KR101295727B1 - Apparatus and method for adaptive noise estimation - Google Patents

Abstract

The present invention relates to an adaptive noise estimation apparatus and method.
In this specification, a Fourier transform is performed on a noisy speech signal including a voice signal and a noise signal to obtain a Fourier spectrum, differential filtering of the Fourier spectrum to obtain a filtered Fourier spectrum, and a state of the noise present in a filter bank. Determining a first smoothing spectrum by smoothing the filtered Fourier spectrum in a frequency domain using a first smoothing coefficient set based on the determined filter bank, and setting a first smoothing spectrum based on the determined filter bank. Obtaining a second smoothing spectrum by smoothing the first smoothing spectrum in a time domain using two smoothing coefficients, and a forgetting coefficient set based on the determined filter bank, the second smoothing spectrum, and the second smoothing spectrum Using the minimum value of It discloses an adaptive noise estimation method comprising the step of estimating the sound.
According to the present invention, the improved speech can effectively suppress the residual of musical noise and the perception of speech distortion.

Description

Adaptive Noise Estimation Apparatus and Method {APPARATUS AND METHOD FOR ADAPTIVE NOISE ESTIMATION}

TECHNICAL FIELD The present invention relates to speech signal processing, and more particularly, to an apparatus and a method for determining a state of noise and adaptively estimating noise in a noisy speech signal.

Speech recognition generally refers to a process of converting an acoustic signal obtained through a microphone or a telephone into a word, a word set, or a sentence. The first step to improve the accuracy of the speech recognition is to efficiently extract only the speech components, which are acoustic signals, from the input signal where the noise and the voice coexisted through a single channel. In order to improve the sound quality of the voice signal in which the noise and the voice input through a single channel coexist, the noise component must be effectively weakened or eliminated without damaging the speech component.

Therefore, in the process of processing a noisy speech signal input through a single channel, a noise estimation is performed to accurately grasp the state of noise in the input noisy speech signal and also to obtain a noise component from the input noisy speech signal. The procedure is basically included. The estimated noise signal is used to improve sound quality by attenuating or removing noise components from the noisy speech signal.

Noise estimation is fundamentally required and the most important step in single channel speech enhancement. For example, speech improved using falsely estimated noise involves the following problems. First, musical noise remains when the estimated noise is evaluated lower than the original noise. Musical noise is a random frequency component and artifacts that perceptually annoy the listener. Second, if the estimated noise is evaluated higher than the actual noise, it causes speech distortion. Voice distortion is an unnatural hearing caused by attenuation of voice signals. It is very difficult to perform noise estimation so that it does not involve musical noise and distortion from voices contaminated with non-stationary noise.

One of the noise estimation methods is voice activity detection (VAD) and a VAD based noise estimation method using the same. According to the VAD-based noise estimation method, the state of noise and the noise are also estimated by using statistical information obtained from previously detected explicit noise frames or long past frames. An explicit noise frame refers to a silent frame or speech-absent frame that does not include speech or a noise-dominant frame in which the noise component is overwhelmingly superior to speech in the entire noisy speech signal.

This conventional VAD-based noise estimation method shows a very good performance when the background noise does not change with time. However, if the background noise is non-stationary or level-varying, if the signal to noise ratio (SNR) is low, or if the energy of the speech signal is weak, VAD-based The noise estimation method has a disadvantage in that it is difficult to obtain reliable information about the state of noise or the current noise level. This is due to the application of incorrectly detected VAD information. In addition, the VAD-based method has a problem of requiring a relatively high computational cost at various stages for noise estimation.

In order to overcome the shortcomings of the VAD based method, several new methods have been proposed. Among them, a well-known method is a weighted average (WA) method based on recursive averaging. The WA method estimates noise in the frequency domain without introducing VAD and continuously updates the estimated noise. According to the WA method, noise is estimated by applying a fixed forgetting factor between the magnitude spectrum of the speech signal contaminated in the current frame and the magnitude spectrum size of the noise estimated in the previous frame. However, since the WA method applies a fixed forgetting coefficient, it cannot reflect the noise change in various noise environments or non-static noise environments, and as a result, there is a limit in that it is impossible to perform a proper noise estimation.

