l Belongs to
Vector is N_hAnd the number of them is n_h
= N_k+ N_lIs.

Similarly, step S is repeated until there is one centroid for the new cluster or the distance between the nearest centroids becomes larger than a predetermined value.
The above process is repeated while making a determination in S23 and S25.

On the contrary, when the distance L (N _k , N _l ) is larger than a predetermined value, the cluster C is a set of average vectors of noise. In this case, the process shifts from step S25 to step S28 to calculate the standard deviation Vm of each cluster.

That is, V _m (ω) = {Σ _τ (N (ω, τ) -N _m (ω)) ² / n _m } ^{1 / 2} τ represents the frame number of the noise spectrum belonging to the cluster m, so that the standard deviation V _m = (V
_m (ω ₁ ), V _m (ω ₂ ), ..., V (ω _d )) are obtained (step S 28).

When selecting the average vector N and the standard deviation V when calculating the noise suppression amount for the input vector X (t), e = argmin _m (L (X (t) , N _m )), N = N _e , V = V _e, and the suppression amount D (t) is calculated using these.

As described above, in this embodiment, the noise spectrum is clustered, and the cluster used for calculating the standard deviation is determined based on which cluster the spectral pattern of the input signal is most similar to. Thereby, the calculation accuracy of the noise suppression amount can be improved, and the noise suppression effect can be further improved.

Next, a third embodiment of the present invention will be described.

In the first and second embodiments, the noise suppression amount D (ω, t) is not limited to the current input spectrum X (t) and the noise mean value N, but also the noise standard deviation (variance). The point calculated based on the (square root of) V. First and second
In the embodiment, any method can be adopted as the method of calculating the noise suppression amount as long as the spectrum residual can be reduced.

In this embodiment, the following equation (8) is adopted as a concrete calculation equation of the noise suppression amount D (ω, t).

D (ω, t) = α (ω, t) V (ω) + βN (ω) (when X (ω, t)> α (ω, t) V (ω) + βN (ω) ) = X (ω, t) (when X (ω, t) ≦ α (ω, t) V (ω) + βN (ω)) (8) In the above equation (8), the input signal X (ω , t) and βN (ω), which is β times the average value of noise, are less than or equal to α (ω, t) times the standard deviation, D (ω, t) = X (ω, t) To do. That is, S (ω,
t) = X (ω, t) -D (ω, t) = X (ω, t) -X (ω, t) =
Since it is 0, it means that the voice component of the input signal is estimated to be 0.

On the other hand, when the input signal is sufficiently large and the difference from βN (ω), which is β times the average value of noise, is α (ω, t) times the standard deviation or more, the estimated speech spectrum S (ω, t) ) Can be expressed by the following equation (9).

S (ω, t) = X (ω, t) -βN (ω) -α (ω, t) V (ω) (9) That is, in this case, the noise suppression amount is the standard deviation V It is proportional with respect to (ω).

Therefore, the variance of noise (= square of standard deviation)
This is a noise suppression method that clearly realizes the purpose of increasing the noise suppression amount when is large, and decreasing the suppression amount when is small.

Here, as α (ω, t), α (ω, t) = F (ω, t) / V (ω) (10) F (ω, t) is V (ω) Note that the form of independent functions or constants is excluded. Because this α
Substituting (ω, t) into the above equation (8), D (ω, t) = βN (ω) + F (ω, t) (X (ω, t)> βN (ω) + F (ω, t) ) = X (ω, t) (X (ω, t) ≦ βN (ω) + F (ω, t)) (11), and the noise suppression amount depends on the standard deviation V (ω). Because it will not do.

The noise suppression method based on the equation (11) is that the noise suppression amount is proportional to the standard deviation of noise.
The noise suppression amount that depends on the variance of noise (square of standard deviation) is most easily realized.

FIGS. 5 and 6 are graphs showing the fourth embodiment of the present invention. In FIG. 5, x is on the horizontal axis and D is on the vertical axis.
7 is a graph showing the noise suppression amount D (ω, t) by taking (ω, t), and FIG. 6 shows X (ω, t) on the horizontal axis and S (ω, t) on the vertical axis. 6 is a graph showing a speech spectrum S (ω, t).

This embodiment adopts the same noise suppression algorithm as that of the first embodiment, and calculates the noise suppression amount D (ω, t) shown in the above equation (8) and the following (1
The calculation formula shown in Formula 2) is adopted.

[0068] The noise suppression amount D (ω, t) and the estimated speech spectrum S (ω, t) when the above equations (8) and (12) are adopted.
Can be expressed by the following equations (13) and (14), respectively.

