Single-channel frequency-domain speech denoising method based on variable-length filter

Abstract

The application relates to a single-channel frequency-domain speech noise reduction method based on a variable-length filter, which comprises the following steps: collecting a noisy speech signal, pre-processing the noisy speech signal; estimating the statistical characteristics of the noisy speech signal and a noise signal based on the pre-processed noisy speech signal, estimating a variable-length speech noise reduction filter through the statistical characteristics; and performing noise reduction filtering according to the variable-length speech noise reduction filter to obtain an estimated value of a clean speech signal. Through research on the interframe correlation coefficient vector of a speech signal, a set of variable-length speech noise reduction filters are derived by utilizing the time-varying characteristics of the speech signal. The method can utilize the interframe correlation of the speech signal, flexibly determine the length of the filter, and has relatively low calculation complexity.

Description

Technical Field

[0001] This invention relates to the field of speech denoising technology, and in particular to a single-channel frequency domain speech denoising method based on a variable-length filter. Background Technology

[0002] Noise is a significant factor affecting the quality of voice communication in daily life. Voice noise reduction technology improves voice quality and intelligibility by suppressing noise and extracting clean speech signals, playing an important role in applications such as mobile communication, conferencing systems, and voice interaction.

[0003] Based on the assumption that the short-time Fourier transform coefficients of adjacent speech frames are uncorrelated, most early short-time Fourier transform domain denoising methods typically employed filter gain, meaning the filter length at each frequency point was 1. However, since speech frames need to overlap during denoising, and the speech signal itself is strongly autocorrelated, the short-time Fourier transform coefficients of the current speech frame and adjacent speech frames are highly correlated. Based on these facts, to improve the performance of speech denoising methods, researchers have proposed various denoising methods that utilize the inter-frame correlation of the speech signal, where the filter length at each frequency point is greater than 1.

[0004] Theoretically, longer filters generally result in better noise reduction performance. However, in practice, the correlation between the current frame and distant frames may be weak. Using excessively long filters can introduce too much interference, increasing the estimation error of signal statistical characteristics and potentially leading to a decrease in filter performance. Furthermore, computational complexity increases rapidly with filter length. Conversely, using excessively short filters fails to fully utilize the inter-frame correlation of the speech signal. Moreover, since speech signals are non-stationary, the correlation between the current speech frame and adjacent frames varies across different frequency bands. Even within the same frequency band, the correlation between the current and adjacent frames changes over time (e.g., when the phoneme types in the speech signal (such as vowels, consonants, plosives, fricatives, etc.) change, the correlation between the current and adjacent frames also changes). Therefore, determining an optimal filter length is difficult. Summary of the Invention

[0005] The purpose of this invention is to provide a single-channel frequency domain speech denoising method based on variable-length filters to address the problems existing in the prior art. By studying the inter-frame correlation coefficient vector of the speech signal and utilizing the time-varying characteristics of the speech signal, a set of variable-length speech denoising filters is derived. This method can utilize the inter-frame correlation of the speech signal, flexibly determine the filter length, and has low computational complexity.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A class of single-channel frequency domain speech denoising methods based on variable-length filters includes:

[0008] Acquire noisy speech signals and preprocess the noisy speech signals;

[0009] Based on the preprocessed noisy speech signal, the statistical characteristics of the noisy speech signal and the noise signal are estimated, and the variable-length speech denoising filter is estimated using the statistical characteristics.

[0010] The variable-length speech denoising filter is used to perform denoising filtering to obtain an estimate of the clean speech signal.

[0011] Optionally, preprocessing the noisy speech signal includes: performing frame segmentation, windowing, and Fourier transform on the noisy speech signal, and introducing inter-frame correlation. The method for preprocessing the noisy speech signal is as follows:

[0012] In speech denoising, the time-domain signal model is represented as:

[0013] y(t) = x(t) + v(t)

[0014] Where t represents a discrete time point, x(t) and v(t) represent the desired signal and additive noise, respectively, and y(t) represents the noisy speech signal;

[0015] After framing, windowing, and Fourier transform, the short-time Fourier transform domain signal model is as follows:

[0016] Y(k,n)=X(k,n)+V(k,n)

[0017] Where Y(k,n), X(k,n), and V(k,n) represent the short-time Fourier transform coefficients of y(t), x(t), and v(t) at the k-th frequency point in the n-th frame, respectively.

