Sound source separation method, device and storage medium using local sensitive hashing

Abstract

The present invention discloses a sound source separation method, device and storage medium using local sensitive hashing. The method of the present invention comprises: (1) obtaining a plurality of single-frame mixed sound signals as training set samples, constructing a spectrum dictionary and an IBM matrix Y of the training set; (2) using a two-class classifier to learn a plurality of projection vectors, and using the learned projection vectors to convert the spectrum dictionary into a hash binary code to obtain a hash spectrum dictionary; (3) extracting the spectrum of the mixed signal to be separated, and using the learned projection vectors to convert the spectrum of the mixed signal to be separated into a hash binary code; (4) finding the index values of the #imgabs0# spectrums that are closest to the hash binary code of the spectrum of the mixed signal to be separated in the hash spectrum dictionary, finding the corresponding #imgabs1# IBM vectors in the IBM matrix Y, and using the weighted average of the #imgabs2# IBM vectors to estimate the IBM vector #imgabs3# of the source signal in the mixed signal to be separated; and (5) restoring the source signal based on #imgabs4#. The present invention has higher accuracy and is more lightweight.

Description

Technical Field

[0001] The present invention relates to sound processing signals, and in particular to a sound source separation method, device and storage medium using local sensitive hashing. Background Art

[0002] Over the past decade, the rapid development of deep learning has significantly improved the performance of single-channel sound source separation. However, this performance improvement comes at the cost of increased model complexity. In real-world scenarios, sound source separation algorithms are applied to devices with limited computing resources, such as mobile phones, headphones, laptops, and hearing aids. Therefore, existing large-model algorithms are not easily deployed on edge AI devices to improve voice quality.

[0003] Extensive research has focused on designing algorithms that strike a balance between complexity and accuracy for mobile and embedded applications. For example, researchers have proposed various optimization algorithms for convolution operations to build compact, low-latency models. Additionally, methods such as model parameter pruning, model compression, and low-bitrate quantization are also important approaches to reducing model computational complexity. However, the efficiency of these models needs to be improved, and their complexity needs to be reduced.

[0004] Locality-sensitive hashing (LSH) is commonly used to generate hash codes. It projects sample points from the source space into a hash space by constructing a series of hash functions, while preserving the similarity between the sample points. That is, similar samples in the source space are more likely to collide in the hash space than dissimilar samples. While LSH is simple and effective, its randomly selected hash functions are not trainable, which limits its performance on specific problems. Summary of the Invention

[0005] Purpose of the invention: To address the problems existing in the prior art, the present invention provides a more efficient and lightweight sound source separation method, device and storage medium.

[0006] Technical solution: The sound source separation method using local sensitive hashing described in the present invention includes:

[0007] (1) Obtain several single-frame mixed sound signals as training set samples, extract the spectrum of each sample to construct the spectrum dictionary and IBM matrix Y of the training set;

[0008] (2) Using a two-class classifier, we learn several optimal projection vectors, and use the learned optimal projection vectors to convert the spectrum dictionary into a hash binary code to obtain a hash spectrum dictionary;

[0009] (3) extracting the spectrum of the mixed signal to be separated, and converting the spectrum of the mixed signal to be separated into a hash binary code using the optimal projection vector obtained by learning;

[0010] (4) Find the spectrum hash binary code closest to the mixed signal to be separated in the hash spectrum dictionary The index value of the spectrum, using The index value is found in the IBM matrix Y. IBM vectors, and using The IBM vector of the source signal in the mixed signal to be separated is estimated by weighted average of IBM vectors

[0011] (5) According to IBM vector Recover the source signal from the mixed signal to be separated.

[0012] Furthermore, step (1) specifically includes:

[0013] (1-1) Obtain several single-frame mixed sound signals as training set samples, extract the short-time Fourier spectrum of each single-frame mixed sound signal, and perform L2 norm normalization;

[0014] (1-2) The short-time Fourier spectrum of all samples normalized by L2 norm is composed of a set As a spectral dictionary, where N represents the number of samples in the training set and F represents the short-time Fourier spectrum dimension;

[0015] (1-3) Calculate the IBM values of all samples in the training set and form the IBM matrix of the spectrum dictionary with the IBM values of all samples Among them, if the current spectrum of the sample is dominated by the target signal, the corresponding IBM value is 1, and if the interference signal is dominant, the corresponding IBM value is 0.