Another noise estimation method proposed to overcome the shortcomings of the VAD-based noise estimation method is to use a minimum statistics (MS) method. According to this, the minimum value of the smoothed power spectrum of the noisy speech signal is tracked over the search window, and the noise is estimated by multiplying the minimum value by the compensated constant. Here, the search window means the length of the past frame corresponding to about 1.5 seconds. The MS method generally shows good performance, but it requires a large amount of memory because the information of the long past frame corresponding to the search window length is constantly needed, especially in the noise dominant signal, which is mostly noise. The disadvantage is that it is not possible to track changes in the noise level quickly. In addition, since the MS method basically uses the estimated noise information of the past frame, there is a limit in that a reliable result cannot be obtained when the level of the noise is severe or changes to another noise environment.

In order to solve the shortcomings of the conventional MS method, a minimal controlled recursive averaging (MCRA) method has been proposed, which has a smoothing coefficient adjusted by signal presence probability at the current frequency position. Noise estimation is performed by regression averaging. In addition, several types of modified MS methods are proposed, two of which are common to the following. The first is to include a VAD method for continuously distinguishing whether voice is present or silent in the current frame or frequency bin under consideration, and the second feature is Recursive Averaging (RA). Based noise estimator is used. However, most of the modified schemes have similar problems because they are based on tracking the minimum of the MS method.

An object of the present invention is to provide an apparatus and method for adaptively estimating noise without using statistical information presented by various explicit noise frames or long past frames detected by the VAD.

Another technical problem of the present invention is Magnitude Signal to Noise Ratio (MSNR) for determining the state of noise in a signal subjected to filtering using a differentiator in the frequency domain, smoothing in the frequency domain, and adaptive smoothing in the time domain. An apparatus and method for adaptively estimating noise using SNR) and forward tracking signal-to-noise ratio (FTSNR) are provided.

According to one aspect of the present invention, an adaptive noise estimation method is provided. The adaptive noise estimating method includes Fourier transforming a noisy speech signal including a speech signal and a noise signal to obtain a Fourier spectrum, differentially filtering the Fourier spectrum to obtain a filtered Fourier spectrum, Determining a state of the noise present in a filter bank, and smoothing the filtered Fourier spectrum in a frequency domain using a first smoothing coefficient set based on the determined filter bank. Obtaining a smoothing spectrum, obtaining a second smoothing spectrum by smoothing the first smoothing spectrum in a time domain using a second smoothing coefficient set based on the determined filter bank, and based on the determined filter bank The forgetting coefficient set, the second smoothing spectrum and the 2 using the minimum value of the smoothed spectrum and a step of estimating the noise.

The state of the noise may be determined by magnitude signal to noise ratio (MSNR) and forward search signal to noise ratio (FSSNR).

The estimating of the noise may further estimate the noise using the noise estimated in the previous frame.

The minimum value of the second smoothed spectrum may be the minimum spectrum detected from a constant past search window.

The first smoothing coefficient, the second smoothing coefficient, and the forgetting coefficient are first exponents for determining total fast fourier transform (FFT) points and second exponents for determining the number of filter banks. The filter bank index determined by the difference between the filter bank index and the second filter bank index obtained by dividing the entire filter bank into low, mid, and high frequencies may be variably set.

The first smoothing coefficient, the second smoothing coefficient, and the forgetting coefficient may be variably set according to a result of comparing the magnitude signal-to-noise and the omnidirectional search signal-to-noise with a threshold.

According to another aspect of the present invention, an apparatus for adaptive noise estimation is provided. The adaptive noise estimator is a Fourier transform unit for Fourier transforming a noisy speech signal including a speech signal and a noise signal to obtain a Fourier spectrum, and differentially filtering the Fourier spectrum to obtain a filtered Fourier spectrum. Frequency of the filtered Fourier spectrum using a filtering unit, a noise state determining unit for determining a state of the noise present in a filter bank, and a first smoothing coefficient set based on the determined filter bank. Smoothing the first smoothing spectrum in the time domain by smoothing the first smoothing spectrum in a time domain using a second smoothing coefficient set based on the first smoothing unit and the determined filter bank by smoothing in a region. Unit, and a network set up based on the determined filterbank Coefficient, comprises the second smoothing and a noise spectrum estimating unit for estimating noise using the second minimum value of the smoothed spectrum.

The noise state determination unit may determine the state of the noise by magnitude signal to noise ratio (MSNR) and forward search signal to noise ratio (FSSNR).

