Voice signal blind source separation method and device, equipment and storage medium

Abstract

The invention discloses a voice signal blind source separation method, device and equipment and a storage medium, and relates to the technical field of voice separation, and the method comprises the steps: collecting a mixed sound signal of a plurality of source signals through a plurality of sound collection devices, and obtaining a plurality of time domain observation signals; converting the time domain observation signal into a frequency domain observation signal, and creating a frequency domain separation signal model based on the frequency domain observation signal and a preset separation matrix; obtaining the direction of arrival of the source signal to obtain direction of arrival information, and performing iterative updating on a preset separation matrix in the frequency domain separation signal model based on an independent vector analysis algorithm of an auxiliary function to perform blind source separation on the frequency domain observation signal to obtain a frequency domain separation signal; the geometric constraint of the iterative update of the preset separation matrix is the constraint created based on the direction of arrival information; and converting the frequency domain separation signal into a time domain separation signal. According to the invention, the efficiency of voice signal blind source separation can be improved, and the accuracy of blind source separation is remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of speech separation technology, and in particular to a speech signal blind source separation method, device, equipment and storage medium. Background Art

[0002] With the rapid development of information technology, blind source separation (BSS) technology has been widely used in many fields, such as speech separation, biomedical signal analysis, wireless communication, and audio and video signal processing.

[0003] Blind source separation technology is a technology that recovers the original signal from multiple mixed signals without knowing the source signal and the mixing process. Traditional blind source separation methods usually use the Independent Component Analysis (ICA) algorithm to separate source signals. This method is based on the assumption that source signals are statistically independent of each other and only uses the observed signal to estimate the source signal. Since it does not require much prior knowledge of the source signal, this method has a wide range of applications and has received widespread attention in the industry. However, the signals in traditional methods are usually processed in the time domain and involve multi-channel convolution operations, which leads to slow algorithm convergence, and the component arrangement is difficult to distinguish, requiring post-processing, and it is difficult to achieve the optimal solution.

[0004] ‌At present, a blind source separation method based on independent vector analysis (IVA) is proposed for traditional blind source separation methods. This method converts the blind source separation problem to the frequency domain, converts the convolution in the time domain into multiplication in the frequency domain, and achieves signal separation by independently iterating at each frequency point to obtain the separation matrix. Moreover, when dealing with the blind source separation problem, this method not only focuses on the independence of the source signal, but also takes into account the correlation between different channel signals. In addition, a specific statistical model is introduced to more accurately describe the characteristics of the signal, thereby improving the accuracy of separation.

[0005] However, in practical applications, the current offline IVA is a time-invariant model and cannot separate speech in real time. Moreover, its separation matrix is ​​updated on all time frames. In actual production scenarios, real-time performance is usually required, and mobile source signals need to be tracked and separated. This requires that the separation matrix can be updated time frame by time frame, and the global arrangement problem in tracking needs to be solved. At the same time, in practical applications, the processing delay of the model needs to be reduced. Summary of the invention

[0006] In view of this, the purpose of this application is to provide a method, device, equipment and storage medium for blind source separation of speech signals, which can improve the efficiency of blind source separation of speech signals and significantly improve the accuracy of blind source separation. The specific scheme is as follows:

[0007] In a first aspect, the present application discloses a method for blind source separation of speech signals, comprising:

[0008] Collecting mixed sound signals of the second number of source signals by using a first number of sound collection devices respectively to obtain the first number of time domain observation signals;

[0009] Convert the time domain observation signal into a frequency domain signal to obtain a frequency domain observation signal, and create a frequency domain separation signal model based on the frequency domain observation signal and a preset separation matrix;

[0010] Acquire the direction of arrival of the source signal to obtain direction of arrival information, and iteratively update the preset separation matrix in the frequency domain separation signal model based on the independent vector analysis algorithm of the auxiliary function to perform blind source separation on the frequency domain observation signal in the frequency domain separation signal model to obtain a frequency domain separation signal; the geometric constraints for iteratively updating the preset separation matrix are constraints created based on the direction of arrival information;

[0011] The frequency domain separation signal is converted into a time domain signal to obtain a time domain separation signal.