[0016] Furthermore, step (2) specifically includes:

[0017] (2-1) The self-similar matrix weight The initial value of the uniform distribution is initialized, and the projection matrix composed of all projection vectors Perform random initialization and set the classifier weight vector The initial value of is assigned to the zero vector, where N represents the number of samples in the training set, L represents the number of projection vectors, F represents the dimension of the short-time Fourier spectrum, and : represents taking all elements of this dimension;

[0018] (2-2) The spectrum dictionary that introduces context information is calculated according to the following formula

[0019]

[0020]

[0021] Among them, H *,: represents the first row of the spectrum dictionary H, Represents a spectrum dictionary that introduces contextual information The first row of η l,* =exp(-‖*-l‖ / τ) represents the context window, τ is the window width, and concat represents feature concatenation along the column;

[0022] (2-3) Set the number of loops l = 1;

[0023] (2-4) Calculate according to the following formula:

[0024] A.

[0025] B.

[0026] C.

[0027] D.

[0028] Where, Represents a hash spectrum dictionary in iteration Column * of P l,: represents the lth row of P; T is the matrix transpose operation; sgn(·) is the sign function; is the bias term, Represent the self-similar matrix weight matrix respectively The lth and l-1th submatrices of , ⊙ represents element-by-element multiplication, β l-1 , β l Represent the weights of the l-1th and lth classifiers respectively, Dist{·} represents the cross entropy cost function calculated according to each element in the matrix, ε l represents the cost function of the l-th projection;

[0029] (2-5) The calculated P l,: As the optimal projection vector , determine whether l is less than L, if so, set l=l+1 and return to step (2-4), if not, execute step (2-6);

[0030] (2-6) According to the following formula, the optimal projection vector obtained by learning is used Convert the spectrum dictionary into a hash binary code to obtain the optimal hash spectrum dictionary

[0031]

[0032]

[0033] Furthermore, step (3) specifically includes:

[0034] (3-1) Extract the short-time Fourier spectrum of the mixed signal to be separated and perform L2 norm normalization;

[0035] (3-2) According to the following formula, the L2 norm-normalized spectrum of the mixed signal to be separated is converted into a hash binary code using the learned projection vector:

[0036]

[0037] b={b l |l=1,…,L}

[0038] Where, represents the spectrum hash binary code of the mixed signal to be separated, x represents the L2 norm normalized spectrum of the mixed signal to be separated; sgn(·) is the sign function; is the bias term, P * Represents the optimal projection matrix composed of the learned optimal projection vectors.

[0039] Furthermore, step (4) specifically includes:

[0040] (4-1) Set up the collection Random selection at initialization Codeword index filling set up Representing a collection All code words in The maximum distance;

[0041] (4-2) Set the number of iterations t = 1;

[0042] (4-3) Calculate the spectrum hash binary code of the mixed signal to be separated and hash spectrum dictionary The tth row of The Hamming distance is assigned to D t , and judge D t Is it less than If yes, go to step (4-4), if no, go to step (4-6);

[0043] (4-4) Replace t The index corresponding to the maximum distance;

[0044] (4-5) Update according to the following formula

[0045]

[0046] In the formula, Hm() represents the Hamming distance, Represents a hash spectrum dictionary The kth row of

[0047] (4-6) Determine whether t is less than N, where N represents the hash spectrum dictionary If so, set t=t+1 and return to step (4-3), otherwise, execute step (4-7);

[0048] (4-7) is normalized according to the following formula:

[0049]

[0050] (4-8) Estimate the masking vector of the source signal in the mixed signal to be separated according to the following formula:

[0051]

[0052] Where Y k,: represents the k-th row of the IBM matrix Y.