The noise estimation unit may further estimate the noise using the noise estimated in the previous frame.

The minimum value of the second smoothed spectrum may be the minimum spectrum detected from a constant past search window.

The noise state determination unit includes a filterbank index determined by a difference between a first exponent for determining a total fast fourier transform (FFT) point and a second exponent for determining the number of the filter banks; The first smoothing coefficient, the second smoothing coefficient, and the forgetting coefficient may be variably set according to a result of comparing the second filter bank index obtained by dividing the entire filter bank into low, mid, and high pitches.

The noise state determination unit may variably set the first smoothing coefficient, the second smoothing coefficient, and the forgetting coefficient according to a result of comparing the magnitude signal-to-noise and the omnidirectional search signal-to-noise with a threshold.

According to the present invention, the improved speech can effectively suppress the residual of musical noise and the perception of speech distortion.

1 is a flowchart illustrating an operation of a noise estimation apparatus according to an exemplary embodiment of the present invention.
FIG. 2 is a block diagram illustrating a noise estimation apparatus for performing the noise estimation method of FIG. 1.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. The following embodiments are intended to illustrate the technical concept of the present invention, and therefore the technical idea of the present invention should not be interpreted as being limited by this embodiment. In the description of the present embodiment and the drawings, reference numerals added to respective components are merely described for convenience of description, and like reference numerals refer to like elements throughout the specification.

In the following, the speech and noise models of contaminated speech in the frequency domain by Fourier transform are examined, and then the MA method, WA method, MS method, and MCRA, which perform noise estimation at the current frame or frequency location The method is disclosed.

The noisy speech y (n) contaminated from the noise n (n) added to the clean voice signal x (n) in the frequency domain is expressed by the following equation.

Referring to Equation 1, n is a discrete time index. y (n) may be approximated to a Fourier Spectrum (FS) Y _i (f) by a short-time Fourier transform as shown in Equation (2).

은 다음의 수학식과 같이 표현된다. Referring to Equation 2, i and f are the frame and frequency position index, respectively. X _i (f) is the FS of clean speech, and N _i (f) is the FS of noise. Also, Fourier Power Spectrum (FPS) squared with Equation 2

Is expressed as the following equation.

를 의미한다. 수학식 3에서 cos(△θ_i)가 1 또는 0을 가지면 오염된 음성의 모델은 각각 수학식 4와 수학식 5로 표현될 수 있다. Referring to Equation 3, Δθ _i is a phase difference between speech and noise.

. If cos (Δθ _i ) in Equation 3 has 1 or 0, the contaminated speech models may be represented by Equations 4 and 5, respectively.

는 푸리에 크기 스펙트럼(Fourier Magnitude Spectrum: FMS)이다. 푸리에 크기 스펙트럼이 수학식 4와 같이 유도되는 것은

이기 때문이며, 수학식 5에서 푸리에 크긴 스펙트럼이 수학식 5와 같이 유도되는 것은

이기 때문이다. In Equation 4,

Is the Fourier Magnitude Spectrum (FMS). The Fourier magnitude spectrum is derived as in Equation 4

This is because the Fourier-long spectrum in Equation 5 is derived as in Equation 5.

.

의 평균을 취해 잡음추정을 얻는 방식이며 다음의 수학식으로 표현된다. MA method frequency-position presented in more noise frame M ₁ is detected by the VAD based noise estimation method

A noise estimate is obtained by taking the average of and is expressed by the following equation.

는 추정된 잡음의 FMS이다. 그러나 MA 방법은 비정적인 잡음에 오염된 음성으로부터 신뢰적인 VAD를 보장하기 어렵기 때문에 올바른 잡음추정을 얻기 어렵다. 또한 VAD는 잡음추정을 위한 여러 단계에서 상대적으로 높은 계산비용을 빈번히 요구하는 측면이 있다.Referring to Equation 6,

Is the FMS of the estimated noise. However, the MA method is difficult to obtain a correct noise estimate because it is difficult to guarantee a reliable VAD from speech contaminated with non-static noise. VAD also frequently requires relatively high computational cost at various stages for noise estimation.

The WA method obtains a noise estimate by first-order regression averaging with a fixed forgetting coefficient a ₁ , and estimates the noise by the following equation.