[0018] Optionally, introducing the inter-frame correlation includes:

[0019] Combine N consecutive time frames together:

[0020] y(k,n)=[Y(k,n)Y(k,n-1)…Y(k,n-N+1)] T =x(k,n)+v(k,n)

[0021] x(k,n)=[X(k,n)X(k,n-1)…X(k,n-N+1)] T

[0022] v(k,n)=[V(k,n)V(k,n-1)…V(k,n-N+1)] T

[0023] Where y(k,n) represents a noisy speech signal vector of length N, x(k,n) is a clean speech signal vector, and v(k,n) is a noisy signal vector, with superscript (·). T The expression represents the transpose, and X(k,n) is the desired signal. That is, the goal of noise reduction is to estimate the signal X(k,n) through the noisy signal vector.

[0024] Optionally, estimating the statistical properties of the noisy speech signal includes:

[0025] Estimate the correlation matrix of the noise signal vector and the correlation matrix of the noisy speech signal vector, and obtain the correlation matrix of the clean speech signal vector by subtracting the correlation matrix of the noise signal vector from the correlation matrix of the noisy speech signal vector.

[0026] The method for estimating the correlation matrix of the noisy speech signal vector is as follows:

[0027] Φ y (k,n)=α y Φ y (k,n-1)+(1-α y )y(k,n)y H (k,n)

[0028] Where, Φ y (k,n) is the correlation matrix, where k represents the k-th frequency point, n represents the n-th frame, y(k,n) is the noisy speech signal vector, αy is the forgetting factor, and (·) H It is the conjugate transpose;

[0029] The method for obtaining the correlation matrix of the clean speech signal vector is as follows:

[0030] Φ x (k,n)=Φ y (k,n)-Φ v (k,n)

[0031] Where, Φ x (k,n) is the correlation matrix for estimating the speech signal vector x(k,n), Φ v (k,n) is the correlation matrix of the noise signal vector.

[0032] Optionally, estimating the variable-length speech noise reduction filter includes:

[0033] Based on the correlation coefficient vector and the threshold parameter δ, where 0 ≤ δ ≤ 1, a vector of variable-length correlation coefficients is obtained. and vectors Length;

[0034] The method for obtaining the correlation coefficient vector is as follows:

[0035] Through the correlation matrix Φ x The correlation coefficient vector ρ can be obtained from (k,n). X (k,n)=[ρ X (k,n) ρ X (k,n-1) …ρ X (k, n-N+1)] T , where the correlation matrix Φ x The first column of (k,n) divided by the correlation matrix Φ x The element in the first row and first column of (k,n) is the correlation coefficient vector ρ. X (k,n);

[0036] Get vector The method is as follows:

[0037] Based on the correlation coefficient vector ρ X The variable-length correlation coefficient vector can be obtained by comparing the elements in (k,n) with the threshold parameter δ.

[0038]

[0039] in, Let P be the length of the variable-length correlation coefficient vector, where 1 ≤ P ≤ N, (·) T The superscript (·) indicates that the vector length is variable;

[0040] Set the maximum length threshold for the variable-length speech noise reduction filter based on the vector. The length P of the variable-length filter is used to determine the variable-length filter. Variable length vector Variable length vector and variable length vectors Length:

[0041]

[0042]

[0043]

[0044]

[0045] in, This indicates a variable-length noise reduction filter. Let P be a variable-length correlation coefficient vector. It is a clean speech signal vector of variable length. It is a variable-length noise signal vector.

