Residual noise suppression method based on characteristic component filtering

Abstract

The invention relates to the technical field of noise suppression, in particular to a residual noise suppression method based on characteristic component filtering. A residual noise suppression method based on feature component filtering is characterized in that noise is suppressed by increasing correlation of voice frames, and adjacent small feature components are replaced with large feature components or the adjacent small feature components are removed. In most cases, the three methods provided by the invention all have a relatively high SegSNR score, which reflects the strong ability of noise reduction under the condition of a low signal-to-noise ratio; in the aspect of segmentation signal-to-noise ratio scoring, the three methods provided by the invention are obviously improved, and the PESQ score of auditory speech quality is almost the highest.

Description

Technical Field

[0001] The present invention relates to the technical field of noise suppression, and in particular to a residual noise suppression method based on characteristic component filtering. Background Art

[0002] Speech enhancement is an important preprocessing method to improve the clarity of voice communication. Boll first used spectral subtraction, which is achieved by subtracting the estimated noise spectrum from the speech amplitude spectrum of the original noise. Since the estimated noise spectrum is inaccurate, a large amount of music noise will be generated, which greatly reduces its effect. In order to remove music noise as much as possible, SIM et al. explored a short-time amplitude spectrum estimation method based on two parameters and the minimum mean square error criterion. Wiener filtering assumes that both pure speech and additive noise have Gaussian distribution, and filters the noisy speech signal on the basis of MMSE to obtain an enhanced speech signal.

[0003] Subspace-based filtering methods are another attractive nonparametric approach, which considers the noisy speech signal as a Euclidean vector space consisting of two orthogonal vector subspaces, the mixed subspace and the noise-only subspace. The mixed subspace contains speech information and noise, while the noise-only subspace contains only noise. Ephraim and Malah proposed an initial subspace decomposition that is only applicable to the white noise case by removing the noise in the mixed subspace and the noise-only subspace with the help of Kalhun-LoVE transform (KLT) or singular value decomposition (SVD). Rezayee and Gazor extended it to colored noise, which also has a similar noise covariance matrix, but it is suboptimal. Subsequently, Hu and Loizou explored the optimal linear estimation of colored noise, based on the generalized eigenvalue decomposition (GEVD) of the covariance matrix of clean speech and noise. In 2014, Saadoune et al. proposed a new reconstruction error (VRE) criterion to balance noise suppression and speech distortion. Then, Sun et al. tried to recover clean speech signals from noise using the principle of constrained low-rank and sparse matrix decomposition, which relies on the assumption that background noise can be regarded as a low-rank subspace and speech signal can be regarded as a sparse subspace. However, the GEVD-based method only considers noise suppression in the noise subspace and ignores the noise in the mixed subspace, resulting in large residual noise at the same level of speech distortion.

[0004] Therefore, the present invention proposes three different filtering algorithms to suppress the residual noise in the mixed subspace. Summary of the invention

[0005] In order to overcome the deficiencies in the prior art, the present invention proposes three different filtering algorithms to suppress the residual noise in the mixed subspace.

[0006] The residual noise suppression method based on characteristic component filtering suppresses noise by increasing the correlation of speech frames, and is achieved by replacing adjacent small characteristic components with large characteristic components or removing adjacent small characteristic components.

[0007] Preferably, the following three methods are included:

[0008] Method a: After jointly diagonalizing the clean speech and noise covariance matrices, filter the small eigenvalues ​​less than zero in descending order of eigenvalues, and replace the adjacent large eigenvalues ​​with small eigenvalues;

[0009] Method b: Adjacent small eigencomponents and small eigencomponents with eigenvalues ​​less than zero are removed together;

[0010] Method c: replace large eigencomponents with adjacent small eigencomponents and introduce the generalized inverse matrix transform (GIMT) into the algorithm.

[0011] Preferably, the linear estimators corresponding to the three filtering methods can be respectively modified as follows:

[0012] Method a:

[0013] Method b:

[0014] Method c:

[0015] Where Λ1, Λ2, Λ3 represent eigenvalue matrices, and V1, V2, V3 represent the corresponding eigenvector matrices.

[0016] Preferably, after descending order of eigenvalues, the similarity between adjacent eigenvalues ​​can be expressed as Y = |λ i -λ i-1 |(i>1), where λ i is the matrix R s and R n is the generalized eigenvalue of , where i is the eigenvalue index.

[0017] Preferably, the elements of Λ1, Λ2, and Λ3 are all above zero and not less than a predetermined preset threshold γ.

[0018] Preferably, the generalized inverse matrix transform (GIMT) is introduced into the algorithm to prove the existence and uniqueness of the Moore-Penrose inverse matrix transform of a random matrix.

[0019] Preferably, γ is a balance parameter between speech distortion and residual noise. In the traditional subspace-based method, the signal-to-noise ratio Where L is the speech frame length; γ is related to the signal-to-noise ratio of the noisy speech signal.