[0012] Optionally, converting the time domain observation signal into a frequency domain signal to obtain the frequency domain observation signal includes:

[0013] Performing short-time Fourier transform on the time-domain observation signal to obtain a frequency-domain observation signal.

[0014] Optionally, converting the frequency domain separated signal into a time domain signal to obtain the time domain separated signal includes:

[0015] The frequency domain separation signal is scaled restored by back projection to obtain a restored separation signal, and the restored separation signal is inversely short-time Fourier transformed to obtain a time domain separation signal.

[0016] Optionally, the auxiliary function-based independent vector analysis algorithm iteratively updates a preset separation matrix in the frequency domain separation signal model to perform blind source separation on the frequency domain observation signal in the frequency domain separation signal model to obtain a frequency domain separation signal, including:

[0017] Preprocessing the preset separation matrix and the frequency domain observation signal in the frequency domain separation signal model to obtain a processed separation matrix and a processed observation signal;

[0018] The processed separation matrix is ​​iteratively updated based on an independent vector analysis algorithm of an auxiliary function to perform blind source separation on the processed observation signal to obtain a frequency domain separation signal.

[0019] Optionally, preprocessing the preset separation matrix and the frequency domain observation signal in the frequency domain separation signal model to obtain a processed separation matrix and a processed observation signal includes:

[0020] The preset separation matrix and the frequency domain observation signal in the frequency domain separation signal model are respectively subjected to averaging processing and whitening processing to obtain a processed separation matrix and a processed observation signal; wherein the whitening processing adopts a PCA dimensionality reduction method.

[0021] Optionally, the auxiliary function-based independent vector analysis algorithm iteratively updates the processed separation matrix to perform blind source separation on the processed observation signal to obtain a frequency domain separation signal, including:

[0022] Iteratively updating the processed separation matrix in an iterative source-guided update manner and using an auxiliary function in an independent vector analysis algorithm based on an auxiliary function to obtain an updated separation matrix, and determining whether the updated separation matrix converges during the iterative update process;

[0023] If the updated separation matrix converges, blind source separation is performed on the processed observation signal frequency-by-frequency point based on the updated separation matrix in an online source-guided updating manner and using the independent vector analysis algorithm to obtain a frequency domain separated signal.

[0024] Optionally, acquiring the direction of arrival of the source signal to obtain direction of arrival information includes:

[0025] The direction of arrival of the source signal is fitted based on a positioning algorithm to obtain direction of arrival information.

[0026] In a second aspect, the present application discloses a speech signal blind source separation device, comprising:

[0027] A collection module, used for collecting mixed sound signals of a second number of source signals through a first number of sound collection devices respectively, to obtain the first number of time domain observation signals;

[0028] A frequency domain conversion module, used for converting the time domain observation signal into a frequency domain signal to obtain a frequency domain observation signal, and creating a frequency domain separation signal model based on the frequency domain observation signal and a preset separation matrix;

[0029] An acquisition module, used to acquire the direction of arrival of the source signal to obtain direction of arrival information;

[0030] A blind source separation module, used for iteratively updating a preset separation matrix in the frequency domain separation signal model based on an independent vector analysis algorithm of an auxiliary function, so as to perform blind source separation on the frequency domain observation signal in the frequency domain separation signal model to obtain a frequency domain separation signal; the geometric constraints for iteratively updating the preset separation matrix are constraints created based on the wave direction information;

[0031] The time domain conversion module is used to convert the frequency domain separation signal into a time domain signal to obtain a time domain separation signal.

[0032] In a third aspect, the present application discloses an electronic device, comprising a processor and a memory; wherein the processor implements the aforementioned method for blind source separation of speech signals when executing a computer program stored in the memory.

[0033] In a fourth aspect, the present application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the aforementioned method for blind source separation of speech signals is implemented.

