JP2000105599A - Noise level time variation coefficient calculating method, device thereof, and noise reducing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To effectively operate even against a voice superposed with noise varying the level with the lapse of time. SOLUTION: An :put signal X is cut out into a time section by being multiplied by a correlation function by a cut-out means 2, a calculation means 3 calculates an autocorrelation coefficient from the time section, and an estimation means 4 estimates the noise level of the time section from difference of the maximum value of the autocorrelation coefficient and an autocorrelation coefficient of maximal value having a second size. When the newest cut-out time section is judged to be a noise section by a judging means 6, a calculation means 7 updates the average noise level. A calculation means 5 calculates the ratio of the noise level of the time section to the average noise level, and the ratio of the magnitude of noise varying from the noise section to the newest time section is output as the coefficient of time variation K. By multiplying this by the estimated value of the characteristic quantity of estimated noise, the noise reducing quantity is adjusted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention relates to a speech generated in a noisy environment or a speech to which electrical noise is added, and how much the noise added to the speech changes with time. The present invention relates to a method for calculating a time variation rate of a noise level indicating a noise level, a noise reduction method for extracting only voice from a voice signal on which a noise whose noise level changes with time is superimposed, and a noise reduction apparatus.

[0002]

2. Description of the Related Art A Boll reference (S.
F.Boll: "Suppression of acoustic noise in speech u
sing spectral subtraction ", IEEE Trans. On Acous
t., Speech, and Sig. Proc., vol.ASSP-27, no.2, pp.
113-120, 1979.4.) And a Wiener filter that has been used for a long time. In addition, various noise reduction methods such as comb filters and adaptive filters have been researched and developed, but the most representative ones are Lim's compiled reference book (JSLim
By "Speech enhancement", Prentice-Hall, Englewood
Cliffs, 1983.).

A spectrum subtraction method and a Wiener filter widely used in a noise reduction apparatus generally estimate a noise characteristic in a noise section which is a time section containing no speech, and subtract the noise component from an input signal. Noise reduction is performed as a basic principle (see FIGS. 15 and 16). However, in a voice section which is a time section including voice, since the estimated value of the noise spectrum is not generally updated, there is a drawback that it cannot cope with time-varying noise. Examples of means to solve this problem are described in the literature (Taniguchi, Tsumura, Fukudome: "Spectral subtraction method for adaptively estimating noise", IEICE Technical Report, SP94-116, pp.63-68. , 1995.3). In this method, it is configured not to determine whether the time section to be subjected to the noise reduction processing is a speech section or a noise section, and to estimate and update the noise spectrum in the speech section using an adaptive algorithm. I have.

However, the algorithm for estimating and updating noise in a voice section in this method is based on the assumption that a short-time voice such as a word is input, and cannot cope with a long voice section. On the other hand, the method described in the literature (Takagi, Yoshida, Watanabe: "Continuous Speech Recognition under Noise by Two-Stage Spectral Subtraction Method", Proceedings of the 1992 IEICE Spring Conference, A-238, 1992.) A dedicated microphone for inputting only noise is used in addition to the input microphone, and the noise spectrum can be estimated and updated sequentially even in the voice section. However, it is not always possible to create a situation in which two microphones are installed, and in order to obtain a noise reduction effect, it is necessary to consider a microphone installation method in which only noise is input.

[0005]

The noise reduction apparatus performs processing for estimating noise characteristics in a noise section which is a time section containing no voice, and reducing a noise component estimated from an input signal. However, since the noise component is an estimated value of the noise spectrum obtained in the noise section, the noise component cannot follow the time-varying noise.

SUMMARY OF THE INVENTION The present invention solves such a conventional problem. A method and apparatus for calculating a noise level temporal fluctuation rate and a noise reduction method and apparatus which are effective even when the magnitude of noise changes with time. The purpose is to provide.

[0007]

SUMMARY OF THE INVENTION In order to achieve the above object, the present invention provides a method of sequentially extracting an input signal as a time interval of a fixed time, and calculating a noise level in the time interval using an autocorrelation coefficient. Then, by calculating the average value (average noise level) and the ratio of the noise level in the time section for which the noise characteristics are estimated, the noise level is expressed as the time variation of the noise level, and the noise subtraction amount is calculated using the noise level variation rate. Is adjusted at any time.

In this case, the noise level is calculated by calculating the autocorrelation coefficient from the input signal and calculating the difference between the autocorrelation coefficient of the peak having the maximum value and the autocorrelation coefficient having the second maximum value peak. Is calculated as the noise level added to the input signal. The specific principle will be described.

First, assuming that an input signal is x (i), an autocorrelation coefficient R (j) is expressed by the following equation (1).

(Equation 1)

Here, j represents the delay time, N is a constant representing the range in which the product sum of the signals is calculated when obtaining the autocorrelation coefficient, and M is the maximum delay time (the range in which the autocorrelation coefficient is calculated).
Is a constant. This autocorrelation coefficient is obtained as the autocorrelation coefficient of the signal section cut out as a time section, but retains the property that the value of the autocorrelation coefficient does not decay with time among the properties of the autocorrelation coefficient of the infinite section To calculate. For example, the value of x (i) from the time of calculating the autocorrelation coefficient to the input signal after the time of (N + M) may be used.

The autocorrelation coefficient R (j) is a general feature that, when the input signal is a periodic signal, a local maximum peak appears at each cycle T, and is also used for detecting the cycle of speech. (See FIG. 13). Further, R (0) is a value indicating the average power of the input signal. If the signal is a periodic signal on which no noise is superimposed, its autocorrelation coefficient is Rs (0)
Then, a relationship of Rs (0) = Rs (T) (2) is established for the autocorrelation coefficient in the cycle T. This property is the property of the autocorrelation function that integrates in infinite time, but if it is the autocorrelation coefficient that does not attenuate with time shown above, it is established as an approximate expression. The audio signal is not a perfect periodic signal, but can be seen as a periodic signal when viewed locally, and is considered to satisfy this property.

Next, the case of a periodic signal on which random noise is superimposed will be considered by expressing the input signal as x (i) = s (i) + n (i) (3). Here, s (i) is a periodic signal,
n (i) represents a random noise component. The autocorrelation coefficient Rn (j) at this time is

(Equation 2) (An example of the autocorrelation coefficient is shown in FIG. 14). Here, if n (i) is a random signal, s
Since (i) and n (i) are uncorrelated,

(Equation 3) Is expected to be zero. further,

(Equation 4) Is also closer to 0 if j> 0. Therefore, the autocorrelation coefficient satisfying j> 0 is

(Equation 5) , And approaches the autocorrelation coefficient when noise is not superimposed.