을 만족하는 경우 잡음추정을 갱신하기 위함이며 일반적으로 1에 근접한 값을 사용한다. β(

)는 음성과 잡음 사이를 묵시적(implicit)으로 구분하기 위한 임계치(threshold)이다. 그러나 비정적인 잡음환경에서 WA 방법은 잡음의 변화나 성분을 고려할 수 없는 고정된 망각 계수가 사용되기에 올바른 잡음추정을 수행하지 못하는 문제가 있다. Where α ₁ (0≤α ₁ ≤1)

If is satisfied, it is used to update the noise estimate. Generally, a value close to 1 is used. β (

) Is a threshold for implicitly distinguishing between speech and noise. However, in a non-static noise environment, the WA method has a problem in that it does not perform a proper noise estimation because a fixed forgetting coefficient is used, which cannot take into account changes or components of noise.

의 평활된 파워 스펙트럼 D_i(f)에서 추적된 최소값 D_i ^min(f)에 보상상수 cf=1.5를 곱하여 주어지는 잡음추정이다. D_i(f)와 D_i ^min(f),

는 각각 수학식 8, 9, 10과 같이 정의된다. MS method is across the navigation window

The noise estimate given by multiplying the minimum value D _i ^min (f) traced from the smoothed power spectrum D _i (f) by cf = 1.5. D _i (f) and D _i ^min (f),

Are defined as Equations 8, 9, and 10, respectively.

Referring to Equation 8 to _{10, α 2 (0≤α 2 ≤1} ) is the smoothing factor (smoothing factor) for obtaining D _i (f), M ₂ denotes a length of the frame in the past for about 1.5 s . Although the MS method generally provides good noise estimation performance, there is a problem that statistical information of long past frames corresponding to the search window is required.

와 α_3,i(f), p_i(f)는 각각 수학식 11, 12 및 13에 의해 정의된다. The noise estimation by MCRA, which is a variation of the MS method, is given by regression averaging with the forgetting coefficient α _{3, i} (f) adjusted by the signal presence probability p _i (f).

And α _{3, i} (f) and p _i (f) are defined by equations (11), (12) and (13), respectively.

Referring to Equations 11 to 13, α ₄ and α ₅ are smoothing coefficients of α _{3, i} (f) and p _i (f), respectively. I _i (f) is defined by the following equation.

는 임계치 간의 비교를 통해 잡음과 음성 사이를 연속적으로 구분하기 위한 식별자이다. D_i(f)와 D_i ^min(f)는 각각 수학식 8에서 평활된 파워 스펙트럼과 수학식 9에서 추적된 최소값이다. Referring to Equation 14,

Is an identifier for continuously distinguishing between noise and speech through a comparison between thresholds. D _i (f) and D _i ^min (f) are the power spectrum smoothed in Equation 8 and the minimum value tracked in Equation 9, respectively.

Hereinafter, a noise estimation method according to the present invention will be described in detail.

1 is a flowchart illustrating an operation of a noise estimation apparatus according to an exemplary embodiment of the present invention.

Referring to FIG. 1, a method of estimating noise of a noisy speech signal by a noise estimating apparatus includes a Fourier transform (S100), an filtering step using a differentiator (S105), on an input noisy speech signal, Frequency smoothing step (Frequency Smoothing, S110), time domain adaptive smoothing step (S115), step of obtaining magnitude SNR and omnidirectional search SNR (S120) and magnitude SNR and omnidirectional SNR Adaptive noise estimation step (S125) using a. Hereinafter, each step of configuring an embodiment of the present invention for processing an input noisy speech signal and estimating noise will be described in more detail.

The noise estimating apparatus performs a Fourier transform on the input noisy speech signal y (n) (S100). y (n) may be defined as the sum of the clean speech signal component x (n) and the noise signal component n (n) as shown in Equation (1). Fourier transform is performed continuously on the short-time signal of the input noisy speech signal y (n), resulting in an input noisy speech signal. y (n) may be approximated to Fourier Spectrum (FS) Y _i (f) as shown in Equation 2.

The noise estimating apparatus filters the Fourier spectrum Y _i (f) based on the derivative and outputs the filtered Fourier spectrum Y ' _i (f) (S105). Differential based filtering is defined as in Equation 15.