[0034] It can be seen that the present application first collects the mixed sound signals of the second number of source signals through the first number of sound collection devices respectively to obtain the first number of time domain observation signals, and then converts the time domain observation signals into frequency domain signals to obtain frequency domain observation signals, and creates a frequency domain separation signal model based on the frequency domain observation signals and the preset separation matrix, and then obtains the wave arrival direction of the source signal to obtain the wave arrival direction information, and iteratively updates the preset separation matrix in the frequency domain separation signal model based on the independent vector analysis algorithm of the auxiliary function to perform blind source separation on the frequency domain observation signals in the frequency domain separation signal model to obtain a frequency domain separation signal, the geometric constraints of the iterative update of the preset separation matrix are the constraints created based on the wave arrival direction information, and finally the frequency domain separation signal is converted into a time domain signal to obtain a time domain separation signal. This application converts the time domain observation signal into a frequency domain observation signal, and then uses an independent vector analysis algorithm based on an auxiliary function to perform blind source separation on the frequency domain observation signal, and uses the wave direction information as a geometric constraint for iterative update. Through the above method, not only the efficiency of blind source separation of speech signals can be improved, but also the accuracy of blind source separation can be significantly improved. At the same time, it is suitable for deployment on devices with limited computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0036] Figure 1 A flow chart of a method for blind source separation of speech signals disclosed in this application;

[0037] Figure 2 A flow chart of a specific method for blind source separation of speech signals disclosed in this application;

[0038] Figure 3 This is a schematic diagram of the structure of a speech signal blind source separation device disclosed in this application;

[0039] Figure 4 This is a structural diagram of an electronic device disclosed in this application. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0041] The present application discloses a method for blind source separation of speech signals. Figure 1 As shown, the method includes:

[0042] Step S11: using a first number of sound collection devices to respectively collect mixed sound signals of a second number of source signals to obtain the first number of time domain observation signals.

[0043] It is understandable that due to the time difference of sound propagation and the multipath effect caused by reflection, the signal received by the microphone is actually a mixed signal of multiple source signals. In this embodiment, when it is necessary to perform blind source separation on speech signals from multiple signal sources, firstly, M source signals (i.e., original signals) are respectively separated by N sound collection devices (such as microphones). The mixed sound signal is collected to obtain the N time domain observation signals .

[0044] Step S12: Convert the time domain observation signal into a frequency domain signal to obtain a frequency domain observation signal, and create a frequency domain separation signal model based on the frequency domain observation signal and a preset separation matrix.

[0045] In this embodiment, after collecting a mixed sound signal of multiple source signals, multiple time domain observation signals are obtained. After that, further, the above time domain observation signal The signal is converted into a frequency domain signal to obtain a corresponding frequency domain observation signal, and then a frequency domain separation signal model is created based on the frequency domain observation signal and a preset separation matrix.

[0046] Specifically, converting the time domain observation signal into a frequency domain signal to obtain the frequency domain observation signal may include: performing a short-time Fourier transform on the time domain observation signal to obtain the frequency domain observation signal. In this embodiment, the time domain observation signal may be converted into a frequency domain signal by a short-time Fourier transform (STFT). Transformed into frequency domain observation signal ,in, is the confusion matrix, Source signal The frequency domain signal is obtained after the short-time Fourier transform. Further, based on the preset separation matrix Create a frequency domain separation signal model, which can be specifically expressed as:

[0047] ;

[0048] In the formula, , preset separation matrix , T is transpose, H is conjugate transpose; , and are the observed signal, predicted signal and original signal in the frequency domain respectively, where is known, and is unknown. The purpose of blind source separation of speech signals is to find As close as possible to the predicted signal .

[0049] Step S13: Acquire the direction of arrival of the source signal to obtain the direction of arrival information, and iteratively update the preset separation matrix in the frequency domain separation signal model based on the independent vector analysis algorithm of the auxiliary function to perform blind source separation on the frequency domain observation signal in the frequency domain separation signal model to obtain a frequency domain separation signal; the geometric constraints for iteratively updating the preset separation matrix are constraints created based on the direction of arrival information.