The properties of these autocorrelation coefficients are described in the literature (Takasugi, Suzuki, Tanaka: “Functions and basic characteristics of speech processing system (SPAC) using autocorrelation function)”, IEICE Transactions (A) , No. 62, no., Pp. 175-182, 1979-3.), Used in sound reduction devices, and literatures (Kunieda, Shimamura,
Suzuki: "An Improved Method of LPC Analysis for Degraded Voice", IEICE Transactions on Electronics (A), No.J80-A, No.9,
pp. 1564-1566, 1997.09.), which has been used for speech analysis, and has been effective. If this property is used, Rn (T) becomes Rs of the audio signal on which noise is not superimposed.
Since it can be considered to approach (0), it can be seen that the audio power PS can be approximated by PS = Rs (0) = Rn (T) (8) further,

(Equation 6) It can be seen that the noise power PN can be approximated as PN = Rn (0) -Rn (T) (10) Power PN of this noise
As the noise level. Then, the time variation rate K of the noise level is calculated from the noise level NLx superimposed on the input signal calculated from the autocorrelation coefficient and the average noise level NLn in the noise section as follows: K = NLx / NLn (11) . Here, the average noise level NLn in the noise section is obtained by averaging the noise levels calculated in the time section (noise section) where it is determined that there is no voice.

Since the noise characteristic of the noise reduction device is generally obtained in a noise interval, the noise characteristic is multiplied by the noise level fluctuation rate from the noise interval to the current time to be processed, so that the time-varying noise is obtained. It can be estimated. By incorporating this algorithm into the noise reduction processing unit of the noise reduction device, effective noise reduction can be realized even for voices in which noise whose noise level changes with time is superimposed.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The invention according to claim 1 of the present invention is characterized in that a speech including noise or the like is used as an input signal, and the input signal is multiplied by a window function to cut it out into a signal of a time section, and the cut-out time is obtained. Calculate the autocorrelation coefficient of the signal of the section, estimate the noise level of the noise superimposed on the signal of the cut time section from the autocorrelation coefficient, and the cut time section is either a voice section or a noise section. Has a process of calculating the average noise level of the time section determined to be a noise section from the autocorrelation coefficient, and the ratio of the noise level of the extracted time section to the average noise level calculated in the noise section. The noise level temporal fluctuation rate calculation method calculates the magnitude of the noise that has changed in the time from the nearest noise section to the cut-out time section as the temporal fluctuation rate of the noise level by calculating , And the can follow effectively even when a change in the magnitude of the noise over time.

According to a second aspect of the present invention, when the noise level of the noise superimposed on the signal in the cut-out time section is estimated from the autocorrelation coefficient, the delay time of the autocorrelation coefficient is reduced. 2. The noise level according to claim 1, wherein the difference between the maximum value of the autocorrelation coefficient in the sequence and the autocorrelation coefficient of the local maximum having the second magnitude is calculated to estimate the noise level. This is a time variation calculation method, and can effectively follow the case where the magnitude of noise changes with time.

According to a third aspect of the present invention, in the first or second aspect, when the product-sum time of the calculated autocorrelation coefficient is N and the maximum delay time of the calculated autocorrelation coefficient is M, Is a method for calculating a noise level temporal fluctuation rate in which a window function for extracting an input signal as a signal in a time section is a square wave having a time length (M + N) or more, and effectively follows even when the magnitude of noise changes with time. be able to.

According to a fourth aspect of the present invention, there is provided a means for extracting a speech including a noise or the like as an input signal, and multiplying the input signal by a window function to cut out the signal into a time interval signal; Means for calculating the autocorrelation coefficient of the signal, means for estimating the noise level of noise superimposed on the signal in the cut time section from the autocorrelation coefficient, and means for selecting whether the cut time section is a speech section or a noise section. And means for calculating the average noise level of the time section determined to be a noise section from the autocorrelation coefficient, and the noise level of the cut time section and the average calculated in the noise section. By calculating the ratio of the noise level, the noise is calculated as the temporal fluctuation rate of the noise level, the magnitude of the noise that has changed in the time from the nearest noise section to the cut-out time section. A bell time variation rate computing unit,
Even when the magnitude of the noise changes with time, it can effectively follow.

According to a fifth aspect of the present invention, the means for estimating the noise level of the noise superimposed on the signal in the cut-out time section from the autocorrelation coefficient includes: The maximum value of the autocorrelation coefficient in the series and the second
The noise level temporal fluctuation rate calculation apparatus according to claim 4, wherein a noise level is estimated by calculating a difference between autocorrelation coefficients of local maxima having a magnitude of: Can be effectively followed even when the value changes.

The invention according to claim 6 of the present invention is characterized in that, in claim 4 or 5, when the product-sum time of the calculated autocorrelation coefficient is N and the maximum delay time of the calculated autocorrelation coefficient is M A noise level temporal fluctuation rate calculating device that converts a window function for extracting an input signal as a signal of a time section into a square wave having a time length (M + N) or more, and effectively follows even when the magnitude of noise changes with time. be able to.

According to a seventh aspect of the present invention, there is provided a method for reducing a noise added to an input signal by using a speech including noise or the like as an input signal. This is a noise reduction method in which the time variation rate of the noise level is calculated and the obtained time variation rate of the noise level is subtracted from the input signal to adjust the subtraction amount of the noise. Effective noise reduction for voice can be realized.

According to an eighth aspect of the present invention, a speech including noise or the like is used as an input signal, and the input signal is multiplied by a window function to cut it out into a signal of a time section. Or, it is determined whether it is a noise section, the signal of the cut time section is converted into an acoustic parameter, the acoustic parameter of the noise component is estimated from the acoustic parameter of the signal of the time section determined as the noise section, and the latest 4. A noise reduction method for calculating an acoustic parameter in which a noise component is reduced from an acoustic parameter of a time section cut out at a time and an acoustic parameter of a noise component estimated in a noise section, according to any one of claims 1 to 3. The time variation of the noise level of the input signal is calculated by the method described above, and the acoustic parameter of the estimated noise component is multiplied by the time variation of the noise level. This is a noise reduction method in which the calculated result is subtracted from the acoustic parameter in the latest time section as the acoustic parameter of the noise component to obtain an acoustic parameter obtained by performing noise reduction processing on the input signal. Effective noise reduction for voice can be realized.

According to a ninth aspect of the present invention, in the noise reduction method according to the eighth aspect, there is provided a noise reduction method using an acoustic parameter as a frequency spectrum. , Effective noise reduction can be realized.

According to a tenth aspect of the present invention, in the noise reduction method according to the eighth aspect, a noise reduction method using acoustic parameters as linear prediction coefficients, wherein noise whose noise level changes with time is superimposed. Effective noise reduction for voice can be realized.