[0053] Furthermore, step (5) specifically includes:

[0054] According to the masking vector Recover the source signal ss from the mixed signal to be separated:

[0055]

[0056] Where ⊙ represents element-by-element multiplication, and x represents the short-time Fourier spectrum of the mixed signal to be separated.

[0057] The sound source separation device using local sensitive hashing of the present invention includes:

[0058] one or more processors;

[0059] a memory for storing one or more programs;

[0060] When the one or more programs are executed by the one or more processors, the one or more processors implement the above method.

[0061] The storage medium containing computer-executable instructions of the present invention is characterized in that the computer-executable instructions are used to perform the above method when executed by a computer processor.

[0062] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0063] 1. Considering that when the spectra of a certain frame of two observation signals are similar, their latent source signals should also be similar, so that their ideal binary masks (IBM) are also consistent. Therefore, the present invention defines the sound source separation problem as a The nearest neighbor search task is to find the closest spectrum of the mixed signal to be separated in the spectrum dictionary composed of the training set, and then find the sound signal corresponding to the spectrum to achieve source separation, which is more accurate than existing methods;

[0064] 2. During the spectrum dictionary search process, the spectrum of the signal to be separated and the spectrum dictionary constructed by the training set are converted into hash codes, which improves the retrieval efficiency;

[0065] 3. Considering that the hash function randomly selected by LSH is not trainable and its performance needs to be improved, the present invention extends LSH and introduces a learnable hash code generation strategy. It adopts a sequential learning method to continuously learn and generate a series of complementary hash projection functions. In order to control the discriminability and stability of binary codes, the perceptron is used as a two-class classifier. At the same time, drawing on the idea of spectral hashing, a training objective is constructed to minimize the difference between the self-similarity matrix between codes in the hash space and the self-similarity matrix between original spectrum samples. This allows the hash code to retain the similarity between samples in the original spectrum space to the greatest extent, ensuring the accuracy of the final separation.

[0066] 4. The binary spectrum dictionary embedding of the present invention can be stored cheaply, and The nearest neighbor search can be accelerated by bit operations, so the present invention is a lightweight sound source separation algorithm that is easy to promote. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The present invention defines the sound source separation problem as a For the nearest neighbor search task, by searching for the spectrum closest to the spectrum of the mixed signal to be separated in the spectrum dictionary formed by the training set, the sound signal corresponding to the spectrum is found to achieve source separation. If directly searching in the spectrum dictionary, the computational complexity of the algorithm is O(FQT), where F and T represent the dimension of the feature vector and the number of feature vectors respectively, and Q represents the floating-point precision (Q = 64 for double precision). To reduce the computational complexity, it is necessary to reduce F and Q. This invention uses Locality-Sensitive Hashing (LSH) to achieve complexity reduction. Through the LSH transformation, the dimension of the embedded vector can be reduced to L (L < F), and the floating-point precision can be reduced to 1 (1 bit). Since the LSH projection matrix is randomly selected, it cannot guarantee that the generated hash codes have strong discriminability, resulting in usually longer required codewords. To increase the discriminability and stability of the hash codes, a learning strategy is introduced. This invention models the construction process of the hash codes as a series of binary classification problems, enabling the generation of the codewords to consider the distribution characteristics of the training data, thereby generating hash binary codes adapted to specific problems.

[0069] Based on the above analysis, an embodiment of this invention provides a sound source separation method using Locality-Sensitive Hashing, as Figure 1 shown, including the following steps:

[0070] (1) Obtain a number of single-frame mixed sound signals as training set samples, and extract the spectrum of each sample to construct the spectrum dictionary and the IBM matrix Y of the training set.

[0071] This step specifically includes:

[0072] (1-1) Obtain a number of single-frame mixed sound signals as training set samples, extract the short-time Fourier spectrum of each single-frame mixed sound signal, and perform L2 norm normalization;

[0073] (1-2) Combine the short-time Fourier spectra of all samples after L2 norm normalization into a set as the spectrum dictionary, where N represents the number of samples in the training set, and F represents the dimension of the short-time Fourier spectrum;

[0074] (1-3) Calculate the IBM values of all samples in the training set, and form the IBM matrix of the spectrum dictionary with the IBM values of all samples where, if the current spectrum of the sample is dominated by the target signal, the corresponding IBM value is 1, and if it is dominated by the interference signal, the corresponding IBM value is 0.