[0050] In this embodiment, the directions of arrival (DOA) of multiple source signals can be first obtained to obtain direction of arrival information, and then the preset separation matrix in the frequency domain separation signal model can be iteratively updated based on the independent vector analysis (IVA) algorithm of auxiliary functions to perform blind source separation on the frequency domain observation signal in the frequency domain separation signal model to obtain the corresponding frequency domain separation signal. Among them, the geometric constraints for iterative updates of the preset separation matrix are constraints created based on the above-mentioned direction of arrival information (i.e., DOA information). It should be pointed out that in order to reduce the number of iterations, speed up the convergence speed of the model, and solve the global arrangement problem, this solution creates iteratively updated geometric constraints (Geometrically Constrained, GC) based on DOA information.

[0051] Specifically, the acquiring the direction of arrival of the source signal to obtain the direction of arrival information may include: fitting the direction of arrival of the source signal based on a positioning algorithm to obtain the direction of arrival information. In this embodiment, for a semi-blind system with a known direction of arrival of a source signal, the direction of arrival of the source signal may be directly acquired, and for a fully blind system with an unknown direction of arrival of a source signal, the positioning algorithm may be used to fit and estimate the DOA information of the unknown source signal.

[0052] Specifically, according to the actual application scenario, the TDOA (Time Difference of Arrival) algorithm, GCC-PATH (Generalized Cross Correlation-Phase Transform) or Delay and Sum (DS) algorithm can be used to estimate the direction of arrival of the source signal. Among them, the TDOA algorithm can enhance the signal from a specific direction by aligning the phase of the wave by using the arrival time difference (TDOA) between microphones. The algorithm can estimate the direction of arrival of the source signal from the estimated confusion matrix based on the assumption of plane wave arrival. and the steering vector Calculate the angle of arrival , the specific calculation formula is:

[0053] .

[0054] Among them, the independent vector analysis algorithm based on the auxiliary function iteratively updates the preset separation matrix in the frequency domain separation signal model to perform blind source separation on the frequency domain observation signal in the frequency domain separation signal model to obtain the frequency domain separation signal. Specifically, it may include: preprocessing the preset separation matrix and the frequency domain observation signal in the frequency domain separation signal model to obtain a processed separation matrix and a processed observation signal; the independent vector analysis algorithm based on the auxiliary function iteratively updates the processed separation matrix to perform blind source separation on the processed observation signal to obtain the frequency domain separation signal. In this embodiment, after obtaining the frequency domain separation signal model, the frequency domain observation signals in the model are respectively preprocessed. Perform value and whitening processing, and use the preset separation matrix Orthogonal, thus maintaining After averaging and whitening, the processed separation matrix and the processed observation signal are obtained; then, the independent vector analysis algorithm based on the auxiliary function is used to iteratively update the processed separation matrix, thereby performing blind source separation on the processed observation signal to obtain the frequency domain separation signal.

[0055] Specifically, the preprocessing of the preset separation matrix and the frequency domain observation signal in the frequency domain separation signal model to obtain the processed separation matrix and the processed observation signal may include: respectively performing mean processing and whitening processing on the preset separation matrix and the frequency domain observation signal in the frequency domain separation signal model to obtain the processed separation matrix and the processed observation signal; wherein the whitening processing adopts PCA (Principal Component Analysis) dimensionality reduction method. In this embodiment, when performing whitening processing, PCA dimensionality reduction method may be used for whitening processing, the purpose of which is to remove the correlation between signals, that is, to ensure that the preset separation matrix The rows of are kept orthogonal, thereby reducing complexity and achieving better convergence of the model. In addition, the whitening process can eliminate the autocorrelation in the signal and simplify the frequency domain characteristics of the signal, making the subsequent signal processing steps clearer and easier to execute.