[0046] Optionally, obtaining an estimate of the clean speech signal includes:

[0047] The variable-length speech noise reduction filter includes: a variable-length Wiener filter, a variable-length minimum variance distortionless response filter, and a variable-length compromise filter;

[0048] Based on the variable-length Wiener filter, the variable-length minimum variance distortionless response filter, or the variable-length compromise filter, combined with the noisy speech signal vector Obtain an estimate of the clean speech signal in For the variable length Wiener filter The variable-length minimum variance distortionless response filter and the variable length compromise filter Any one of them, and the estimated value of the clean speech signal. Convert to the time domain.

[0049] Optionally, estimating the variable-length Wiener filter includes:

[0050] Based on variable length filter Define the subband mean square error of the expected signal estimate, and decompose the subband mean square error into the subband mean square error based on subband speech distortion and the subband mean square error based on residual interference and noise. Take the derivative and set the result to zero;

[0051] Obtaining a variable-length Wiener noise reduction filter:

[0052]

[0053] in, Let X(k,n) be a variable-length Wiener filter, and φX(k,n) be the variance of the desired signal X(k,n). For matrix The inverse matrix, A variable-length noisy signal vector The correlation matrix is ​​the correlation matrix Φ of the noisy speech signal. y A matrix consisting of the first P rows and first P columns of (k,n), with superscript (·). * Indicates complex conjugation.

[0054] Optionally, the method for estimating the variable-length minimum variance distortionless response filter is as follows:

[0055]

[0056] in, It is a variable-length minimum variance distortionless response filter.

[0057] Alternatively, the method for estimating the variable-length compromise filter is as follows:

[0058]

[0059] in, It is a variable-length compromise filter. For vectors Correlation matrix, Φ in (k,n) To estimate the correlation matrix of the interference signal plus the noise signal, μ is a freely set parameter, μ≥0.

[0060] The beneficial effects of this invention are as follows:

[0061] This invention derives a set of variable-length speech denoising filters by studying the inter-frame correlation coefficient vector of speech signals and utilizing the time-varying characteristics of speech signals. This method can utilize the inter-frame correlation of speech signals, flexibly determine the filter length, and also has low computational complexity.

[0062] The concept proposed in this invention can be easily extended to the time domain or other transform domains. Compared with the prior art, this invention can reduce computational complexity while maintaining the output speech quality of traditional fixed-length filters. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a schematic diagram of a single-channel frequency domain speech denoising method based on a variable-length filter, according to an embodiment of the present invention.

[0065] Figure 2 This is a schematic diagram illustrating how the performance of the variable-length Wiener filter varies with parameter δ when the input signal-to-noise ratio is different, according to an embodiment of the present invention; wherein, Figure 2 (a) Output signal-to-noise ratio. Figure 2 (b) Speech distortion coefficient, Figure 2 (c) PESQ score. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] like Figure 1 As shown, this invention discloses a single-channel frequency domain speech denoising method based on a variable-length filter, specifically including the following steps:

[0069] Step 1: Acquire noisy speech signals;

[0070] Step 2: Perform preprocessing on the noisy speech signal, including framing, windowing, and Fourier transform, and introduce inter-frame correlation.

[0071] Step 3: Estimate the statistical characteristics of the noisy speech signal and the noise signal;

[0072] Step 4: Estimate the variable-length speech noise reduction filter;

[0073] Step 5: Filter the noisy speech signal in the frequency domain to obtain an estimate of the clean speech signal;

[0074] Step 6: Convert the filtered frequency domain speech signal to the time domain.

[0075] The specific method for step 2 is as follows:

[0076] In speech denoising, the time-domain signal model is represented as: y(t)=x(t)+v(t);

[0077] Where t represents a discrete time point, x(t) and v(t) represent the desired signal (usually clean speech) and additive noise, respectively, x(t) and v(t) are uncorrelated, and y(t) represents the noisy speech signal.

[0078] Assuming all signals are zero-mean, wideband real signals, after framing, windowing, and Fourier transform, the short-time Fourier transform domain signal model can be written as:

[0079] Y(k,n)=X(k,n)+V(k,n)

[0080] Where Y(k,n), X(k,n), and V(k,n) represent the short-time Fourier transform coefficients of y(t), x(t), and v(t) at the k-th frequency point in the n-th frame, respectively.