According to an eleventh aspect of the present invention, a speech including noise or the like is used as an input signal, and the input signal is multiplied by a window function to cut out a signal in a time section. Or, it is determined whether it is a noise section, the signal of the cut time section is converted into an acoustic parameter, the acoustic parameter of the noise component is estimated from the acoustic parameter of the signal of the time section determined as the noise section, and the latest Subtracts the acoustic parameter of the noise component estimated in the noise section from the acoustic parameter of the time section extracted at the time of, converts the acoustic parameter obtained by subtracting the acoustic parameter of the noise component into a time signal, and connects the converted time waveform And a method for calculating a time variation rate of a noise level of an input signal by the method according to any one of claims 1 to 3. After subtracting the result of multiplying the estimated noise parameter acoustic parameter by the time variation rate of the noise level as the noise component acoustic parameter from the acoustic parameter of the latest time section, converting the subtracted acoustic parameter into a time signal This is a noise reduction method characterized in that the noise is output in such a manner that the effective noise reduction can be realized even for a voice on which noise whose noise level changes with time is superimposed.

According to a twelfth aspect of the present invention, in the noise reduction method according to the eleventh aspect, a noise reduction method using an acoustic parameter as a frequency spectrum is provided. , Effective noise reduction can be realized.

According to a thirteenth aspect of the present invention, there is provided the noise reduction method according to the eleventh aspect, wherein the acoustic parameter is a linear prediction coefficient, wherein noise whose time varies with time is superimposed. Effective noise reduction for voice can be realized.

According to a fourteenth aspect of the present invention, a speech including noise or the like is used as an input signal, and the input signal is multiplied by a window function to cut out a signal in a time section. Or it is determined whether it is a noise section, the signal of the cut time section is converted into a frequency spectrum, and the noise spectrum is estimated from the frequency spectrum of the signal of the time section determined as the noise section,
Subtract the noise spectrum estimated in the noise section from the cut-out frequency spectrum of the latest time section to design the frequency characteristics of the filter to realize noise reduction, calculate the filter coefficient from the filter frequency characteristics and configure the filter A noise reduction method for realizing noise reduction of an input signal by passing the input signal through a filter having the filter coefficient, wherein the noise level of the input signal is reduced by the method according to any one of claims 1 to 3. Calculating the time variation rate, determining the frequency characteristic of the filter by subtracting the result of multiplying the noise spectrum by the time variation rate of the noise level as an estimated value of the noise spectrum from the frequency spectrum of the signal in the latest time section. A noise reduction method characterized by the following: Effective noise reduction can be achieved.

According to a fifteenth aspect of the present invention, there is provided an apparatus for reducing a noise added to an input signal by using speech including noise or the like as an input signal. The input signal is passed to a noise reduction processing unit and a noise level time variation rate calculation unit, and the noise level time variation rate calculation unit calculates the noise level with the device according to any one of claims 4 to 6. A noise reduction device which calculates a time variation rate, passes the obtained time variation rate of the noise level to a noise reduction processing unit, and adjusts a noise subtraction amount by means for subtracting a noise component from an input signal. In addition, effective noise reduction can be realized even for a voice on which noise whose noise level changes with time is superimposed.

The invention according to claim 16 of the present invention is characterized in that a voice including noise or the like is used as an input signal, and the input signal is multiplied by a window function to cut out a signal in a time section. Means for determining whether the signal is a voice section or a noise section, means for converting a signal of the cut-out time section into acoustic parameters, and sound of a noise component from the acoustic parameters of the signal of the time section determined as the noise section. A noise reduction apparatus comprising: means for estimating a parameter; and means for subtracting a noise component from an acoustic parameter in a time section cut out at the latest time and an acoustic parameter of a noise component estimated in a noise section, wherein 6. The sound parameter of the time section cut out at the latest time based on the time variation rate of the noise level of the input signal calculated by the apparatus described in any one of 6. From the data to the means for subtracting the acoustic parameter estimated in the noise section, and multiplying the estimated acoustic parameter of the noise component by the time variation rate of the noise level as the acoustic parameter of the noise component from the acoustic parameter of the latest time interval. This is a noise reduction apparatus characterized by obtaining an acoustic parameter obtained by performing a noise reduction process on an input signal by subtraction, and can realize effective noise reduction even for a voice on which noise whose noise level changes with time is superimposed.

According to a seventeenth aspect of the present invention, there is provided the noise reducing apparatus according to the sixteenth aspect, wherein the noise parameter has a frequency spectrum as an acoustic parameter. , Effective noise reduction can be realized.

[0032] The invention according to claim 18 of the present invention is the noise reduction device according to claim 16, wherein the noise parameter is a linear prediction coefficient as an acoustic parameter. Effective noise reduction for voice can be realized.

[0033] According to a nineteenth aspect of the present invention, there is provided means for taking a speech including noise or the like as an input signal, and multiplying the input signal by a window function to cut out the signal into a time section signal; Means for determining whether the signal is a voice section or a noise section, means for converting a signal of the cut-out time section into acoustic parameters, and sound of a noise component from the acoustic parameters of the signal of the time section determined as the noise section. Means for estimating the parameter, means for subtracting the acoustic parameter of the noise component estimated in the noise section from the acoustic parameter of the time section cut out at the latest time,
A noise reduction device comprising: means for converting an acoustic parameter obtained by subtracting an acoustic parameter of a noise component into a time signal; and means for connecting and outputting the converted time waveform,
Passing the time variation rate of the noise level of the input signal calculated by the apparatus according to claim 4 to means for subtracting the acoustic parameter estimated in the noise section from the acoustic parameter in the time section cut out at the latest time; The result of multiplying the estimated noise component acoustic parameter by the time variation rate of the noise level is subtracted from the acoustic parameter of the latest time section as the noise component acoustic parameter, and then the subtracted acoustic parameter is converted into a time signal and output. The noise reduction device is characterized in that the noise reduction device can perform effective noise reduction even for a voice on which noise whose noise level changes with time is superimposed.

According to a twentieth aspect of the present invention, there is provided the noise reduction apparatus according to the nineteenth aspect, wherein the noise parameter has a frequency spectrum as an acoustic parameter. , Effective noise reduction can be realized.

According to a twenty-first aspect of the present invention, in the noise reducing apparatus according to the nineteenth aspect, the noise reducing apparatus uses the acoustic parameter as a linear prediction coefficient. Effective noise reduction for voice can be realized.