[0056] In a specific implementation, the auxiliary function-based independent vector analysis algorithm iteratively updates the processed separation matrix to perform blind source separation on the processed observation signal to obtain a frequency domain separation signal. Specifically, it may include: iteratively updating the processed separation matrix in an iterative source-guided update manner and using the auxiliary function in the auxiliary function-based independent vector analysis algorithm to obtain an updated separation matrix, and judging whether the updated separation matrix converges during the iterative update process; if the updated separation matrix converges, blind source separation is performed on the processed observation signal frequency by frequency point in an online source-guided update manner, based on the updated separation matrix, and using the independent vector analysis algorithm to obtain a frequency domain separation signal. In this embodiment, after analyzing the separated signal, it is known that the separated signal obeys a multivariate super-Gaussian distribution. Therefore, when performing blind source separation, in order to improve the stability of the iteration, this solution uses an auxiliary function to improve the stability of the iteration. Specifically, the KL (Kullback-Leibler) divergence between two functions can be selected to measure the independence of the signal. The specific calculation formula is:

[0057] ;

[0058] In order to ensure the real-time performance of blind source separation, The spatial covariance matrix in Updated by time frame, the specific formula is:

[0059] ;

[0060] In the formula, , , is a hyperparameter, and the loss of the geometric constraint is:

[0061] ;

[0062] In the formula, To adjust the weight parameters of the constraints, is the steering vector pointing to 𝜃, is a non-negative constraint value.

[0063] It should be pointed out that when When , the spatial filter is forced to form a beam in the direction of arrival, thereby retaining the target source; and when When is small, a spatial null is created in this direction so that multiple spatial null constraints toward all interfering directions can be used to suppress all interference and preserve the target source.

[0064] Furthermore, when all sources are stationary, the preset separation matrix is ​​replaced by estimating a vector substitution, namely ,in, If the updated channel is , then when the update vector is not a diagonal element (i.e. )hour, ;

[0065] When the update vector is a diagonal element (i.e. )hour, ;

[0066] In the formula, .

[0067] It should be pointed out that if you need to Find the inverse and update it with time frame t, then the update complexity is If you keep updating like this, Figure 2 The overall complexity of the process on the right is .

[0068] In this embodiment, in the source tracking stage (i.e., when only a single source is moving), an online source update method is adopted and the geometric constraints are discarded. The complexity is only After analysis, it is found that when only one source moves (such as k), as long as the preset separation matrix has converged, it is no longer necessary to provide DOA information to track the moving source. And because w is related to GC, the source tracking stage and the source separation stage use different update methods. Specifically, in the source tracking stage without geometric constraints, if , then the tracking result (i.e. prediction signal) is:

[0069] ;

[0070] if , then the tracking result (i.e. prediction signal) is:

[0071] ;

[0072] .

[0073] Through the above update method, it is avoided The inverse process greatly reduces the complexity and reduces the overall complexity to ,In addition, this scheme is based on the independent vector analysis algorithm ,based on the auxiliary function, which makes full use of the ,correlation of the source signal between frequency bands, and ,at the same time, applies the characteristics of the angle prior, and ,combines two different update methods (i.e., iterative source-guided update method ,and online source-guided update method), thereby improving the separation ,efficiency and robustness of the independent vector analysis algorithm.

[0074] Step S14: converting the frequency domain separation signal into a time domain signal to obtain a time domain separation signal.

[0075] In this embodiment, after the frequency domain separation signal is obtained, a time domain transformation is performed on it, so as to convert the frequency domain separation signal into a time domain signal to obtain a time domain separation signal.

[0076] In this embodiment, the frequency domain separation signal is converted into a time domain signal to obtain a time domain separation signal, including: using a reverse projection method to restore the scale of the frequency domain separation signal to obtain a restored separation signal, and performing an inverse short-time Fourier transform on the restored separation signal to obtain a time domain separation signal. It can be understood that the above steps perform a preprocessing operation on the signal in the model, such as a whitening operation, so that the scale of the signal changes. Therefore, after blind source separation, the signal needs to be scaled. Specifically, the frequency domain separation signal can be scaled first by using a reverse projection method to restore the scale to obtain a restored separation signal, and then the restored separation signal is subjected to an inverse short-time Fourier transform to obtain a time domain separation signal. Specifically, a microphone signal can be used as a reference, for example, the first microphone signal can be used as a reference, and its reverse projection is calculated, and then the signal of the first microphone channel in the reverse projection output signal is used as the final prediction signal.

