A method for acoustic signal pattern classification based on eigenvalue clustering

Abstract

A method for acoustic signal pattern classification based on eigenvalue clustering belongs to the technical field of acoustic signal pattern classification. It includes the following steps: S1, using a fixed time length to divide the time domain sound signal into continuous frame segments, and using Fourier transform to convert each segment into a spectrum segment; S2, using the similarity / dissimilarity metrics between the vectors of the spectrum segments to construct a similarity / dissimilarity metric matrix; S3, performing spectral decomposition on the metric matrix to obtain a spectral matrix consisting of an eigenvector matrix P and eigenvalues; S4, performing k-means clustering on the spectral eigenvalue sequence to obtain k groups of results; S5, mapping the eigenvalue clustering results to a set of spectrum segments. The present invention directly uses eigenvalues ​​for k-means clustering, and the clustering vector is a one-dimensional scalar, so there is no need to use PCA technology for dimensionality reduction processing, thereby achieving optimal computational efficiency. The method is suitable for technical fields such as acoustic noise reduction, DTD detection of acoustic echoes, echo separation, and sound source localization.

Description

Technical Field

[0001] The present invention belongs to the technical field of acoustic signal pattern classification, and in particular relates to an acoustic signal pattern classification method based on eigenvalue clustering. Background Art

[0002] Clustering is an unsupervised pattern classification technique that generally does not require manual data labeling, but does require a pre-determined number of cluster types. Among various clustering techniques, spectral clustering has seen rapid growth in recent years. Initially, it is somewhat similar to traditional hierarchical clustering: a mutual similarity measure (or mutual dissimilarity measure) is constructed between all participating data. This measure is an N×N symmetric matrix, where N is the number of original data points participating in the cluster. The spectral decomposition of this matrix is ​​then calculated, and the eigenvector matrix (i.e., the similarity matrix) is normalized (Schmidt orthogonalized and normalized). Dimensionality reduction is then performed based on traditional PCA statistical analysis. Based on the distribution of eigenvalues, the eigenvector matrix corresponding to a subset of eigenvalues ​​(accounting for more than 80% of the total) is clustered. The number of rows in this eigenvector matrix remains the same as the number of original data points participating in the cluster, but the number of columns is now less than N. This is the result of PCA dimensionality reduction. Perform k-means clustering on the row vectors of all partial eigenvector matrices, and finally correspond the clustering results to the original data to obtain the classification of the original data that is actually needed.

[0003] A drawback of spectral clustering is that when the number of raw data points involved in clustering is large (i.e., N is very large), even after PCA dimensionality reduction, the dimensions of the rows of the partial eigenvector matrix participating in k-means clustering are still sufficiently large, which undoubtedly imposes a burden on k-means clustering. Furthermore, how should the clustering of the rows of the partial eigenvector matrix be mapped to the classification of the raw data? It is generally assumed that the number of rows of the partial eigenvector matrix is ​​the same as the number of raw data points involved in clustering, and the first row is directly mapped to the first raw data point, the second row to the second raw data point, and so on.

[0004] Acoustic signal type identification is a key issue in all types of acoustic engineering solutions. For example, in the field of acoustic noise reduction, both active and passive noise reduction involve distinguishing between noise and effective acoustic signals. For single-channel blind noise reduction (a passive noise reduction mode), it is necessary to distinguish between background noise and foreground sound based on signal spectrum analysis; for multi-channel passive noise reduction, it is necessary to distinguish the direction of the acoustic bins and track them; for multi-channel active noise reduction, it is necessary to identify background noise and foreground sound, and then generate background sound with the opposite phase.

[0005] In acoustic echo cancellation, using signal separation technology to achieve echo cancellation or DTD voice duplex detection requires identifying and distinguishing near-end and far-end signals. Acoustic signal localization requires distinguishing source signals from interference signals in different locations. The cocktail effect requires separating human voices from musical sounds, male and female voices, or speech in different languages, similarly requiring different signal identification. These are all examples of using different technologies to achieve acoustic signal recognition. Summary of the Invention

[0006] In response to the above-mentioned problems existing in the prior art, the purpose of the present invention is to provide an acoustic signal pattern classification method based on eigenvalue clustering, which can reduce the dimension of the clustering vector and does not require the use of PCA technology for dimensionality reduction processing, thereby achieving optimal computational efficiency.

[0007] The present invention provides the following technical solutions:

[0008] An acoustic signal pattern classification method based on eigenvalue clustering comprises the following steps:

[0009] S1. Use a fixed duration to divide the time domain sound signal into continuous frame segments, and use Fourier transform to convert each segment into a spectrum segment;

[0010] S2. constructing a similarity / dissimilarity measure matrix using the similarity / dissimilarity measures between the vectors of the spectrum segments;

[0011] S3, performing spectral decomposition on the metric matrix to obtain a spectral matrix consisting of an eigenvector matrix P and eigenvalues;

[0012] S4, performing k-means clustering on the spectral eigenvalue sequence to obtain k groups of results;

[0013] S5. Map the result of eigenvalue clustering to a set of spectrum segments.