According to a twenty-second aspect of the present invention, there is provided a means for extracting a speech including a noise or the like as an input signal and multiplying the input signal by a window function to cut out the signal into a time interval signal; A means for determining whether the signal is a voice section or a noise section, a means for converting a cut-out signal in a time section into a frequency spectrum, and an estimation of a noise spectrum from the frequency spectrum of the signal in the time section determined to be a noise section Means for designing a filter frequency characteristic for realizing noise reduction by subtracting the noise spectrum estimated in the noise section from the cut-out frequency spectrum of the latest time section, and a filter coefficient from the filter frequency characteristic. Means for calculating and configuring a filter, and passing an input signal through a filter having the calculated filter counts Therefore, the present invention is a noise reduction device that realizes noise reduction of an input signal, and calculates a time variation rate of a noise level of an input signal obtained by the device according to any one of claims 4 to 6 to a signal of a signal in a time section cut out. The noise spectrum estimated from the frequency spectrum is subtracted and passed to means for designing a filter, and the result of multiplying the noise spectrum by the time variation rate of the noise level is used as an estimated value of the noise spectrum from the frequency spectrum of the signal in the latest time section. This is a noise reduction device characterized by determining a frequency characteristic of a filter by subtraction, and can realize effective noise reduction even for a voice on which noise whose noise level changes with time is superimposed.

(Embodiment 1) Claims 1 to 6 of the present invention
A configuration example of a method and an apparatus for calculating a time variation rate of a noise level described in (1) will be described as a first embodiment with reference to FIG. The speech including noise and the like is input as an input signal X, and is cut out as a signal in a time section by multiplying a window function by a time section cut-out means 2, and then the autocorrelation coefficient of the cut out signal in the time section is calculated. It is passed to the calculating means 3 and the means 6 for judging whether the extracted time section is a voice section or a noise section.

In the means 3 for calculating the autocorrelation coefficient, the autocorrelation coefficient R (j) is calculated by the above equation (1) using the input signal as x (i). However, the autocorrelation coefficient R (j) calculated by the autocorrelation coefficient calculation means 3 is calculated in a time section extracted by multiplying the input signal by a window function, and the window function has a time length ( (M + N) square wave. This means that of the data following the infinite length, only the signal of the time length necessary for calculating the autocorrelation coefficient is cut out and calculated, and the calculated autocorrelation coefficient value attenuates with time. No window function is used. The value of N is set to about 20 to 40 ms, and the value of M is set to about 15 to 20 ms.

The autocorrelation coefficient calculated by the means 3 for calculating the autocorrelation coefficient in the extracted time interval is passed to the noise level estimating means 4. The calculation of the noise level in the noise level estimating means 4 is performed by the above equation (10). The noise level in the latest time section calculated by the noise level estimating means 4 is passed to a means 5 for calculating a time variation rate of the noise level. At the same time, if the result of the means 6 for determining whether the cut-out time section is a voice section or a noise section is a noise section, the noise level in the latest time section is calculated and the average noise level in the noise section is calculated. And averages the average noise level by averaging.

The noise level NLx in the latest time section calculated by the means 4 for estimating the noise level superimposed on the input signal from the autocorrelation coefficient and the noise level calculated in the means 7 for calculating the average noise level in the noise section. Average noise level NL
n is passed to the means 5 for calculating the time variation rate of the noise level, and calculates and outputs the time variation rate K of the noise level changed from the noise section to the latest cut-out time section by the formula (11).

(Embodiment 2) Claims 7 and 1 of the present invention
FIG. 2 shows a basic configuration example of the noise reduction method and apparatus described in FIG. The input signal X on which noise or the like is superimposed is passed to a noise reduction processing unit 41 and a noise level time variation rate calculation unit 42, and the method and apparatus described in the first embodiment, or any one of claims 1 to 6. The time variation rate K of the noise level by the method or apparatus described in
Is calculated. Time variation rate K of the calculated noise level
Is passed to the noise reduction processing unit 41 and is used to adjust the amount of noise reduction by multiplying the acoustic parameter of the noise predicted when performing the noise reduction processing.

(Embodiment 3) Next, claim 8 of the present invention will be described.
Embodiments 3 to 10 and 16 to 18 will be described with reference to FIG. 3 as an embodiment for obtaining noise-reduced acoustic parameters from an input signal X on which noise is superimposed. An audio signal on which noise or the like is superimposed is received as an input signal X, and is passed to an acoustic parameter processing unit 400 and a noise level time variation rate estimation unit 100, respectively. The acoustic parameter processing unit 400 first extracts the latest time section by multiplying the input signal X by a window function, and calculates acoustic parameters. At the same time, the extracted time section is passed to the means 6 for determining whether the section is a voice section or a noise section. If the result is determined to be a noise section, the acoustic parameter is transmitted to the acoustic parameter estimating means 32 for the noise component. The acoustic parameters of the noise components are obtained by passing and averaging, and are passed to the acoustic parameter subtracting means 33.

On the other hand, the noise level temporal fluctuation rate estimating unit 10
0, the time variation rate K of the noise level is determined by the method and apparatus described in the first embodiment in parallel with the above processing.
Ask for. First, the input signal X is multiplied by a window function and cut out as the latest time section, and the autocorrelation coefficient R (j) is calculated. Then, the noise level estimating means 4 calculates R (0) and R
If the difference of (T) is estimated as the noise level NLs from the above equation (10) and the result of the means 6 for judging whether the extracted time section is a speech section or a noise section is a noise section, The noise level NLs is passed to means 7 for calculating an average noise level in a noise section, and is averaged to obtain an average noise level NLn. The means 5 for calculating the time variation rate of the noise level calculates the ratio between the noise average level NLn at the time of calculating the acoustic parameter of the noise component and the noise level NLs in the latest time section by the above equation (11). The variation rate K is passed to the acoustic parameter subtraction means 13. The acoustic parameter subtraction means 13 subtracts the acoustic parameter from the acoustic parameter of the latest time section, the acoustic parameter of the noise component, and the value of the noise level temporal variation rate K, and outputs the acoustic parameter P after the subtraction.

(Embodiment 4) Next, claim 1 of the present invention will be described.
An example for obtaining a noise-reduced time signal in the noise reduction methods and apparatuses described in 1 to 13 and 19 to 21 will be described as a fourth embodiment with reference to FIG. here,
An example of a method using a frequency spectrum as an acoustic parameter used in the noise reduction processing (the method and apparatus according to claims 12 and 20) is a combination of a spectrum subtraction method and a means for calculating a noise level temporal variation rate. FIG.

An audio signal on which noise or the like is superimposed is input signal X
As the spectrum subtraction processing unit 2
00 and the noise level temporal fluctuation rate estimating unit 100. The spectrum subtraction processing unit 200 first cuts out the latest time section by multiplying the input signal X by a window function.
Calculate (ω). At the same time, the extracted time section is passed to means 6 for determining whether the section is a voice section or a noise section. If the result is determined to be a noise section, the frequency spectrum X (ω) is transmitted to the noise estimating section 16. To obtain the noise spectrum N (ω) by averaging
It is passed to the spectrum subtraction means 14.