[0077] It can be seen that the embodiment of the present application first collects the mixed sound signals of the second number of source signals through the first number of sound collection devices respectively to obtain the first number of time domain observation signals, and then converts the time domain observation signals into frequency domain signals to obtain frequency domain observation signals, and creates a frequency domain separation signal model based on the frequency domain observation signals and the preset separation matrix, and then obtains the wave arrival direction of the source signal to obtain the wave arrival direction information, and iteratively updates the preset separation matrix in the frequency domain separation signal model based on the independent vector analysis algorithm of the auxiliary function to perform blind source separation on the frequency domain observation signals in the frequency domain separation signal model to obtain a frequency domain separation signal, the geometric constraints of the iterative update of the preset separation matrix are the constraints created based on the wave arrival direction information, and finally the frequency domain separation signal is converted into a time domain signal to obtain a time domain separation signal. The embodiment of the present application converts the time domain observation signal into the frequency domain observation signal, and then uses the independent vector analysis algorithm based on the auxiliary function to perform blind source separation on the frequency domain observation signal, and uses the wave direction information as the geometric constraint for iterative update. Through the above method, not only the efficiency of blind source separation of speech signals can be improved, but also the accuracy of blind source separation can be significantly improved. In addition, it can be understood that the IVA algorithm has a strong anti-noise ability, and can well separate the source signal even in an environment with a certain amount of noise, and can solve the frequency arrangement problem caused by the overall modeling of the signal. In addition, IVA can also handle non-stationary signals well and adapt to signal changes by dynamically updating model parameters. By combining the auxiliary function with the independent vector analysis algorithm, this solution can reduce the amount of calculation and is suitable for deployment on devices with limited computing resources (such as edge devices).

[0078] Correspondingly, the present application also discloses a speech signal blind source separation device, see Figure 3 As shown, the device comprises:

[0079] The acquisition module 11 is used to acquire mixed sound signals of a second number of source signals through a first number of sound acquisition devices respectively, to obtain the first number of time domain observation signals;

[0080] A frequency domain conversion module 12 is used to convert the time domain observation signal into a frequency domain signal to obtain a frequency domain observation signal, and to create a frequency domain separation signal model based on the frequency domain observation signal and a preset separation matrix;

[0081] An acquisition module 13 is used to acquire the direction of arrival of the source signal to obtain direction of arrival information;

[0082] A blind source separation module 14 is used to iteratively update a preset separation matrix in the frequency domain separation signal model based on an independent vector analysis algorithm of an auxiliary function, so as to perform blind source separation on the frequency domain observation signal in the frequency domain separation signal model to obtain a frequency domain separation signal; the geometric constraints for iteratively updating the preset separation matrix are constraints created based on the wave direction information;

[0083] The time domain conversion module 15 is used to convert the frequency domain separation signal into a time domain signal to obtain a time domain separation signal.

[0084] Among them, the specific working processes of the above-mentioned modules can refer to the corresponding contents disclosed in the aforementioned embodiments, which will not be repeated here.

[0085] It can be seen that in the embodiment of the present application, the mixed sound signals of the second number of source signals are first collected by a first number of sound collection devices respectively to obtain the first number of time domain observation signals, and then the time domain observation signals are converted into frequency domain signals to obtain frequency domain observation signals, and a frequency domain separation signal model is created based on the frequency domain observation signals and the preset separation matrix, and then the arrival direction of the source signal is obtained to obtain the arrival direction information, and the preset separation matrix in the frequency domain separation signal model is iteratively updated based on the independent vector analysis algorithm of the auxiliary function to perform blind source separation on the frequency domain observation signals in the frequency domain separation signal model to obtain a frequency domain separation signal, the geometric constraints of the iterative update of the preset separation matrix are the constraints created based on the arrival direction information, and finally the frequency domain separation signal is converted into a time domain signal to obtain a time domain separation signal. The embodiment of the present application converts the time domain observation signal into the frequency domain observation signal, and then uses the independent vector analysis algorithm based on the auxiliary function to perform blind source separation on the frequency domain observation signal, and uses the wave direction information as the geometric constraint for iterative update. Through the above method, not only the efficiency of blind source separation of speech signals can be improved, but also the accuracy of blind source separation can be significantly improved. At the same time, it is suitable for deployment on devices with limited computing resources.