[0020]

[0021] Among them, -5 to -0.5 represents a low signal-to-noise ratio condition, and -0.5 to 7 represents a high signal-to-noise ratio environment.

[0022] Preferably, among the objective indicators considered, PESQ is a perceptual speech quality evaluation method for mobile phones and narrowband speech codecs recommended by ITU-T, which is calculated as follows: a linear combination of the average interference value D and the average asymmetric interference to obtain the impact of γ on the segmented SNR, PESQ, and the total score;

[0023] PESQ=a0+a1D+a2A

[0024] a0=4.5

[0025] a1=-0.1

[0026] a2=-0.0309

[0027] Another accurate assessment of background distortion (noise suppression) is the SEGSNR measurement, which is defined as:

[0028]

[0029] where s p and represents the signal frames of clean speech and enhanced speech in the time domain, and p is the index representing all clean speech frames.

[0030] The present invention has the following advantages: 1. In most cases, the three methods proposed in the present invention have higher segSNR scores, which reflects the powerful ability of noise reduction under low signal-to-noise ratio conditions.

[0031] 2. In terms of segmental SNR scores, the three proposed methods all show significant improvements, while the PESQ score of auditory speech quality is almost the highest.

[0032] 3. Compared with the GEVD-based method, methods a, b, and c have fewer glitches in the time domain and are more similar to clean speech in the frequency domain, showing better enhancement performance. Compared with other methods, this method has obvious advantages in noise suppression and speech distortion. In short, this method is more suitable for speech enhancement under low signal-to-noise ratio conditions.

[0033] 4. This paper proposes three speech enhancement methods based on feature component filtering. Through the influence of the number of feature components on speech enhancement and the influence of frame correlation, and the introduction of GIMT to solve the irreversible problem of the residue eigenvector matrix, experiments are carried out to make it more general. It is proved that this method can further suppress noise and maintain good speech quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a graph showing how the PESQ and segsrn scores vary with the number of eigenvalues ​​M according to the present invention.

[0035] Figure 2 This is a correlation comparison diagram between clean speech and noisy speech of the present invention.

[0036] Figure 3 This is a feature component filtering block diagram for speech enhancement of the present invention.

[0037] Figure 4 This is a diagram showing the influence of γ on segmented SNR, PESQ, and total score of the present invention.

[0038] Figure 5 This is a comparison diagram of the correlation between adjacent speech frames of the GEVD method of the present invention and the c method.

[0039] Figure 6 This is the average score diagram of PESQ and segSNR under different SNRS and methods of the present invention.

[0040] Figure 7 Spectrum diagram and waveform comparison diagram of different methods of speech enhancement of the present invention.

[0041] Figure 8 The figure is a performance score diagram of various methods and environments of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described below in conjunction with specific examples. These examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0043] Example 1

[0044] Residual noise suppression methods based on characteristic component filtering, such as Figure 1-2 As shown, the perceptual evaluation of speech quality (PESQ) (ITU2001) and segmental signal-to-noise ratio (segSNR) scores increase with Λ GIt changes with the increase of the number of eigenvalues. Obviously, as the number of eigenvalues ​​increases, the scores of PESQ and segSNR tend to be higher, where PESQ and segSNR represent speech distortion and residual noise, respectively. After reaching a maximum point, the scores of PESQ and SEGSRN decrease slightly or remain unchanged. This (Sun and Mu 2015) motivates us to select an appropriate number of main feature components to obtain higher PESQ and SEGSRN scores.

[0045] Three eigencomponent filtering methods were explored to further compress the residual noise while maintaining good speech distortion. Method a jointly diagonalizes the clean speech and noise covariance matrices, filters small eigencomponents with eigenvalues ​​less than zero in descending order of eigenvalue size, and replaces adjacent small eigencomponents with large eigencomponents; method b only calculates adjacent small eigencomponents and small eigencomponents with eigenvalues ​​less than zero; method c only calculates adjacent small eigencomponents. The speech signal is then restored using the filtered eigencomponents.

[0046] like Figure 3 As shown in the figure, the noisy speech signal is separated from the frame and window, converted into an observation matrix, and then filtered using methods a, b, and c. The reason for introducing GIMT is that the residual feature vector matrix after filtering is irreversible. Then, the implementation of superposition is to reorganize the frame-based speech fragments into a continuously enhanced speech signal. The dotted line surrounds the main work of this paper.

[0047] The linear estimators corresponding to the three filtering methods can be corrected as follows:

[0048]

[0049] Wherein Λ1, Λ2, Λ3 represent eigenvalue matrices, whose elements are all above zero and not less than a preset threshold γ determined later, and V1, V2, V3 represent corresponding eigenvector matrices.