On the other hand, the noise level temporal fluctuation rate estimating unit 10
0, the input signal X is multiplied by a window function in parallel with the above processing to extract the latest time section, and the autocorrelation coefficient R
Calculate (j). Then, the noise level estimating means 4 estimates the difference between R (0) and R (T) as the noise level NLs from the equation (10), and determines whether the extracted time section is a speech section or a noise section. Judgment means 6
Is the noise interval, this noise level NLs
Is passed to means 7 for calculating an average noise level in a noise section, and is averaged to obtain an average noise level NLn.

Means 5 for calculating time variation of noise level
Then, the ratio between the noise average level NLn at the time of noise spectrum calculation and the noise level NLs in the latest time section is calculated by the above equation (11), and is passed to the spectrum subtracting means 13 as the noise level time variation rate K. In the spectrum subtracting means 13, the frequency spectrum X in the latest time section is
From the value of (ω), the estimated value N (ω) of the noise spectrum, and the value of the noise level temporal variation rate K, S (ω) = X (ω) −KN
(Ω) According to (12), the frequency spectrum is subtracted, and the frequency spectrum S (ω) after the subtraction is converted into a time domain signal by the time signal converting means 14. Repeat the above process,
An output time signal W is obtained by connecting the obtained time signal waveforms.

(Embodiment 5) Next, Embodiment 1 of the present invention will be described.
Embodiment 5 is an example of noise reduction processing in the noise reduction method and apparatus described in Embodiments 4 and 22, and FIG. 5 shows an example of processing in which a Wiener filter and a noise level time variation rate are combined.
It will be explained by. The input signal X is passed to the Wiener filter processing unit 300 and the noise level time variation rate estimation unit 100. In the Wiener filter processing unit 300, first, the input signal X is multiplied by a window function and cut out to obtain the latest time section, which is passed to the frequency spectrum conversion means 12. At the same time, the extracted time section is passed to means 6 for determining whether the time section is a voice section or a noise section,
If it is determined that the noise section is a noise section, the spectrum determined by the frequency spectrum converting means 12 is determined as a noise spectrum, passed to the noise spectrum estimating section 16 and averaged, and the noise spectrum N (ω ) Update the estimate. Estimated value N of the calculated noise spectrum
(Ω) is passed to the filter frequency characteristic design means 21.

On the other hand, the noise level fluctuation estimating section 100
After the input signal 1 is cut out as a signal in a time section by multiplying the input signal 1 by a square wave as a window function, the autocorrelation coefficient calculation means 3 calculates the autocorrelation coefficient R (j). Next, R (0) and R
The difference of (T) is calculated by the above equation (10), and is estimated as the noise level NLs. Means 11 for determining whether the extracted time section is a voice section or a noise section
Is determined to be a noise section, the calculated noise level NLs is passed to the noise level calculation means 7 in the noise section, and the noise level NLs is averaged.
Ask for. The ratio between the noise level NLs in the latest time section and the noise level NLn in the noise section is calculated by the above equation (11).
Is calculated as the noise level time variation rate K. The obtained noise level time variation rate K is passed to the filter frequency characteristic design means 21.

Thus, means 1 for converting to a frequency spectrum
2, the spectrum X (ω) of the latest time section and the noise spectrum N calculated by the noise spectrum estimation unit 16
Based on (ω) and the noise level temporal variation rate K calculated by the noise level temporal variation rate estimating section 100, the filter frequency characteristic designing section 21 shows a filter frequency characteristic for passing only an audio signal from an input signal as H ( ω) = {X (ω) −KN (ω)} / N (ω) (13) The filter coefficient of the Wiener filter is obtained by the means 22 for calculating the filter coefficient from H (ω), and is passed to the Wiener filtering means 23. The input signal X is transmitted to the Wiener filtering unit 23.
And an output time signal W is obtained.

(Embodiment 6) FIG. 6 shows a specific configuration example of a noise reduction device according to a fifteenth embodiment of the present invention as a sixth embodiment. The audio signal 1 on which noise or the like is superimposed is input from a microphone 51 and is converted into a digital signal X by an AD converter 52, and then a noise reduction processing unit 53
Passed to. At the same time, the digital signal X is passed to the noise level time variation calculation unit 54, where the noise and the noise are reduced by the method and apparatus described in the first embodiment or the method or apparatus described in any one of claims 1 to 6. The time variation rate K of the level is calculated. The calculated time variation rate K of the noise level is passed to the noise reduction processing unit 53, and is used to adjust the noise reduction amount by multiplying the acoustic parameter of the noise predicted when performing the noise reduction processing. .
The signal processed by the noise reduction processing unit 53 is the noise reduction signal Y
Is output as Y is output as an acoustic parameter P or a time waveform W according to the purpose.

Next, the noise level time fluctuation rate calculation unit 54
Will be described with reference to FIG. In the noise level time variation rate calculating section 54, first, the AD converted signal X
Is cut out as a time section signal by an input signal cutout section 71, and an autocorrelation coefficient calculation section 72 and a speech section /
This is passed to the noise section determination unit 74. Autocorrelation coefficient calculator 7
The autocorrelation coefficient obtained in step 2 is calculated by the noise level calculator 7.
3 and used to calculate the noise level according to equation (10). The voice section / noise section determination section 74 determines whether the cut-out time section is a voice section or a noise section, and if the time section is a noise section, connects the switch 75 to the terminal N to calculate the noise level. The noise level calculated by the unit 73 is passed to the average noise level updating unit 76, and if it is a voice section, the changeover switch 75
To the terminal S so that the noise level is not passed to the average noise level updating unit 76. Average noise level updating unit 76
Then, the average noise level is updated by averaging with the average noise level 77 in the past noise section, and the updated average noise level 7 is updated.
7 is passed to the noise level time variation calculation unit 78. The noise level temporal fluctuation rate calculating section 78 obtains the noise level temporal fluctuation rate K from the noise level calculated in the time section and the average noise level, and passes the noise level temporal fluctuation rate K to the noise reduction processing section 53.