Referring to Equation 15, i is a frame index and ψ (f) is a weight for enhancing signal components for each frequency band. Voice is divided into voiced sound that occurs while passing through the body's gates and unvoiced sound that does not pass through the gate, and the voiceless voice has a lower energy than the voiced sound. white) Similar to noise, it is an aperiodic signal and has a characteristic of being distributed over the entire band. As a result, the unvoiced sound is more difficult to distinguish from the contaminated voice signal than the voiced sound. However, using the filtering of the differentiator as shown in Equation 15, it is possible to efficiently classify and extract the characteristics of the unvoiced sound from the speech signal contaminated with noise.

The noise estimation apparatus performs smoothing in the frequency domain (S110), whereby a first smoothing spectrum Y ″ _i (f) is obtained. Smoothing in the frequency domain is defined by the following equation.

Referring to Equation 16, a ₁ , a ₂ , ..., a _v1 are v 1st order smoothing coefficients that smoothly transform the non-static signal on the frequency axis, and a ₁ + a ₂ + ... + a _v1 = 1 to be.

By performing smoothing in the frequency domain, the noise estimator can estimate noise more accurately from speech contaminated with various forms of non-static noise, and use the incorrectly estimated noise to improve the speech. It can reduce the incidence of accompanying musical noise or voice distortion.

On the other hand, when noise is estimated by the prior art from speech contaminated by a signal in which the noise is variably developed on the time axis, i) residual noise is caused by the estimated evaluation lower than actual noise in the area where the noise increases. , ii) speech distortion is caused by the estimated estimates higher than actual noise in areas where noise is reduced. Therefore, the noise estimating apparatus may use a smoothing coefficient fixed on the time axis as shown in Equation 8 to perform reliable noise estimation. However, the problem of attenuation of speech characteristics and non-static maintenance of noise change according to an approximation of the smoothing coefficient remains.

Therefore, the noise estimator uses an adaptive smoothing coefficient calculated in units of several filter banks divided from the entire Fast Fourier Transform (FFT) points, thereby providing the first smoothing spectrum Y " _i (f). Adaptive smoothing in the time domain is performed in operation S 115. As an example, the noise estimation apparatus performs a second smoothing spectrum S _i using the adaptive smoothing coefficient b ₁ (j) in the time domain as shown in Equation 17. _{, j} (k) is extracted.

Referring to Equation 17, j (0 ≦ j ≦ 2 ^Pp −1) is the filterbank 2 ^Pp index divided by 2 ^p from the total FFT points 2 ^P , and the uppercase P is an exponent for determining the FFT point. , Lowercase p is an index for determining the number of filter banks, k (0 ≦ k ≦ j · 2 ^p −1) is a spectral bin index existing in the filter bank, b ₁ (j), b ₂ ( j), ..., b _v2 (j) is the v2 order adaptive smoothing coefficient. Here, b ₁ (j) + b ₂ (j) + ... + b _v2 (j) = 1. b ₁ (j), b ₂ (j), ..., b _v2 (j) are set by Table 1 to be described later.

When a low weight is assigned to the smoothing coefficient b ₁ (j) in the speech-like filter bank, the characteristics of the speech can be effectively represented. When a high weight is assigned to the smoothing coefficient b ₁ (j) in the noise-like filter bank, noise characteristics can be effectively suppressed. According to this, it is possible to maintain the characteristics of the voice while stably converting to the variable noise change. As a result, the conventional method of applying a fixed smoothing coefficient without considering the change of noise does not effectively reduce the effect of variable noise, and when the smoothing coefficient close to 1 is applied, the change in noise is stable. The problem of attenuation of and the smoothing coefficient approaching zero can be solved.

The noise estimating apparatus determines a state of noise existing in the filter bank (S120). For example, the determination of the state of the noise may be made by magnitude SNR or omnidirectional SNR indicating the state of the noise. First, the magnitude SNR may be obtained by Equation 18.

In addition, the omnidirectional SNR may be calculated by Equation 19.