[0050] Example 2

[0051] It is proved that the Moore-Penrose inverse matrix transformation (Zhibin 2010) of a random matrix is ​​unique.

[0052] Theorem: If A∈m×n matrix, B∈n×m matrix, the following equation is satisfied

[0053]

[0054] We call B the Moore-Penrose inverse matrix of A. Assume that both B and C satisfy the above equation

[0055] B=BAB=A T B T B=(AT C T )A T B T

[0056] Equation (1): B = (CA) A H B H B=CABAB=CAB

[0057] Equation (2): C = C (AC) = CC H A H =(CC H )A H B H A H =CAB

[0058] According to equations (1) and (2), we can conclude that B is equal to C, which proves the uniqueness of the Moore-Penrose inverse matrix. In addition, according to (Adukov 1998), we can also prove the existence of the Moore-Penrose inverse matrix of a random matrix.

[0059] As far as we know, the introduction of GIMT makes subspace-based methods more general. The linear estimate can be redefined as

[0060] H opt ≈UGU T

[0061] G=Δ x (Δ x +μΔ n ) -1

[0062] Where G is the gain matrix, here the eigenvector matrix U is orthogonal, the noisy speech is projected to the orthogonal direction, and after linear weighting G, U can be restored to its original form.

[0063] The final estimate can be expressed as

[0064]

[0065] G=Λ g (Λ g +μI) -1

[0066] Where i=1…M, i=(M+1)…1 [Rank(V)] is the subscript, and G is the gain matrix. is a matrix The orthogonal part of yes The non-orthogonal part of .

[0067] It should be emphasized that due to Non-orthogonality, noisy speech is partially projected to non-orthogonal directions for linear weighting and corresponding non-orthogonal restoration, which may cause some residual noise and speech distortion. In this method, by inputting highly correlated feature components (similar eigenvalue amplitudes), this can be largely avoided.

[0068] Example 3

[0069] The parameter γ is estimated by sorting the eigenvalues ​​in descending order. The similarity between adjacent eigenvalues ​​can be expressed as γ = |λ i -λ i-1 |(i>1), where λ i is the matrix R s and R n is the generalized eigenvalue of , where i is the eigenvalue index.

[0070] For example, γ=0.01, if |λ i -λ i-1 |<0.01, we perform corresponding eigencomponent filtering; otherwise we do nothing.

[0071] In fact, γ is a balance parameter between speech distortion and residual noise. If γ is too small, there will be a lot of residual noise in the enhanced speech signal. On the contrary, if γ is too large, the enhanced speech signal will have serious speech distortion. In the traditional subspace-based method, the signal-to-noise ratio Where L is the speech frame length. In fact, γ is related to the signal-to-noise ratio estimation of the noisy speech signal.

[0072]

[0073] Among them, -5 to -0.5 represents a low signal-to-noise ratio condition, and -0.5 to 7 represents a high signal-to-noise ratio environment.

[0074] Example 4

[0075] The enhanced speech signals are evaluated on segmented sentences (sp01-sp30) from the NOISEX-92 (Sun and Mu 2015) database. The sampling frequency of each test speech is 8 kHz and the sampling length is 21000. Rectangular windows with 50% overlap are used in these experiments.

[0076] like Figure 4 As shown, among the objective indicators considered, PESQ is a perceptual speech quality assessment method recommended by ITU-T for mobile phones and narrowband speech codecs. Its calculation method is as follows: a linear combination of the average interference value D and the average asymmetric interference.

[0077] PESQ=a0+a1D+a2A

[0078] a0=4.5

[0079] a1=-0.1

[0080] a2=-0.0309

[0081] Another accurate assessment of background distortion (noise suppression) is the SEGSNR measurement, which is defined as:

[0082]

[0083] where s p and represents the signal frames of clean speech and enhanced speech in the time domain, and p is the index representing all clean speech frames.

[0084] To further improve the reliability of the assessment, a composite target measure using a nonlinear and nonparametric regression model (Hu and Loizou 2008) was also used, which showed higher correlations with subjective speech and global distortion than traditional target measures.

[0085] Overall=1.594+0.805×PESQ-0.512×LLR-0.007×WSS

[0086] like Figure 4 As shown in the figure, the parameter γ is a flexible and adjustable parameter, which directly determines the enhancement performance of the proposed method. segsrn and the total score increase with the increase of γ until they reach certain peaks and then continue to decrease, which indicates that the enhancement performance will also encounter similar changes. At the same time, the PESQ score almost decreases with the increase of γ. Therefore, it is a meaningful issue to choose a suitable γ to obtain good perceived speech quality and less residual noise.