[0075] (2) Use a binary classifier to learn a number of optimal projection vectors, and use the learned optimal projection vectors to convert the spectrum dictionary into hash binary codes to obtain a hash spectrum dictionary. <00002及44>

[0076] Specifically including:

[0077] (2-1) The self-similar matrix weight The initial value of the uniform distribution is initialized, and the projection matrix composed of all projection vectors Perform random initialization and set the classifier weight vector The initial value of is assigned to the zero vector, where N represents the number of samples in the training set, L represents the number of projection vectors, D represents the dimension of the short-time Fourier spectrum, and : represents taking all elements of this dimension;

[0078] (2-2) The spectrum dictionary that introduces context information is calculated according to the following formula

[0079]

[0080]

[0081] Among them, H *,: represents the first row of the spectrum dictionary H, Represents a spectrum dictionary that introduces contextual information The first row of η l,* =exp(-‖*-l‖ / τ) represents the context window, τ is the window width, and concat represents feature concatenation along the column;

[0082] (2-3) Set the number of loops l = 1;

[0083] (2-4) Calculate according to the following formula:

[0084] A.

[0085] B.

[0086] C.

[0087] D.

[0088] Where, Represents a hash spectrum dictionary in iteration Column * of P l,: represents the lth row of P; T is the matrix transpose operation; sgn(·) is the sign function; is the bias term, Represent the self-similar matrix weight matrix respectively The lth and l-1th submatrices of , ⊙ represents element-by-element multiplication, β l-1 , β l Respectively represent the l-1th and lth classifier weights, ε lrepresents the cost function of the lth projection; Dist{·} represents the cross entropy cost function calculated according to each element in the matrix, specifically:

[0089]

[0090] in, and are all self-similar matrices; the target dominance of each training sample is calculated using the IBM matrices Y and H: γ = {∑ f [Y⊙H-(1-Y)⊙H]}≥0, and calculate its autocorrelation matrix γγ T ∈{0,1} N×N , divide the N×N elements into two groups, 1 represents the same type of sample and 0 represents the different type of sample, and define their sets as and In order to reduce the impact of the imbalance of the two types of training data, the normalization factor λ0 is introduced. The above cost function shows that the feature vectors in the hash code space should retain the similarity between the feature vectors in the original spectrum space;

[0091] (2-5) The calculated P l,: As the optimal projection vector Determine whether l is less than L. If so, set l = l + 1 and return to step (2-4). If not, execute step (2-6).

[0092] (2-6) According to the following formula, the optimal projection vector obtained by learning is used Convert the spectrum dictionary into a hash binary code to obtain the optimal hash spectrum dictionary

[0093]

[0094]

[0095] (3) Extract the spectrum of the mixed signal to be separated, and use the optimal projection vector obtained by learning to convert the spectrum of the mixed signal to be separated into a hash binary code.

[0096] Specifically include:

[0097] (3-1) Extract the short-time Fourier spectrum of the mixed signal to be separated and perform L2 norm normalization;

[0098] (3-2) According to the following formula, the L2 norm-normalized spectrum of the mixed signal to be separated is converted into a hash binary code using the learned projection vector:

[0099]

[0100] b={b l|l=1,…,L}

[0101] Where, represents the spectrum hash binary code of the mixed signal to be separated, x represents the L2 norm normalized spectrum of the mixed signal to be separated; sgn(·) is the sign function; is the bias term, P * Represents the optimal projection matrix composed of the learned optimal projection vectors.