[0081] The variance of Y(k,n) is φY(k,n)=E[|Y(k,n)| 2 ]=φX(k,n)+φV(k,n), where φX(k,n)=E[|X(k,n)| 2 ] and φV(k,n)=E[|V(k,n)| 2 Let E[·] represent the variances of X(k,n) and V(k,n), respectively, and let |·| represent the expected value.

[0082] To introduce inter-frame correlation, N consecutive time frames are combined together, represented as:

[0083] y(k,n)=[Y(k,n)Y(k,n-1)…Y(k,n-N+1)] T =x(k,n)+v(k,n)

[0084] Where y(k,n) represents a noisy speech signal vector of length N. The clean speech signal vector x(k,n) and the noisy signal vector v(k,n) are defined similarly to the vector y(k,n) in the above formula, i.e., x(k,n)=[X(k,n)X(k,n-1)…X(k,n-N+1)] T ,v(k,n)=[V(k,n)V(k,n-1)…V(k,n-N+1)] T , superscript (·) T This indicates transpose.

[0085] The specific method for step 3 is as follows:

[0086] The correlation matrix Φ of the noise signal vector v(k,n) is estimated using existing noise estimation algorithms. v (k,n), the correlation matrix Φ of the noisy speech signal vector y(k,n) is estimated by a recursive method. y (k,n):

[0087] Φ y (k,n)=α y Φ y (k,n-1)+(1-α y )y(k,n)y H (k,n)

[0088] Where α y Forgetting factor (0 < α) y <1), superscript (·) HIndicates conjugate transpose; since the speech signal and noise signal are independent, through Φ x (k,n)=Φ y (k,n)-Φ v Estimate the correlation matrix Φ of the speech signal vector x(k,n) (k,n). x (k,n); through Φ x The correlation coefficient vector ρ can be obtained from (k,n). X (k,n)=[ρ X (k,n) ρ X (k,n-1) … ρ X (k, n-N+1)] T (matrix Φ) x The first column of (k,n) divided by matrix Φ x The element in the first row and first column of (k,n) is the vector ρ. X (k,n)); Φ y (k,n) is the correlation matrix of the noisy speech signal vector y(k,n), where k is the frequency point, n is the frame, and α y Let y(k,n) be the forgetting factor, and y(k,n) be the noisy speech signal vector.

[0089] The specific method for step 4 is as follows:

[0090] In this invention, X(k,n) is set as the desired signal, meaning the goal of noise reduction is to estimate the signal X(k,n) from the noisy signal vector y(k,n), rather than estimating the entire speech signal vector x(k,n). Therefore, in this invention, the vector x(k,n) is decomposed into two orthogonal components: a component related to X(k,n) and an interference component.

[0091]

[0092] in, x i (k,n)=[X i (k,n) X i (k,n-1) … X i (k, n-N+1)] T =x(k,n)-x d (k,n),

[0093] In the above formula, the superscript (·) * Denotes complex conjugation, x d (k,n) is the signal vector associated with the desired signal X(k,n), x i (k,n) is the interference signal vector, and vector ρ X (k,n) represents the normalized inter-frame correlation coefficient vector between X(k,n) and the vector x(k,n), where xd (k,n) is the signal vector associated with the desired signal X(k,n), x i (k,n) is the interference signal vector, and N is the maximum length of the variable length filter.