[0087] like Figure 5 As shown in Figure 2, the correlation between adjacent speech frames of method c and the GEVD-based method is compared. By comparing, it can be found that method c has obvious improvements. Figure 2 As shown in Figure 2, the introduction of noise directly leads to a decrease in the correlation between adjacent speech frames. Compared with the GEVD-based method, method c further improves the correlation between adjacent speech frames, making the inter-frame correlation close to that of clean speech.

[0088] like Figure 6As shown in Figure 2, the PESQ and segSNR average scores for different SNRs and methods are shown. In terms of segmental SNR scores, the three proposed methods have obvious improvements, and the PESQ score of auditory speech quality is almost the highest. The results show that the characteristic component is the basic component of the noisy speech signal, and the appropriate characteristic component is selected to reconstruct the speech signal from the two aspects of noise removal and speech intelligibility.

[0089] like Figure 7 As shown in the figure, the waveforms and spectrograms of the three proposed methods and the other seven methods are shown. Compared with the GEVD-based method, methods a, b, and c have fewer glitches in the time domain and are more similar to clean speech in the frequency domain, showing better enhancement performance. Compared with other methods, this method has obvious advantages in both noise suppression and speech distortion. In short, this method is more suitable for speech enhancement under low signal-to-noise ratio conditions.

[0090] Table 8 shows the PESQ and segSNR scores under different noise environments from -5dB to 0dB. Bold and italic indicate the highest and second highest PESQ or segSNR scores. Obviously, in most cases, the three proposed methods have higher segSNR scores, which reflects the strong ability of noise reduction under low SNR conditions.

[0091] The above embodiments and comparative examples describe the preferred embodiments of the present invention in detail, but the present invention is not limited thereto. Within the technical concept of the present invention, the technical solution of the present invention can be subjected to a variety of simple modifications, including the combination of various technical features in any other suitable manner, and these simple modifications and combinations should also be regarded as the contents disclosed by the present invention and belong to the protection scope of the present invention.

Claims

1. A residual noise suppression method based on characteristic component filtering, characterized in that: The noise is suppressed by increasing the correlation of speech frames, replacing adjacent small feature components with large feature components or removing adjacent small feature components.

2. The residual noise suppression method based on characteristic component filtering according to claim 1 is characterized in that it includes the following three methods: Method a: After jointly diagonalizing the clean speech and noise covariance matrices, filter the small eigenvalues ​​less than zero in descending order of eigenvalues, and replace the adjacent large eigenvalues ​​with small eigenvalues; Method b: Adjacent small eigencomponents and small eigencomponents with eigenvalues ​​less than zero are removed together; Method c: replace large eigencomponents with adjacent small eigencomponents and introduce the generalized inverse matrix transform (GIMT) into the algorithm.

3. The residual noise suppression method based on characteristic component filtering according to claim 2 is characterized in that the linear estimators corresponding to the three filtering methods can be respectively corrected as: Method a: Method b: Method c: Where Λ1, Λ2, Λ3 represent eigenvalue matrices, and V1, V2, V3 represent the corresponding eigenvector matrices.

4. The residual noise suppression method based on characteristic component filtering according to claim 3 is characterized in that, after the eigenvalues ​​are sorted in descending order, the similarity between adjacent eigenvalues ​​can be expressed as γ = |λ i -λ i-1 |(i>1), where λ i is the matrix R s and R n is the generalized eigenvalue of , where i is the eigenvalue index.

5. The residual noise suppression method based on characteristic component filtering according to claim 4 is characterized in that the elements of Λ1, Λ2, and Λ3 are all above zero and not less than a determined preset threshold γ.

6. The residual noise suppression method based on characteristic component filtering according to claim 2 is characterized in that a generalized inverse matrix transform (GIMT) is introduced into the algorithm to prove the existence and uniqueness of the Moore-Penrose inverse matrix transform of a random matrix.

7. The residual noise suppression method based on characteristic component filtering according to claim 6 is characterized in that γ is a balance parameter between speech distortion and residual noise; in the traditional subspace-based method, the signal-to-noise ratio Where L is the speech frame length; γ is related to the signal-to-noise ratio of the noisy speech signal. in, -5 to -0.5 indicates a low signal-to-noise ratio condition, and -0.5 to 7 indicates a high signal-to-noise ratio environment.

8. The residual noise suppression method based on characteristic component filtering according to claim 7 is characterized in that, among the objective indicators considered, PESQ is a perceptual speech quality evaluation method for mobile phones and narrowband speech codecs recommended by ITU-T, and its calculation method is as follows: a linear combination of the average interference value D and the average asymmetric interference is used to obtain the influence of γ on the segmented SNR, PESQ, and the total score; PESQ=a0+a1D+a2A a0＝4.5 a1＝-0.1 a2＝-0.0309 Another accurate assessment of background distortion (noise suppression) is the SEGSNR measurement, which is defined as: where sx and represents the signal frames of clean speech and enhanced speech in the time domain, and p is the index representing all clean speech frames.

Images