[0014] Furthermore, in step S1, the spectrum of each segment is the original data to be clustered, which is a vector with a dimension of d. The original data is recorded as a vector group, which is represented as follows:

[0015]

[0016] Where N is the number of fragments, that is, the number of original data.

[0017] Furthermore, in step S2, the matrix is ​​a real symmetric matrix; the dissimilarity measure adopts the Euclidean distance, and the similarity measure adopts the correlation coefficient, the Peasson correlation coefficient or a decreasing function of the Euclidean distance.

[0018] Furthermore, the specific process of step S2 is as follows:

[0019] Select data from the original data and The data and The similarity or dissimilarity measure is denoted as:

[0020]

[0021] in, and

[0022] The measurement matrix is ​​expressed as follows:

[0023]

[0024] Furthermore, in step S3, the spectrum matrix is ​​expressed as follows:

[0025]

[0026] The relationship between the spectral matrix and the metric matrix is ​​as follows:

[0027]

[0028] Furthermore, the specific process of step S4 is as follows:

[0029] For the spectral eigenvalue sequence {λ1,λ2,…,λ N} Perform k-means clustering to obtain k groups of results,

[0030]

[0031]

[0032] …

[0033]

[0034] in, n1+n2+...+n k =N.

[0035] Furthermore, the specific process of step S5 is as follows:

[0036] S5.1 linearly interpolates the vector of the original data so that its dimension reaches the least common multiple of the original dimension and N;

[0037] S5.2 performs inner product analysis on a certain original data with each eigenvector corresponding to each category in the k categories, takes the average projection value of each category, and assigns the original data to the category with the largest average projection value; thus, the mapping of the eigenvalue clustering results to the classification of the original data is completed.

[0038] By adopting the above technology, compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1) The present invention directly uses eigenvalues ​​for k-means clustering. The clustering vector is a one-dimensional scalar, and there is no need to use PCA technology to reduce the dimensionality, thereby achieving optimal computational efficiency. It is applicable to technical fields such as acoustic noise reduction, DTD detection of acoustic echoes, echo separation, and sound source localization.

[0040] 2) In the present invention, the original data is interpolated to make its dimension consistent with the dimension of the feature vector, that is, a sorting method of projecting onto the feature vector is adopted to correspond the specific situation of the eigenvalue clustering to the original data. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 2 is a system block diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0043] On the contrary, the present invention covers any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention as defined by the claims. Furthermore, to facilitate a better understanding of the present invention, certain specific details are described in detail below in the detailed description of the present invention. Those skilled in the art will be able to fully understand the present invention without these details.

[0044] See also Figure 1 This embodiment differs from traditional spectral clustering in that it does not require k-means clustering of the row vectors of the eigenvector matrix, but instead directly uses the eigenvalues ​​for k-means clustering. This has the advantage of making the clustering vectors very low-dimensional, directly a one-dimensional scalar, eliminating the need for PCA techniques for dimensionality reduction, thus achieving optimal computational efficiency. Furthermore, the specific circumstances of eigenvalue clustering, corresponding to the original data, can be determined by using a sorting method that projects onto the eigenvectors.

[0045] As in traditional spectral clustering, the dimension of the original data is inconsistent with the dimension of the row vector of the eigenvalue matrix; the latter is an integer less than N after dimensionality reduction in traditional spectral clustering, while in this case it is N; therefore, when considering the inner product projection of the original data and the eigenvector, the original data must first be interpolated to make its dimension consistent with the dimension of the eigenvector.

[0046] The specific steps are as follows:

[0047] Step S1: Get original data:

[0048]

[0049] Where N=4,

[0050] Step S1: Use the correlation coefficient to construct a symmetric matrix as follows:

[0051]

[0052] The sequence of 4 eigenvalues ​​is {2, 0, 2, 0}; the eigenvector corresponding to eigenvalue 2 is {(1, 0, 1, 0) T , (0, 1, 0, 1) R}; The eigenvector corresponding to eigenvalue 0 is {(-1, 0, 1, 0) R , (0, -1, 0, 1) R};

[0053] Perform k-means clustering on the eigenvalues ​​{2, 0, 2, 0}, with k = 2, and the clustering is {2, 2} as one class and {0, 0} as another class. Now classify the original data α1 and α2;

[0054] α1 and the four eigenvectors are interpolated to 12 dimensions (the least common multiple of 4 and 3 is 12), and the inner products are performed, such as:

[0055] (1,1,1,1,-1,-1,-1,-1,1,1,1,1))·(1,1,1,0,0,0,1,1,1,0,0,0) T =2

[0056] (1,1,1,1,-1,-1,-1,-1,1,1,1,1)·(0,0,0,1,1,1,0,0,0,1,1,1) T =2

[0057] The above is the inner product value (projection) of the two eigenvectors of α1 on the eigenvalue 2, with an average value of 2.