는 이전 프레임에서 추정된 잡음이다. γ_i(j)(또는 φ_i(j))가 1에 가까우면 해당 필터뱅크는 잡음 유사 필터뱅크를 나타내고, φ_i(j)(또는 γ_i(j))가 0에 가까우면 해당 필터뱅크는 음성 유사 필터뱅크를 나타낸다. 이와 같이 잡음 추정 장치는 γ_i(j) 또는 φ_i(j)의 크기에 기반한 필터뱅크의 판단 기준에 따라 필터뱅크의 잡음 상태를 결정할 수 있다. Referring to Equation 18 and Equation 19, FBS is the filter bank size,

Is the noise estimated in the previous frame. If γ _i (j) (or φ _i (j)) is close to 1, the corresponding filter bank represents a noise-like filter bank, and if φ _i (j) (or γ _i (j)) is close to 0, the corresponding filter bank Denotes a negative like filterbank. As such, the noise estimating apparatus may determine the noise state of the filter bank based on a criterion of the filter bank based on the size of γ _i (j) or φ _i (j).

On the other hand, T _{i , j} (k) is the omnidirectional search spectrum. The omni-directional search is for estimating the state of noise present in the second smoothed spectrum _{Si , j} (k) in the entire frame or in several subbands divided in one frame. T _{i , j} (k) is defined by the following equation (20).

Referring to Equation ^{_{20, S min i, j (}} k) is the minimum spectrum detected from the S _{i, j} (k) of constant previous search _{window, c 1 (j), c} 2 (j), ... , c _v2 (j) is the v2 order adaptive forgetting factor. Here, c ₁ (j) + c ₂ (j) + ... + c _v2 (j) = 1. The noise estimation apparatus may assign a relatively low weight to the adaptive forgetting factor c ₁ (j) in the speech like filter bank so that the omnidirectional search spectrum has a noise like spectrum. In addition, the noise estimation apparatus may assign a relatively high weight to the adaptive forgetting factor c ₁ (j) in the noise like filter bank so that the omnidirectional search spectrum has the noise like spectrum (as described in the proposed proposal). C ₁ (j), c ₂ (j), ..., c _v2 (j) and the forgetting factor d (j) are described by the algorithm of Table 1 to be described later. Is set.

Referring to Table 1, FB2 (m) is the secondary filterbank index divided by the user's arbitrary bass, midtone, and treble from the total filterbank, M is the total number of secondary filterbanks, and TH ^{FB2 (m)} _φ ( l) is the threshold of φ _i (j) and TH ^{FB2 (m)} _γ (l) is the threshold of γ _i (j). And L is the total step divided by the user according to TH ^{FB2 (m)} _φ (l) and TH ^{FB2 (m)} _γ (l), W ^{FB2 (m)} _b1 (l), ..., W ^{FB2 (m )} _bv3 (l) is the _weight of b ₁ (j), b ₂ (j), ..., b _v2 (j), W ^{FB2 (m)} _c1 (l), ..., W ^{FB2 (m)} _cv3 (l) is the _weight of c ₁ (j), c ₂ (j), ..., c _v2 (j), and W ^{FB2 (m)} _d (l) is the weight of d (j).

을 아래의 수학식 21에 의해 구한다(S125). The noise estimation apparatus estimates noise using d (j) and the minimum value S ^min _{i, j} (k) calculated by the algorithm of Table 1 above.

Is obtained by the following equation (21) (S125).

As described in detail above, the noise estimation method according to the present invention estimates noise by applying an adaptive forgetfulness coefficient instead of the conventional WA technique of applying a fixed forgetting element for each frame regardless of noise change. As a result, unlike the conventional technology using statistical information presented by various noise frames obtained by VAD, a correct VAD can be obtained even in a low noise environment or a low SNR where the noise environment is variously non-static, or the voice region is low. Estimation can be performed adaptively, resulting in reliable noise estimation. In addition, the present embodiment is relatively easy to implement in real hardware or software because the calculation amount is relatively small and the amount of memory required is not large.

FIG. 2 is a block diagram illustrating a noise estimation apparatus for performing the noise estimation method of FIG. 1. Referring to FIG. 2, the noise estimation apparatus 200 includes a Fourier transform unit 205, a filtering unit 210, a first smoothing unit 215, a second smoothing unit 220, and a noise state for an input noisy speech signal. Determination unit 225 and noise estimation unit 230. Functions of the respective components 205, 210, 215, 220, 225, and 230 included in the noise estimation apparatus 200 according to the embodiment of the present invention are the noise estimation procedure according to the embodiment of FIG. 1 of the present invention described above. Since the descriptions in the steps S100, S105, S110, S115, S120, and S125 may be equally applicable, detailed descriptions thereof will be omitted below. The noise estimating apparatus 200 according to the exemplary embodiment of the present invention is provided in a voice-based application device or a voice recognition system such as a speakerphone or a video call communication device, a hearing aid, a Bluetooth device, and the like, to determine a state of noise from an input noisy voice signal. And can also be used to estimate noise, improve sound quality, and / or use speech recognition.