[0094] In traditional methods, an FIR filter h(k,n) of length N is typically used to denoise the noisy signal vector y(k,n) in order to estimate the desired signal X(k,n), i.e.:

[0095]

[0096] Where Z(k,n) represents the estimated value of the desired signal X(k,n). X represents the filtered speech signal. ri (k,n)=h H (k,n)x i (k,n) represents the filtered interference signal, V rn (k,n)=h H (k,n)v(k,n) represents residual noise, with the superscript (·). H Let h(k,n) denote the conjugate transpose, and be an FIR filter of length N. h(k,n) is defined as:

[0097] h(k,n)=[H(k,n) H(k,n-1) … H(k,n-N+1)] T

[0098] vector ρ X (k,n) represents the inter-frame correlation between the current frame and the past 1 to N-1 frames at the k-th frequency point. Extensive experiments have shown that the vector ρ... X The magnitudes of each element in (k,n) gradually decrease as the distance from the current frame increases (although there may be fluctuations at longer distances, the trend remains unchanged, which also indicates that the correlation between the current frame and distant speech frames is low), that is:

[0099] |ρ X (k,n)|≥|ρ X (k,n-1)|≥…≥|ρ X (k, n-N+1)|

[0100] Let there be a freely set threshold δ, where 0 ≤ δ ≤ 1. When |ρ X When (k,nl)|≥δ(0≤l≤N-1), it can be considered that X(k,n) contains more effective information than interference information when estimating X(k,n). Otherwise, it is considered that X(k,nl) contains less effective information for estimating X(k,n) than interference information. Therefore, the filter length can be determined based on |ρ XThe comparison relationship between (k,nl)| and δ determines the outcome, i.e., if:

[0101] |ρ X (k,n)|≥|ρ X (k,n-1)|≥…≥|ρ X (k,n-P+1)|≥δ>|ρ X (k,nP)|≥…≥|ρ X (k, n-N+1)|

[0102] Then it is acceptable for:

[0103]

[0104] The length of the variable-length filter is determined to be P.

[0105] In the formula, 1 ≤ P ≤ N, meaning N is the maximum length of the variable-length filter, where (·) indicates that the vector length is variable. When the parameter δ = 0, P = N, which is the case for a traditional fixed-length filter. Based on this, we can determine the vector... The length P of the vector determines the vector and The length of is variable between 1 and N. Based on this, the noise reduction process can be rewritten as:

[0106]

[0107] In the formula, This represents an estimate of the clean speech signal X(k,n). This represents the filtered speech signal. This represents the filtered interference signal. Indicates residual noise. Let P be a variable-length correlation coefficient vector. It is a clean speech signal vector of variable length. Let X(k,n) be a variable-length noise signal vector. It can be seen that the distortion-free condition for the desired signal X(k,n) is... in, This represents a variable-length noise reduction filter, (·) H This indicates the conjugate transpose, (·) * This represents complex conjugate. Clearly, the three filtered parts are uncorrelated. Therefore, The variance is:

[0108]

[0109] In the formula:

[0110]

[0111]

[0112]

[0113] in, They represent variance Representing vectors The correlation matrix of length P has a rank of 1 and a dimension of P×P, φX(k,n)=E[|X(k,n)| 2 E[·] represents the mathematical expectation. Representing vectors Correlation matrix ( Represents a variable-length vector (any one of them), with a size of P×P, in this invention it is assumed It is a full-rank matrix.

[0114] To derive the optimal variable-length filter, the desired signal estimate is... The subband mean square error (MSE) is defined as:

[0115]

[0116] The MSE in the formula can be further decomposed into the sum of the MSE based on subband speech distortion and the MSE of residual interference and noise:

[0117]

[0118] In the formula, For MSE based on speech distortion, The residual interference and noise after filtering is represented by the MSE, defined as:

[0119]

[0120] Wherein, φX(k,n), and Let X(k,n) and the filtered interference signal be represented respectively. and residual noise signal The variance.

[0121] Pair In The variable-length Wiener filter can be derived by taking the derivative and setting the result to zero.

[0122]

[0123] in, It is a variable-length Wiener filter. For matrix Take the correlation matrix Φ of the noisy speech signal y The matrix can be obtained by taking the first P rows and first P columns of (k,n). The inverse matrix of ).

[0124] In the formula Can be written as

[0125]

[0126] in, It represents the correlation matrix between the interference signal and the noise.

[0127] right Applying the matrix inverse lemma, It can be represented as:

[0128]

[0129] in, For matrix Φ in The inverse matrix of (k,n).