[0086] In some specific embodiments, the frequency domain conversion module 12 may specifically include:

[0087] The Fourier transform unit is used to perform short-time Fourier transform on the time domain observation signal to obtain a frequency domain observation signal.

[0088] In some specific embodiments, the time domain conversion module 15 may specifically include:

[0089] A scale recovery unit, used for recovering the scale of the frequency domain separated signal by back-projection to obtain a recovered separated signal;

[0090] The inverse Fourier transform unit is used to perform an inverse short-time Fourier transform on the recovered separated signal to obtain a time-domain separated signal.

[0091] In some specific embodiments, the blind source separation module 14 may specifically include:

[0092] A first preprocessing unit, used for preprocessing the preset separation matrix and the frequency domain observation signal in the frequency domain separation signal model to obtain a processed separation matrix and a processed observation signal;

[0093] The first iterative updating unit is used to iteratively update the processed separation matrix based on an independent vector analysis algorithm of an auxiliary function, so as to perform blind source separation on the processed observation signal to obtain a frequency domain separation signal.

[0094] In some specific embodiments, the first pre-processing unit may specifically include:

[0095] The second preprocessing unit is used to perform averaging and whitening on the preset separation matrix and the frequency domain observation signal in the frequency domain separation signal model, respectively, to obtain a processed separation matrix and a processed observation signal; wherein the whitening process adopts a PCA dimensionality reduction method.

[0096] In some specific embodiments, the first iterative updating unit may specifically include:

[0097] A second iterative updating unit is used to iteratively update the processed separation matrix in an iterative source-guided updating manner and by using an auxiliary function in an independent vector analysis algorithm based on an auxiliary function to obtain an updated separation matrix;

[0098] A judging unit, used to judge whether the updated separation matrix converges during the iterative updating process;

[0099] The blind source separation unit is used to perform blind source separation on the processed observation signal frequency by frequency point based on the updated separation matrix and using the independent vector analysis algorithm in an online source-guided update manner if the updated separation matrix converges, so as to obtain a frequency domain separation signal.

[0100] In some specific embodiments, the acquisition module 13 may specifically include:

[0101] The fitting unit is used to fit the direction of arrival of the source signal based on a positioning algorithm to obtain direction of arrival information.

[0102] Furthermore, the present application also discloses an electronic device. Figure 4 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content in the diagram cannot be regarded as any limitation on the scope of use of the present application.

[0103] Figure 4 A schematic diagram of the structure of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the method for blind source separation of speech signals disclosed in any of the aforementioned embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0104] In this embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device, and the communication protocol it follows is any communication protocol that can be applied to the technical solution of the present application, and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs and is not specifically limited here.

[0105] In addition, the memory 22, as a carrier for storing resources, can be a read-only memory, a random access memory, a disk or an optical disk, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0106] The operating system 221 is used to manage and control the hardware devices and computer program 222 on the electronic device 20, which can be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program that can be used to complete the speech signal blind source separation method performed by the electronic device 20 disclosed in any of the aforementioned embodiments, the computer program 222 can further include a computer program that can be used to complete other specific tasks.

[0107] Furthermore, the present application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the aforementioned blind source separation method for speech signals is implemented. The specific steps of the method can be referred to the corresponding contents disclosed in the aforementioned embodiments, and will not be repeated here.

[0108] Furthermore, an embodiment of the present application also discloses a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the speech signal blind source separation method disclosed above.