Next, an example of processing in the noise reduction processing section 53 will be described with reference to FIG. First, the AD converted signal X is cut out as a time section signal by the input signal cutout unit 61. The extracted time section signal is passed to the acoustic characteristic calculation section 62 and the speech section / noise section determination section 63. The acoustic characteristic calculation unit 62 converts the cut-out time signal into acoustic parameters such as a frequency spectrum by FFT, for example, and passes the converted signal to the noise reduction processing operation unit 66. At the same time, the voice section / noise section determination section 63 determines whether the cut-out time section is a voice section or a noise section, and if the time section is a noise section, connects the changeover switch 64 to the terminal N. The acoustic parameters calculated by the acoustic characteristic calculation unit 62 are passed to the noise characteristic update unit 65,
If it is determined that the voice section is the voice section, the changeover switch 64 is connected to the terminal S, and the acoustic parameter is updated to the noise characteristic updating section 65.
Do not pass to. The noise characteristic update unit 65 estimates the acoustic parameters at the time of superimposing noise by averaging the acoustic parameters in the past noise section, and uses the result as the noise reduction processing unit 6
Pass to 6. The noise reduction processing unit 66 combines the acoustic parameter calculated in the time section, the acoustic parameter calculated in the noise section, and the time variation rate K of the noise level calculated by the time variation rate calculation unit 54 of the noise level, Estimate acoustic parameters without noise. The acoustic parameter without noise is converted into a time section signal by the time signal conversion unit 67, and then connected to the past time section signal and output.

(Embodiment 7) FIG. 9 shows a specific configuration example of a noise reduction device for obtaining an acoustic parameter from an input signal according to the sixteenth to eighteenth aspects of the present invention as a seventh embodiment. The audio signal on which noise or the like is superimposed
After being input from 1 and converted into a digital signal X by the AD converter 52, it is passed to the noise reduction processing unit 53.
At the same time, the digital signal X is passed to the noise level time variation calculation unit 54, where the noise and the noise are reduced by the method and apparatus described in the first embodiment or the method or apparatus described in any one of claims 1 to 6. The time variation rate K of the level is calculated. The calculated time variation rate K of the noise level is passed to the noise reduction processing unit 53, and is used to adjust the noise reduction amount by multiplying the acoustic parameter of the noise predicted when performing the noise reduction processing. . The parameter P of the signal processed by the noise reduction processing unit 53 is passed to, for example, the parameter display unit 60 and used.

An example of the noise reduction processing section 53 in the present apparatus will be described with reference to FIG. First, the signal X subjected to AD conversion is cut out as a time section signal by the input signal cutout unit 61. The extracted time section signal is passed to the acoustic characteristic calculation section 62 and the speech section / noise section determination section 63. The acoustic characteristic calculation unit 62 converts the cut-out time signal into acoustic parameters such as a frequency spectrum by FFT, for example, and passes the converted signal to the noise reduction processing operation unit 66. At the same time, the voice section / noise section determination section 63 determines whether the cut-out time section is a voice section or a noise section, and if the time section is a noise section, connects the changeover switch 64 to the terminal N. Then, the acoustic parameter calculated by the acoustic characteristic calculation unit 62 is passed to the noise characteristic updating unit 65, and if it is determined that the voice section is selected, the changeover switch 64 is connected to the terminal S and the acoustic parameter is not passed to the noise characteristic updating unit 65. To do. The noise characteristic updating unit 65 estimates the acoustic parameters at the time of superimposing noise by averaging the acoustic parameters in the past noise section, and uses the result as a noise reduction processing unit 66
Pass to. The noise reduction processing unit 66 combines the acoustic parameter calculated in the time section and the acoustic parameter calculated in the noise section with the time variation rate K of the noise level calculated by the time variation rate calculation unit 54 of the noise level, An acoustic parameter P on which noise is not superimposed is estimated and output.

(Eighth Embodiment) FIG. 11 shows a specific configuration example of a noise reduction device for outputting a signal obtained by reducing noise from an input signal according to the present invention as an eighth embodiment. The audio signal on which noise or the like is superimposed is input from a microphone 51, converted into a digital signal X by an AD converter 52, and passed to a noise reduction processing unit 53. At the same time, the digital signal X is passed to the noise level time variation calculation unit 54, where the noise and the noise are reduced by the method and apparatus described in the first embodiment or the method or apparatus described in any one of claims 1 to 6. The time variation rate K of the level is calculated. Time variation rate K of the calculated noise level
Is passed to the noise reduction processing unit 53 and is used to adjust the amount of noise reduction by multiplying the acoustic parameter of the noise predicted when performing the noise reduction processing. The signal Y processed by the noise reduction processing unit 53 is output from a speaker after being converted into an analog signal by a DA converter 55.

As an example of the noise reduction processing section 53 in the present apparatus, an example of a method using Wiener filter processing will be described with reference to FIG. First, the signal X subjected to AD conversion is cut out as a time section signal by the input signal cutout unit 61. The extracted time section signal is passed to the acoustic characteristic calculation section 62 and the speech section / noise section determination section 63. The acoustic characteristic calculation unit 62 converts the cut-out time signal into a frequency spectrum by, for example, FFT, and passes the frequency spectrum to the noise reduction processing operation unit 66. At the same time, the voice section / noise section determination section 63 determines whether the cut-out time section is a voice section or a noise section, and if the time section is a noise section, connects the changeover switch 64 to the terminal N. Then, the frequency spectrum calculated by the acoustic characteristic calculating unit 62 is passed to the noise characteristic updating unit 65. If it is determined that the voice section is a voice section, the changeover switch 64 is connected to the terminal S and the frequency spectrum is not passed to the noise characteristic updating unit 65. To do. The noise characteristic updating unit 65 estimates the frequency spectrum at the time of noise superposition by averaging the frequency spectrum of the past noise section, and passes the result to the noise reduction processing unit 66. The noise reduction processing unit 66 combines the frequency spectrum calculated in the time section and the frequency spectrum calculated in the noise section with the time variation rate K of the noise level calculated by the time variation rate calculation unit 54 of the noise level, The frequency spectrum without noise is estimated. The filter characteristic is determined based on the frequency spectrum in which the noise is not superimposed, and the time signal X is converted into a noise-reduced time signal W by the filtering processing unit 68 and output.

[0058]

According to the present invention, as is apparent from the above embodiment, it is possible to calculate an effective time variation rate of a noise level even when the magnitude of noise changes with time.
It is possible to realize a noise reduction device that is effective even for a voice on which noise whose level is changed is superimposed. This makes it possible to reduce the noise of audio signals with noise superimposed on acoustic signals recorded on a recording medium such as a tape recorder to make it easier to hear, and to use a speech recognition device that receives speech uttered in a noisy environment as input. , As a pre-process, it is possible to more effectively prevent the recognition rate from deteriorating due to noise.

[Brief description of the drawings]

FIG. 1 is a block diagram of a method and an apparatus for calculating a time variation rate of a noise level using an autocorrelation coefficient according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a basic configuration of a noise reduction method and device according to a second embodiment of the present invention.

FIG. 3 is a block diagram showing a basic configuration of a noise reduction method and device according to a third embodiment of the present invention.

FIG. 4 is a block diagram illustrating a basic configuration of a noise reduction method and apparatus according to a fourth embodiment of the present invention.

FIG. 5 is a block diagram showing a basic configuration of a noise reduction method and device according to a fifth embodiment of the present invention.