[0102] (4) Find the spectrum hash binary code closest to the mixed signal to be separated in the hash spectrum dictionary The index value of the spectrum, using The index value is found in the IBM matrix Y. IBM vectors, and using The IBM vector of the source signal in the mixed signal to be separated is estimated by weighted average of IBM vectors

[0103] Specifically include:

[0104] (4-1) Set up the collection Random selection at initialization Codeword index filling set up Representing a collection All code words in The maximum distance;

[0105] (4-2) Set the number of iterations t = 1;

[0106] (4-3) Calculate the spectrum hash binary code of the mixed signal to be separated and hash spectrum dictionary The tth row of The Hamming distance is assigned to D t , and judge D t Is it less than If yes, go to step (4-4), if no, go to step (4-6);

[0107] (4-4) Replace t The index corresponding to the maximum distance;

[0108] (4-5) Update according to the following formula

[0109]

[0110] In the formula, Hm() represents the Hamming distance, Represents a hash spectrum dictionary The kth row of

[0111] (4-6) Determine whether t is less than N, where N represents the hash spectrum dictionary If so, set t=t+1 and return to step (4-3), otherwise, execute step (4-7);

[0112] (4-7) is normalized according to the following formula:

[0113]

[0114] (4-8) Estimate the masking vector of the source signal in the mixed signal to be separated according to the following formula:

[0115]

[0116] Where Y k,: represents the k-th row of the IBM matrix Y.

[0117] (5) According to IBM vector Restoring the source signal from the mixed signal to be separated;

[0118]

[0119] Where ⊙ represents element-by-element multiplication, and x represents the short-time Fourier spectrum of the mixed signal to be separated.

[0120] This embodiment also provides a sound source separation device using local sensitive hashing. The embodiment of the present invention provides services for the implementation of the above-mentioned method of the present invention, including one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0121] The system memory may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The device may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the storage system may be used to read and write non-removable, non-volatile magnetic media. The memory may include at least one program product having a set (e.g., at least one) program modules that are configured to perform the functions of various embodiments of the present invention. A program / utility having a set (e.g., at least one) program module may be stored in, for example, a memory, such program modules including but not limited to an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment. The program modules typically perform the functions and / or methods of the embodiments described herein. The processing unit executes various functional applications and data processing, such as implementing the method of embodiment one, by running the program stored in the system memory.

[0122] The present invention also provides a storage medium containing computer-executable instructions, which are used for the above method when executed by a computer processor.

[0123] The computer storage medium of the embodiment of the present invention can adopt any combination of one or more computer-readable media. Computer-readable media can be computer-readable signal media or computer-readable storage media. Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or components, or any combination thereof. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by an instruction execution system, device or device or used in combination with it.

[0124] A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0125] Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0126] The computer program code for performing the operations of the present invention can be written in one or more programming languages, or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0127] The above disclosure is only a preferred embodiment of the present invention and cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A sound source separation method using local sensitive hashing, characterized in that The method comprises the steps of: (1) Obtain several single-frame mixed sound signals as training set samples, extract the spectrum of each sample to construct the spectrum dictionary and IBM matrix Y of the training set; (2) Using a two-class classifier, we learn several optimal projection vectors, and use the learned optimal projection vectors to convert the spectrum dictionary into a hash binary code to obtain a hash spectrum dictionary; (3) extracting the spectrum of the mixed signal to be separated, and converting the spectrum of the mixed signal to be separated into a hash binary code using the optimal projection vector obtained by learning; (4) Find the spectrum hash binary code closest to the mixed signal to be separated in the hash spectrum dictionary The index value of the spectrum, using The index value is found in the IBM matrix Y. IBM vectors, and using The IBM vector of the source signal in the mixed signal to be separated is estimated by weighted average of IBM vectors (5) According to IBM vector Recover the source signal from the mixed signal to be separated.

2. The sound source separation method using locality sensitive hashing according to claim 1, characterized in that: Step (1) specifically includes: (1-1) Obtain several single-frame mixed sound signals as training set samples, extract the short-time Fourier spectrum of each single-frame mixed sound signal, and perform L2 norm normalization; (1-2) The short-time Fourier spectrum of all samples normalized by L2 norm is composed of a set As a spectral dictionary, where N represents the number of samples in the training set and F represents the short-time Fourier spectrum dimension; (1-3) Calculate the IBM values of all samples in the training set and form the IBM matrix of the spectrum dictionary with the IBM values of all samples Among them, if the current spectrum of the sample is dominated by the target signal, the corresponding IBM value is 1, and if the interference signal is dominant, the corresponding IBM value is 0.