[0130] Under distortion-free constraints Below, minimize the output energy of the filter. This allows for the derivation of the Minimum Variance Distortionless Response (MVDR) filter. The above process can be represented as:

[0131]

[0132] Here, "st" is an abbreviation for "subject to".

[0133] By solving the above equation, we can obtain:

[0134]

[0135] Typically, a tradeoff filter can better control the balance between output signal-to-noise ratio and speech distortion. It can be achieved by minimizing speech distortion under the constraint that the residual noise level is less than the input noise level. This process can be expressed as:

[0136]

[0137] In the formula, φV(k,n)=E[V(k,n) 2], 0 < α < 1, by solving the above problem, the variable-length compromise filter can be derived:

[0138]

[0139] In the formula, μ (μ≥0) is a freely set parameter. In practical applications, a suitable value of μ is usually selected according to the following principles. When μ=1, Equivalent to a variable-length Wiener filter When 0 < μ < 1, The output signal-to-noise ratio and speech distortion are both lower than those of the variable-length Wiener filter. When μ > 1, The output signal-to-noise ratio and speech distortion are both higher than those of the variable-length Wiener filter.

[0140] The specific method for step 5 is as follows:

[0141] Based on the filter estimated in step 4, it can be used... Obtain an estimate of the clean speech signal. in For variable length Wiener filters Variable-length minimum variance distortionless response filter and variable length compromise filter Any one of them.

[0142] The specific method for step 6 is as follows:

[0143] The clean speech signal estimate obtained in step 5 can be obtained by using inverse Fourier transform and overlapping addition or overlapping preservation methods. Convert to the time domain.

[0144] To better illustrate the effects of this invention, the following specific examples are provided to verify the effectiveness of the proposed method. In the simulation, the clean speech signal uses corpus from the TIMIT database, and the noise uses white noise (generated by a computer).

[0145] like Figure 2 The performance of the variable-length Wiener filter varies with parameter δ when the input signal-to-noise ratio is different. In the simulation, the maximum filter length N is set to 8. Figure 2 This paper presents the variation of the performance of a variable-length Wiener filter with parameter δ under different input signal-to-noise ratios (iSNR) in a white noise field. Figure 2 As can be seen, under different iSNR conditions, the output signal-to-noise ratio first increases and then decreases with δ. Figure 2a), while the speech distortion coefficient monotonically increases with δ ( Figure 2 (b) When the parameter δ is small, the perceptual evaluation of speech quality (PESQ) score does not change with the increase of parameter δ; after δ reaches a certain value, the PESQ score gradually decreases with the increase of δ. Figure 2 c). Simulation results show that when the value of parameter δ is appropriate (the situation differs depending on the iSNR), shortening the filter length does not affect the speech quality of the output speech signal. However, a shorter filter length means lower complexity in the noise reduction algorithm.

[0146] In summary, when the parameter δ is set appropriately, this invention can reduce computational complexity while maintaining the output speech quality of a traditional fixed-length filter (i.e., when δ = 0).