FIG. 6 is a block diagram illustrating an example of a noise reduction device according to a sixth embodiment of the present invention.

FIG. 7 is a block diagram showing a configuration of a noise level time variation rate calculation unit in the noise reduction apparatus according to the sixth embodiment.

FIG. 8 is a block diagram illustrating a configuration of a noise reduction processing unit in a noise reduction device according to a sixth embodiment.

FIG. 9 is a block diagram illustrating an example of a noise reduction device according to a seventh embodiment of the present invention.

FIG. 10 is a block diagram illustrating a configuration of a noise reduction processing unit in a noise reduction device according to a seventh embodiment.

FIG. 11 is a block diagram illustrating an example of a noise reduction device according to an eighth embodiment of the present invention.

FIG. 12 is a block diagram illustrating a configuration of a noise reduction processing unit in a noise reduction device according to an eighth embodiment.

FIG. 13 is a characteristic diagram showing an example of an audio signal and an autocorrelation coefficient;

FIG. 14 is a characteristic diagram showing an example of a speech signal on which noise is superimposed and an autocorrelation coefficient.

FIG. 15 is a block diagram of a conventional noise reduction method using a spectral subtraction method.

FIG. 16 is a basic block diagram of a conventional noise reduction method and device using a Wiener filter method.

[Explanation of symbols]

Reference Signs List 1 input signal 2, 11 time section extraction means 3 autocorrelation coefficient calculation means 4 noise level estimation means 5 noise level time variation calculation means 6 voice section / noise section determination means 7 average noise level estimation Means 12 Means for converting time signal to frequency spectrum 13 Means for subtracting spectrum in frequency spectrum domain 14 Means for converting frequency spectrum to time signal 15 Time signal waveform connection means 16 Noise spectrum estimating means 20 Output signal 21 Filter frequency characteristics Design means 22 filter coefficient calculation means 23 Wiener filtering means 31 acoustic parameter calculation means 32 noise acoustic parameter estimation means 33 noise parameter subtraction section 41, 53 noise reduction processing section 42, 54 noise level time variation calculation section 51 Input microphone 52 A / D converter 55 D / A converter 56 Output speaker 60 Parameter display section 61, 71 Input signal cutout section 62 Acoustic characteristic calculation section 63, 74 Voice section / noise section determination section 64, 75 Changeover switch 65 Noise characteristic update section 66 Noise reduction processing operation section 67 Time signal conversion section 68 Time signal connection section 69 Filter characteristic determination section 70 Filtering processing section 72 Autocorrelation coefficient calculation section 73 Noise level calculation section 76 Average noise level update section 77 Average noise Level 78 Calculating section for noise level temporal fluctuation rate 100 Calculating section for noise level temporal fluctuation rate 200 Spectrum subtraction processing section 300 Wiener filter processing section 400 Acoustic parameter processing section X AD converted signal Y Output signal after noise reduction processing K Time variation rate of noise level P Miscellaneous Reduction after the acoustic parameters W noise reduction after the time waveform

Claims (22)