3. The sound source separation method using locality sensitive hashing according to claim 1, characterized in that: Step (2) specifically includes: (2-1) The self-similar matrix weight The initial value of the uniform distribution is initialized, and the projection matrix composed of all projection vectors Perform random initialization and set the classifier weight vector The initial value of is assigned to the zero vector, where N represents the number of samples in the training set, L represents the number of projection vectors, F represents the dimension of the short-time Fourier spectrum, and : represents taking all elements of this dimension; (2-2) The spectrum dictionary that introduces context information is calculated according to the following formula Among them, H *,: represents the first row of the spectrum dictionary H, Represents a spectrum dictionary that introduces contextual information The first row of η l,· =exp(-‖*-l‖ / τ) represents the context window, τ is the window width, and concat represents feature concatenation along the column; (2-3) Set the number of loops l = 1; (2-4) Calculate according to the following formula: A、 B、 C、 D、 Where, Represents a hash spectrum dictionary in iteration Column * of P l,: represents the lth row of P; T is the matrix transpose operation; sgn(·) is the sign function; is the bias term, Represent the self-similar matrix weight matrix respectively The lth and l-1th submatrices of , ⊙ represents element-by-element multiplication, β l-1 , β l Represent the weights of the l-1th and lth classifiers respectively, Dist{·} represents the cross entropy cost function calculated according to each element in the matrix, ε l represents the cost function of the l-th projection; (2-5) The calculated P l,: As the optimal projection vector Determine whether l is less than L. If so, set l = l + 1 and return to step (2-4). If not, execute step (2-6). (2-6) According to the following formula, the optimal projection vector obtained by learning is used Convert the spectrum dictionary into a hash binary code to obtain the optimal hash spectrum dictionary 4. The sound source separation method using locality sensitive hashing according to claim 1, characterized in that: Step (3) specifically includes: (3-1) Extract the short-time Fourier spectrum of the mixed signal to be separated and perform L2 norm normalization; (3-2) According to the following formula, the L2 norm-normalized spectrum of the mixed signal to be separated is converted into a hash binary code using the learned projection vector: Where, represents the spectrum hash binary code of the mixed signal to be separated, x represents the L2 norm normalized spectrum of the mixed signal to be separated; sgn(·) is the sign function; is the bias term, P * represents the optimal projection matrix composed of the optimal projection vectors obtained by learning, L represents the number of projection vectors, and T is the matrix transpose operation.

5. The sound source separation method using locality sensitive hashing according to claim 1, characterized in that: Step (4) specifically includes: (4-1) Set up the collection Random selection at initialization Codeword index filling set up Representing a collection All code words in The maximum distance; (4-2) Set the number of iterations t = 1; (4-3) Calculate the spectrum hash binary code of the mixed signal to be separated and hash spectrum dictionary The tth row of The Hamming distance is assigned to D t , and judge D t Is it less than If yes, go to step (4-4), if no, go to step (4-6); (4-4) Replace t The index corresponding to the maximum distance; (4-5) Update according to the following formula In the formula, Hm() represents the Hamming distance, Represents a hash spectrum dictionary The kth row of (4-6) Determine whether t is less than N, where N represents the hash spectrum dictionary If so, set t=t+1 and return to step (4-3), otherwise, execute step (4-7); (4-7) is normalized according to the following formula: (4-8) Estimate the masking vector of the source signal in the mixed signal to be separated according to the following formula: Where Y k,: represents the k-th row of the IBM matrix Y.

6. The sound source separation method using locality sensitive hashing according to claim 1, characterized in that: Step (5) specifically includes: According to the masking vector Recover the source signal ss from the mixed signal to be separated: Where ⊙ represents element-by-element multiplication, and x represents the short-time Fourier spectrum of the mixed signal to be separated.

7. A sound source separation device using local sensitive hashing, characterized in that: include: one or more processors; a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 6.

8. A storage medium containing computer-executable instructions, characterized in that: The computer executable instructions are used to perform the method according to any one of claims 1 to 6 when executed by a computer processor.

Images