[0058] (1,1,1,1,-1,-1,-1,-1,1,1,1,1)·(-1,-1,-1,0,0,0,1,1,1,0,0,0) T =-4

[0059] (1,1,1,1,-1,-1,-1,-1,1,1,1,1)·(0,0,0,-1,-1,-1,0,0,0,1,1,1) T =4

[0060] The above is the inner product value (projection) of α1 on the two eigenvectors of eigenvalue 0, and the average value is 0.

[0061] Therefore, α1 is classified as the category corresponding to eigenvalue 2.

[0062] Interpolate α2 and the four eigenvectors to 12 dimensions (the least common multiple of 4 and 3 is 12), and perform inner products, such as:

[0063] ((-1,-1,-1,-1,1,1,1,1,2,2,2,2)·(1,1,1,0,0,0,1,1,1,0,0,0) T =1

[0064] (-1,-1,-1,-1,1,1,1,1,2,2,2,2)·(0,0,0,1,1,1,0,0,0,1,1,1) T =7

[0065] The above is the inner product value (projection) of α2 on the two eigenvectors of eigenvalue 2, with an average value of 4.

[0066] (-1, -1, -1, -1, 1, 1, 1, 1, 2, 2, 2, 2)·(-1, -1, -1, 0, 0, 0, 1, 1, 1, 0, 0, 0) T =7

[0067] (-1, -1, -1, -1, 1, 1, 1, 1, 2, 2, 2, 2)·(0, 0, 0, -1, -1, 0, 0, 0, 1, 1, 1) T =5

[0068] The above is the inner product value (projection) of α2 on the two eigenvectors of eigenvalue 0, with an average value of 6.

[0069] Therefore, α2 is classified as the category corresponding to the eigenvalue 0.

[0070] Similarly, α3 is classified into the category corresponding to eigenvalue 2; α4 is classified into the category corresponding to eigenvalue 0.

[0071] This completes the entire eigenvalue clustering process.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for acoustic signal pattern classification based on eigenvalue clustering, characterized by: The following steps are involved: S1. Use a fixed duration to divide the time domain sound signal into continuous frame segments, and use Fourier transform to convert each segment into a spectrum segment; S2. constructing a similarity / dissimilarity measure matrix using the similarity / dissimilarity measures between the vectors of the spectrum segments; S3, performing spectral decomposition on the metric matrix to obtain a spectral matrix consisting of an eigenvector matrix P and eigenvalues; S4, performing k-means clustering on the spectral eigenvalue sequence to obtain k groups of results; S5, mapping the result of eigenvalue clustering to a set of spectrum segments; the specific process of step S5 is as follows: S5.1 Perform linear interpolation on the vector of the original data so that its dimension reaches the least common multiple of the original dimension and N; S5.2 Take the inner product of the original data with each eigenvector corresponding to each of the k categories, take the average projection value of each category, and assign the original data to the category with the largest average projection value; thus, the result of eigenvalue clustering is mapped to the classification of the original data.

2. The acoustic signal pattern classification method based on eigenvalue clustering according to claim 1 is characterized in that In step S1, the spectrum of each segment is the original data to be clustered, which is a vector with a dimension of d. The original data is recorded as a vector group, which is represented as follows: ； Where N is the number of fragments, that is, the number of original data.

3. The acoustic signal pattern classification method based on eigenvalue clustering according to claim 2 is characterized in that In step S2, the matrix is ​​a real symmetric matrix; the dissimilarity measure adopts the Euclidean distance, and the similarity measure adopts the correlation coefficient, the Peasson correlation coefficient or a decreasing function of the Euclidean distance.

4. The acoustic signal pattern classification method based on eigenvalue clustering according to claim 3 is characterized in that The specific process of step S2 is as follows: Select data from the original data and , the data and The similarity or dissimilarity measure is denoted as: ； in, and ∈ ; The measurement matrix is ​​expressed as follows: 。 5. The acoustic signal pattern classification method based on eigenvalue clustering according to claim 4 is characterized in that In step S3, the spectrum matrix is ​​expressed as follows: ， The relationship between the spectral matrix and the metric matrix is ​​as follows: 。 6. The acoustic signal pattern classification method based on eigenvalue clustering according to claim 5 is characterized in that The specific process of step S4 is as follows: Spectral eigenvalue sequence Perform k-means clustering to obtain k groups of results. ； in, , .

Images