[Claims] An audio signal containing noise or the like is used as an input signal.
The input signal is multiplied by a window function to cut it out into a signal in a time section, the autocorrelation coefficient of the signal in the cut out time section is calculated, and the noise level of the noise superimposed on the signal in the cut out time section is calculated by the self phase Estimated from the number of relations, it is determined whether the extracted time interval is a voice interval or a noise interval, and the average noise level of the time interval determined to be a noise interval is calculated from the autocorrelation coefficient. And
By calculating the ratio of the noise level in the extracted time interval to the average noise level calculated in the noise interval, the magnitude of the noise that has changed in the time from the nearest noise interval to the extracted time interval is represented by the time variation of the noise level. Noise level time variation rate calculation method to calculate as a rate. 2. A method for estimating a noise level of noise superimposed on a signal in the cut-out time section from an autocorrelation coefficient, comprising: determining a maximum value of an autocorrelation coefficient in a delay time sequence of the autocorrelation coefficient; 2. The method according to claim 1, wherein the noise level is estimated by calculating the difference between the autocorrelation coefficients of the local maximum having the second magnitude. 3. An input signal is cut out as a signal in a time section according to claim 1 or 2, wherein N is the product-sum time of the calculated autocorrelation coefficient and M is the maximum delay time of the calculated autocorrelation coefficient. A noise level temporal fluctuation rate calculation method in which a window function is a square wave having a time length (M + N) or more. 4. A speech including noise or the like is used as an input signal.
Means for cutting out the signal in the time section by multiplying the input signal by a window function, means for calculating the autocorrelation coefficient of the signal in the cut time section, and noise of the noise superimposed on the signal in the cut time section Means for estimating the level from the autocorrelation coefficient, means for determining whether the extracted time section is a speech section or a noise section, and means for calculating the average noise level in the time section determined to be a noise section by the autocorrelation. It has a means to calculate from the relation number, and by calculating the ratio of the noise level of the cut-out time section to the average noise level calculated in the noise section, the time changed from the closest noise section to the cut-out time section A noise level temporal fluctuation rate calculation device that calculates a magnitude of noise as a temporal fluctuation rate of the noise level. 5. A means for estimating a noise level of noise superimposed on a signal in a cut-out time section from an autocorrelation coefficient, comprising: a maximum value of an autocorrelation coefficient in a delay time sequence of the autocorrelation coefficient; 5. The noise level temporal fluctuation rate calculation apparatus according to claim 4, wherein the noise level is estimated by calculating the difference between the autocorrelation coefficients of the local maximum values having the second magnitude. 6. The input signal is cut out as a signal in a time section according to claim 4 or 5, wherein the product-sum time of the calculated autocorrelation coefficient is N and the maximum delay time of the calculated autocorrelation coefficient is M. A noise level temporal fluctuation rate calculating device that uses a window function as a square wave having a time length (M + N) or more. 7. An audio signal including noise or the like is used as an input signal.
4. A method for reducing noise added to an input signal, comprising: calculating a time variation rate of a noise level by the method according to claim 1; and subtracting the determined time variation rate of the noise level from the input signal. A noise reduction method that adjusts the amount of noise subtraction by using 8. A speech including noise or the like is used as an input signal.
The input signal is multiplied by a window function to cut out a signal in a time section, it is determined whether the cut-out time section is a voice section or a noise section, and the cut-out time section signal is converted into an acoustic parameter. Estimate the acoustic component of the noise component from the acoustic parameter of the signal in the time interval determined as the noise interval, and reduce the noise component from the acoustic parameter of the time interval cut out at the latest time and the acoustic parameter of the noise component estimated in the noise interval. A noise reduction method for calculating an estimated acoustic parameter, comprising calculating a time variation rate of a noise level of an input signal by the method according to any one of claims 1 to 3, and adding the noise to an estimated acoustic parameter of a noise component. The result of multiplying the time variation of the level as the acoustic parameter of the noise component from the acoustic parameter of the latest time section And it can calculate the noise reduction method for obtaining acoustic parameters noise reduction processing of the input signal. 9. The noise reduction method according to claim 8, wherein the acoustic parameter is a frequency spectrum. 10. The noise reduction method according to claim 8, wherein the acoustic parameter is a linear prediction coefficient. 11. A speech including noise or the like is used as an input signal, and the input signal is multiplied by a window function to cut out a signal in a time section. Whether the cut out time section is a speech section or a noise section is determined. Judgment, the signal of the extracted time interval is converted into an acoustic parameter, the acoustic parameter of the noise component is estimated from the acoustic parameter of the signal of the time interval determined as the noise interval, and the audio of the time interval extracted at the latest time A noise reduction method comprising: subtracting an acoustic parameter of a noise component estimated in a noise section from a parameter, converting an acoustic parameter obtained by subtracting an acoustic parameter of the noise component into a time signal, and connecting and outputting the converted time waveform. The time variation of the noise level of the input signal is calculated by the method according to any one of claims 1 to 3, and the acoustic parameter of the estimated noise component is calculated. Data obtained by multiplying the data by the time variation rate of the noise level as a noise component acoustic parameter, from the acoustic parameter of the latest time section, and then converting the subtracted acoustic parameter into a time signal and outputting the converted signal. Characteristic noise reduction method. 12. The noise reduction method according to claim 11, wherein the acoustic parameter is a frequency spectrum. 13. The noise reduction method according to claim 11, wherein the acoustic parameter is a linear prediction coefficient. 14. An audio signal containing noise or the like is used as an input signal, and the input signal is multiplied by a window function to cut out a signal in a time section. Whether the cut out time section is a voice section or a noise section is determined. Judge, convert the signal of the extracted time section into a frequency spectrum, estimate the noise spectrum from the frequency spectrum of the signal of the time section determined as the noise section, and estimate the noise section from the frequency spectrum of the latest extracted time section Subtracting the obtained noise spectrum to design a frequency characteristic of a filter for realizing noise reduction, calculating a filter coefficient from the filter frequency characteristic to form a filter, and passing an input signal through a filter having the filter coefficient 4. A noise reduction method for realizing noise reduction of an input signal by using the method described in claim 1. The time variation of the noise level of the input signal is calculated by the method described above, and the result of multiplying the noise spectrum by the time variation of the noise level is subtracted from the frequency spectrum of the signal in the latest time section as a noise spectrum estimation value. A frequency characteristic of the filter. 15. An apparatus for reducing a noise added to an input signal by using a voice including noise or the like as an input signal,
And a noise reduction unit and a noise level time variation calculation unit. The input signal is passed to a noise reduction processing unit and a noise level time variation rate calculation unit. 6. The time variation of the noise level is calculated by the apparatus according to any one of the above items 6, and the obtained time variation of the noise level is passed to the noise reduction processing unit, and the amount of noise subtraction is calculated by means for subtracting the noise component from the input signal. A noise reduction device characterized by adjusting the following. 16. A means for taking a speech including noise or the like as an input signal and multiplying the input signal by a window function to cut out a signal in a time section, and the cut out time section is either a voice section or a noise section. Means for determining whether or not the cut-out signal in the time section is converted into an acoustic parameter; means for estimating the sound parameter of the noise component from the sound parameter of the signal in the time section determined to be the noise section; A noise reduction device comprising means for subtracting a noise component from an acoustic parameter of a time section extracted in the above and an acoustic parameter of a noise component estimated in a noise section. The estimated time variation rate of the noise level of the input signal is estimated in the noise section from the acoustic parameters of the time section cut out at the latest time. The parameter is passed to the means for subtracting the parameter, and the result of multiplying the estimated noise component acoustic parameter by the time variation rate of the noise level is subtracted as the noise component acoustic parameter from the acoustic parameter in the latest time section to reduce the noise of the input signal. A noise reduction device for obtaining processed acoustic parameters. 17. The noise reduction device according to claim 16, wherein the acoustic parameter is a frequency spectrum. 18. The noise reduction device according to claim 16, wherein the acoustic parameter is a linear prediction coefficient. 19. A means for taking a speech including noise or the like as an input signal and multiplying the input signal by a window function to cut out a signal in a time section, and the cut out time section is either a voice section or a noise section. Means for determining whether or not the cut-out signal in the time section is converted into an acoustic parameter; means for estimating the sound parameter of the noise component from the sound parameter of the signal in the time section determined to be the noise section; Means for subtracting the acoustic parameter of the noise component estimated in the noise section from the acoustic parameter of the time section cut out, and means for converting the acoustic parameter obtained by subtracting the acoustic parameter of the noise component into a time signal,
And a means for connecting and outputting the converted time waveform, wherein the time variation rate of the noise level of the input signal calculated by the apparatus according to any one of claims 4 to 6 is calculated based on the latest time. To the means for subtracting the acoustic parameter estimated in the noise section from the acoustic parameter in the time section extracted in the above, and multiplying the estimated acoustic parameter of the noise component by the time variation rate of the noise level as the latest acoustic parameter of the noise component. A noise reduction apparatus characterized in that after subtracting from acoustic parameters in a time section, the subtracted acoustic parameters are converted into a time signal and output. 20. The noise reduction device according to claim 19, wherein the acoustic parameter is a frequency spectrum. 21. The noise reduction device according to claim 19, wherein the acoustic parameter is a linear prediction coefficient. 22. A means for taking a speech including noise or the like as an input signal and cutting the signal into a signal of a time section by multiplying the input signal by a window function, and wherein the cut time section is either a speech section or a noise section. Means for determining whether or not the cut-out time interval signal is converted into a frequency spectrum; means for estimating a noise spectrum from the frequency spectrum of the signal in the time interval determined to be a noise interval; A means for designing a frequency characteristic of a filter for realizing noise reduction by subtracting a noise spectrum estimated in a noise section from a frequency spectrum of the filter, a means for calculating a filter coefficient from the frequency characteristic of the filter to constitute a filter, Achieving noise reduction of the input signal by passing the input signal through a filter having the calculated filter count A noise reduction device which calculates a time variation rate of a noise level of an input signal obtained by the device according to any one of claims 4 to 6, and estimates a noise spectrum estimated from a frequency spectrum of a signal in a cut-out time section. Subtract to the filter design means and multiply the noise spectrum by the noise level time variation to obtain the noise spectrum estimation value and subtract from the frequency spectrum of the signal in the latest time section to obtain the filter frequency characteristics. A noise reduction device characterized by determining.