The above description is merely illustrative of the technical idea of the present invention, and those skilled in the art to which the present invention pertains may make various modifications and variations without departing from the essential characteristics of the present invention. Therefore, the embodiments disclosed in the present invention are intended to illustrate rather than limit the scope of the present invention, and the scope of the technical idea of the present invention is not limited by these embodiments. The scope of protection of the present invention should be construed according to the following claims, and all technical ideas falling within the scope of the same shall be construed as falling within the scope of the present invention.

Claims (12)

Fourier transforming the noisy speech signal including the speech signal and the noise signal to obtain a Fourier spectrum;
Differentially filtering the Fourier spectrum to obtain a filtered Fourier spectrum;
Determining a state of the noise present in a filter bank;
Obtaining a first smoothing spectrum by smoothing the filtered Fourier spectrum in a frequency domain using a first smoothing coefficient set based on the determined filter bank;
Obtaining a second smoothing spectrum by smoothing the first smoothing spectrum in a time domain by using a second smoothing coefficient set based on the determined filter bank; And
Estimating noise using a forgetting coefficient set based on the determined filter bank, a minimum value of the second smoothing spectrum, and the second smoothing spectrum,
And the state of the noise is determined by magnitude signal to noise ratio (MSNR) and forward search signal to noise ratio (FSSNR). delete The method of claim 1, wherein estimating the noise comprises:
And estimating the noise further using the noise estimated in the previous frame. The method of claim 1,
And the minimum value of the second smoothed spectrum is a minimum spectrum detected from a constant past search window. The method of claim 1, wherein the first smoothing coefficient, the second smoothing coefficient and the forgetting coefficient,
A filter bank index determined by a difference between a first exponent for determining an overall fast fourier transform (FFT) point and a second exponent for determining the number of the filter banks, and a bass and midtone for the entire filter bank. The noise estimation method, characterized in that the variable is set in accordance with the result of comparing the filter bank index divided by the treble. The method of claim 1, wherein the first smoothing coefficient, the second smoothing coefficient and the forgetting coefficient,
And variably set according to a result of comparing the magnitude signal-to-noise and the omnidirectional search signal-to-noise with a threshold. A Fourier transform unit for Fourier transforming a noisy speech signal including a speech signal and a noise signal to obtain a Fourier spectrum;
A filtering unit for differentially filtering the Fourier spectrum to obtain a filtered Fourier spectrum;
A noise state determination unit for determining a state of the noise present in a filter bank;
A first smoothing unit for smoothing the filtered Fourier spectrum in a frequency domain by using a first smoothing coefficient set based on the determined filter bank;
A second smoothing unit for smoothing the first smoothing spectrum in a time domain by using a second smoothing coefficient set based on the determined filter bank; And
And a noise estimating unit estimating noise by using the forgetting coefficient set based on the determined filter bank, the second smoothing spectrum, and the minimum value of the second smoothing spectrum.
And the noise state determination unit determines the state of the noise by magnitude signal to noise ratio (MSNR) and forward search signal to noise ratio (FSSNR). Estimation device. delete The method of claim 7, wherein the noise estimation unit,
And estimating the noise further using the noise estimated in the previous frame. The method of claim 7, wherein
And the minimum value of the second smoothed spectrum is a minimum spectrum detected from a constant past search window. The method of claim 7, wherein the noise state determination unit,
A filter bank index determined by a difference between a first exponent for determining an overall fast fourier transform (FFT) point and a second exponent for determining the number of the filter banks, and a bass and midtone for the entire filter bank. And variably setting the first smoothing coefficient, the second smoothing coefficient, and the forgetting coefficient according to a result of comparing the second filter bank index divided by the treble. The method of claim 7, wherein the noise state determination unit,
And variably setting the first smoothing coefficient, the second smoothing coefficient, and the forgetting coefficient according to a result of comparing the magnitude signal-to-noise and the omnidirectional search signal-to-noise with a threshold.

Images