D (ω, t) = βN (ω) + kV (ω) (when x ≦ k) = βN (ω) + (k + a) V (ω) (when x ≧ k + b) = βN (ω) + (A (x−k) / b + k) V (ω) (other cases) where x = (X (ω, t) −βN (ω)) / V (ω) (13) S (ω, t) = 0 (when X (ω, t) <βN (ω) + kV (ω)) = X (ω, t) −βN (ω) − (k + a) V (ω) (X (ω, t) > ΒN (ω) + (k + b) V (ω)) = X (ω, t) −βN (ω) − (a (x−k) / b + k) V (ω) (other cases) (14) Here, k, a, and b are constants obtained experimentally, and a,
b is a non-negative value.

The thick line in FIG. 5 indicates the noise suppression amount D (ω, t) based on the above equation (13), and the thick line in FIG. 6 indicates the speech spectrum S (ω, t) based on the above equation (14). There is.

The above equation (14) is classified into three cases according to the value of the difference between the input signal X (ω, t) and the noise average value N (ω). In the first case, that is, when the difference between the input signal X (ω, t) and the average value N (ω) of noise is less than or equal to k times the standard deviation V (ω), S (ω, t) = It has become 0. This means that the difference between the average value of the input signal and the noise originates from the noise fluctuation (time fluctuation), and the voice component of the input signal is estimated to be zero.

In the second case, that is, when the input signal becomes (k + b) times larger than the variance from the average value, it cannot be said that the signal may have suddenly become large in noise. It is determined that there is a high possibility that a large voice signal has been added to the noise signal, and a relatively large value N (ω) + (k + b) V (ω)
Subtract.

In the third case, that is, the input signal X (ω, t) is larger than the average value of noise by k times the standard deviation, but (k +
If it is smaller than b) times, it means that weak voice noise is likely to be mixed in the noise signal, which means that a so-called moderate noise suppression amount is subtracted according to the difference between the average value of the input signal and the noise. To do.

As described above, in the present embodiment, the above equation (8) can be embodied in the simplest form. Therefore, there is an advantage that it can be easily implemented in practice and the amount of calculation is small.

[0075]

As described above, according to the present invention, the noise suppression precision can be improved by making the suppression coefficient variable according to the variance of noise.

[Brief description of drawings]

FIG. 1 is a flowchart showing a noise suppression method according to an embodiment of the present invention.

FIG. 2 is a spectrum diagram showing an example of an input spectrum time series in which the horizontal axis represents time in frame units and the vertical axis represents frequency.

FIG. 3 is a flowchart showing another embodiment of the present invention.

FIG. 4 is a flowchart for explaining a specific process of clustering in processes S11 to S13 in FIG.

FIG. 5 is a graph showing a fourth embodiment of the present invention.

FIG. 6 is a graph showing a fourth embodiment of the present invention.

[Explanation of symbols]

S1 ... Noise spectrum acquisition processing, S3 ... Average spectrum and standard deviation calculation processing, S4 ... Noise suppression amount calculation, S5
... estimation of the speech spectrum. Proxy Patent Attorney Susumu Ito

Claims (8)