[0147] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A single-channel frequency domain speech denoising method based on a variable-length filter, characterized in that, include: Acquire noisy speech signals and preprocess the noisy speech signals; Preprocessing the noisy speech signal includes: performing frame segmentation, windowing, and Fourier transform on the noisy speech signal, and introducing inter-frame correlation. The method for preprocessing the noisy speech signal is as follows: In speech denoising, the time-domain signal model is represented as: in, Representing discrete time points, and Representing the desired signal and additive noise, respectively. Indicates noisy speech signals; After framing, windowing, and Fourier transform, the short-time Fourier transform domain signal model is as follows: in, , , They represent , and In the Frame number Short-time Fourier transform coefficients at each frequency point; Introducing the inter-frame correlation includes: Will A series of consecutive time frames combined together: in, Indicates length is The noisy speech signal vector, For clean speech signal vectors, The noise signal vector, with superscript Indicates transpose. The desired signal is the signal vector; that is, the goal of noise reduction is to estimate the signal from the noisy signal vector. ; Based on the preprocessed noisy speech signal, the statistical characteristics of the noisy speech signal and the noise signal are estimated, and the variable-length speech denoising filter is estimated using the statistical characteristics. The estimated statistical properties of noisy speech signals include: Estimate the correlation matrix of the noise signal vector and the correlation matrix of the noisy speech signal vector, and obtain the correlation matrix of the clean speech signal vector by subtracting the correlation matrix of the noise signal vector from the correlation matrix of the noisy speech signal vector. The method for estimating the correlation matrix of the noisy speech signal vector is as follows: in, For the correlation matrix, Indicates the first One frequency point, Indicates the first frame, This is a noisy speech signal vector. Forgetting factor, It is the conjugate transpose; The method for obtaining the correlation matrix of the clean speech signal vector is as follows: in, To estimate the speech signal vector The correlation matrix, The correlation matrix of the noise signal vector; The estimated variable-length speech noise reduction filter includes: Based on correlation coefficient vector and threshold parameter ,in, Obtain the vector of variable-length correlation coefficients. and vectors Length; The method for obtaining the correlation coefficient vector is as follows: Through the correlation matrix The correlation coefficient vector can then be obtained. Among them, the correlation matrix Divide the first column by the correlation matrix The element located in the first row and first column is the correlation coefficient vector. ; Get vector The method is as follows: Based on the correlation coefficient vector Elements and threshold parameters The comparison results yield the variable-length correlation coefficient vector. : in, A vector of correlation coefficients of variable length. The length of the variable-length correlation coefficient vector. , The transpose symbol, superscript This indicates that the vector length is variable; Set the maximum length threshold for the variable-length speech noise reduction filter based on the vector. length Determine the variable length filter Variable length vector Variable length vector and variable length vectors Length: in, This indicates a variable-length noise reduction filter. For length is The variable-length correlation coefficient vector, It is a clean speech signal vector of variable length. It is a variable-length noise signal vector; The variable-length speech denoising filter is used to perform denoising filtering to obtain an estimate of the clean speech signal.

2. The single-channel frequency domain speech denoising method based on a variable-length filter according to claim 1, characterized in that, Obtaining an estimate of the clean speech signal includes: The variable-length speech noise reduction filter includes: a variable-length Wiener filter, a variable-length minimum variance distortionless response filter, and a variable-length compromise filter; Based on the variable-length Wiener filter, the variable-length minimum variance distortionless response filter, or the variable-length compromise filter, combined with the noisy speech signal vector Obtain an estimate of the clean speech signal. ,in For the variable length Wiener filter The variable-length minimum variance distortionless response filter and the variable length compromise filter Any one of them, and the estimated value of the clean speech signal. Convert to the time domain.

3. The single-channel frequency domain speech denoising method based on a variable-length filter according to claim 2, characterized in that, Estimating the variable-length Wiener filter includes: Based on variable length filter Define the subband mean square error of the expected signal estimate, and decompose the subband mean square error into the subband mean square error based on subband speech distortion and the subband mean square error based on residual interference and noise. Take the derivative and set the result to zero; Obtaining a variable-length Wiener noise reduction filter: in, It is a variable-length Wiener filter. Desired signal variance For matrix The inverse matrix, A variable-length noisy signal vector The correlation matrix is ​​the correlation matrix of the noisy speech signal. The former Before departure A matrix composed of columns, with superscript Indicates complex conjugation.

4. The single-channel frequency domain speech denoising method based on a variable-length filter according to claim 2, characterized in that, The method for estimating the variable-length minimum variance distortionless response filter is as follows: in, It is a variable-length minimum variance distortionless response filter.

5. The single-channel frequency domain speech denoising method based on a variable-length filter according to claim 2, characterized in that, The method for estimating the variable-length compromise filter is as follows: in, For variable length compromise filters, For vectors ( Correlation matrix, ( To estimate the correlation matrix of the interference signal plus the noise signal, These are parameters that can be freely set. .

Images