[Claims] 1. A procedure for obtaining an average vector and standard deviation of noise from a spectral time series of an input signal containing only noise, and an input signal containing noise and the noise so as to reduce a spectral residual in spectral subtraction. A noise suppression amount determining procedure for determining a noise suppression amount based on an average vector and a standard deviation, and a procedure for suppressing noise of the input signal by subtracting the noise suppression amount from an input signal mixed with noise. A noise suppression method characterized by the above. 2. A step of dividing a spectral time series of a noise-only input signal into a plurality of clusters of a predetermined number of clusters, and a step of obtaining an average vector and standard deviation of noise for each cluster from the noise-only input signal. And a procedure for selecting a cluster that is the closest to the input spectrum using a predetermined distance measure from the input spectrum of the input signal containing noise, and obtaining the average vector and standard deviation of the selected cluster, and the spectrum residual in the spectral subtraction. A noise suppression amount determining procedure for determining a noise suppression amount based on an input signal containing noise and an average vector and standard deviation of the selected clusters so as to reduce the difference; And a step of suppressing the noise of the input signal by subtracting the suppression amount. Noise suppression method. 3. A spectral time series {N (t)} t = 1,2, ..., N of a noise-only signal prior to noise suppression of an input signal.
(t) = (N (ω ₁ , t), N (ω ₂ , t), ..., N (ω _d , t)) (d-dimensional vector, where ω ₁ , ω ₂ etc. are vectors corresponding to frequencies The average vector N = (N of the noise spectrum time series {N (t)}
(ω ₁ ), N (ω ₂ ), ..., N (ω _d )) and the standard deviation vector V = (V (ω ₁ ), V (ω ₂ ), ..., V (ω _d )) for each frequency
, And the current input spectrum X (t) = (X
(ω ₁ , t), X (ω ₂ , t), ..., X (ω _d , t)) and the noise suppression amount D (for the current frame t based on the average vector N and the standard deviation vector V t) = (D (ω ₁ , t), D (ω ₂ ,
t), ..., D (ω _d , t)), and the spectrum of the input signal S (t) = (S (ω ₁ ,
t), S (ω ₂ , t), ..., S (ω _d , t)), where S (t) = X (t)
A noise suppression method comprising: a procedure of estimating by D (t). 4. A spectral time series {N (t)} t = 1,2, ..., N of a noise-only signal prior to noise suppression of an input signal.
(t) = (N (ω ₁ , t), N (ω ₂ , t), ..., N (ω _d , t)) (d-dimensional vector, where ω ₁ , ω ₂ etc. are vectors corresponding to frequencies Of the noise spectrum time series {N
(t)} is divided into M _e clusters that do not exceed a predetermined number M, and for each cluster m (m = 1,2, ..., M _e (≦ M)),
Average vector of noise (centroid) N _m = (N _m (ω ₁ ),
N _m (ω ₂ ), ..., N _m (ω _d )) and the standard deviation vector V _m = (V _m (ω ₁ ), V _m (ω ₂ ), ..., V
_Of the procedure for estimating _m (ω _d )) and N _m (m = 1,2, ..., M _e ), the current input spectrum X (t) = (X (ω ₁ , t), X (ω ₂ ,
t), ..., X (ω _d , t))
And a corresponding standard deviation V, and the current input spectrum X (t) = (X
(ω ₁ , t), X (ω ₂ , t), ..., X (ω _d , t)) and the noise suppression amount D (for the current frame t based on the average vector N and the standard deviation vector V t) = (D (ω ₁ , t), D (ω ₂ ,
t), ..., D (ω _d , t)), and the spectrum of the input signal S (t) = (S (ω ₁ ,
t), S (ω ₂ , t), ..., S (ω _d , t)), where S (t) = X (t)
A noise suppression method comprising: a procedure of estimating by D (t). 5. The noise suppression amount D (t) in the input spectrum X (t) of the current frame and the obtained average spectrum N _m (m = 1, 2, ..., _Me ) of the noise. , A coefficient α (t) = (α (ω ₁ , t), α (ω) calculated from the average vector N of noises closest to X (t) and the standard deviation V corresponding to N for a predetermined distance measure. ₂ , t),…, α (ω _d , t))
And by the predetermined constants β and γ, D (ω, t) = α (ω, t) V (ω) + βN (ω) (X (ω, t)> α (ω, t) V (ω) + βN (ω)) = X (ω, t) (X (ω, t) ≤ α (ω, t) V (ω) + βN (ω)) Item 5. The noise suppression method described in either item 3 or 4. 6. The coefficient α (ω, t) is calculated by using a predetermined constant k, a non-negative constant a and a non-negative constant b, The noise suppression method according to claim 5, wherein the noise suppression method is determined by 7. A process for calculating a mean vector and standard deviation of noise from a spectral time series of an input signal containing only noise, and an input signal containing noise so as to reduce a spectral residual in spectral subtraction. Noise suppression amount determination processing for determining a noise suppression amount based on the average vector and standard deviation of the noise, and processing for suppressing the noise of the input signal by subtracting the noise suppression amount from an input signal containing noise A noise suppression program for executing and. 8. A process for causing a computer to divide a spectral time series of a noise-only input signal into a plurality of clusters having a predetermined number of clusters, and an average vector and standard deviation of noise for each cluster from the noise-only input signal. And a process of selecting a cluster that is the closest to the input spectrum using a predetermined distance measure from the input spectrum of the input signal containing noise, and calculating the average vector and standard deviation of the selected cluster, and spectral subtraction. Noise suppression amount determination processing for determining the noise suppression amount based on the input signal mixed with noise and the average vector and standard deviation of the selected clusters so as to reduce the spectrum residual in the input signal mixed with noise. And a process of suppressing the noise of the input signal by subtracting the noise suppression amount from Because of the noise suppression program.