[0109] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0110] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0111] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0112] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0113] The above is a detailed introduction to a speech signal blind source separation method, device, equipment and storage medium provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for blind source separation of speech signals, characterized in that: include: Collecting mixed sound signals of the second number of source signals by using a first number of sound collection devices respectively to obtain the first number of time domain observation signals; Convert the time domain observation signal into a frequency domain signal to obtain a frequency domain observation signal, and create a frequency domain separation signal model based on the frequency domain observation signal and a preset separation matrix; Acquire the direction of arrival of the source signal to obtain direction of arrival information, and iteratively update the preset separation matrix in the frequency domain separation signal model based on the independent vector analysis algorithm of the auxiliary function to perform blind source separation on the frequency domain observation signal in the frequency domain separation signal model to obtain a frequency domain separation signal; the geometric constraints for iteratively updating the preset separation matrix are constraints created based on the direction of arrival information; The frequency domain separation signal is converted into a time domain signal to obtain a time domain separation signal.

2. The method for blind source separation of speech signals according to claim 1, characterized in that: The step of converting the time domain observation signal into a frequency domain signal to obtain a frequency domain observation signal comprises: Performing short-time Fourier transform on the time-domain observation signal to obtain a frequency-domain observation signal.

3. The method for blind source separation of speech signals according to claim 2, characterized in that: The step of converting the frequency domain separated signal into a time domain signal to obtain a time domain separated signal comprises: The frequency domain separation signal is scaled restored by back projection to obtain a restored separation signal, and the restored separation signal is inversely short-time Fourier transformed to obtain a time domain separation signal.

4. The method for blind source separation of speech signals according to claim 1, characterized in that: The auxiliary function-based independent vector analysis algorithm iteratively updates the preset separation matrix in the frequency domain separation signal model to perform blind source separation on the frequency domain observation signal in the frequency domain separation signal model to obtain the frequency domain separation signal, including: Preprocessing the preset separation matrix and the frequency domain observation signal in the frequency domain separation signal model to obtain a processed separation matrix and a processed observation signal; The processed separation matrix is ​​iteratively updated based on an independent vector analysis algorithm of an auxiliary function to perform blind source separation on the processed observation signal to obtain a frequency domain separation signal.

5. The method for blind source separation of speech signals according to claim 4, characterized in that: The preprocessing of the preset separation matrix and the frequency domain observation signal in the frequency domain separation signal model to obtain the processed separation matrix and the processed observation signal includes: The preset separation matrix and the frequency domain observation signal in the frequency domain separation signal model are respectively subjected to averaging processing and whitening processing to obtain a processed separation matrix and a processed observation signal; wherein the whitening processing adopts a PCA dimensionality reduction method.

6. The method for blind source separation of speech signals according to claim 4, characterized in that: The auxiliary function-based independent vector analysis algorithm iteratively updates the processed separation matrix to perform blind source separation on the processed observation signal to obtain a frequency domain separation signal, including: Iteratively updating the processed separation matrix in an iterative source-guided update manner and using an auxiliary function in an independent vector analysis algorithm based on an auxiliary function to obtain an updated separation matrix, and determining whether the updated separation matrix converges during the iterative update process; If the updated separation matrix converges, blind source separation is performed on the processed observation signal frequency-by-frequency point based on the updated separation matrix in an online source-guided updating manner and using the independent vector analysis algorithm to obtain a frequency domain separated signal.

7. The method for blind source separation of speech signals according to any one of claims 1 to 6, characterized in that: The acquiring the direction of arrival of the source signal to obtain direction of arrival information includes: The direction of arrival of the source signal is fitted based on a positioning algorithm to obtain direction of arrival information.

8. A speech signal blind source separation device, characterized in that: include: A collection module, used for collecting mixed sound signals of a second number of source signals through a first number of sound collection devices respectively, to obtain the first number of time domain observation signals; A frequency domain conversion module, used for converting the time domain observation signal into a frequency domain signal to obtain a frequency domain observation signal, and creating a frequency domain separation signal model based on the frequency domain observation signal and a preset separation matrix; An acquisition module, used to acquire the direction of arrival of the source signal to obtain direction of arrival information; A blind source separation module, used for iteratively updating a preset separation matrix in the frequency domain separation signal model based on an independent vector analysis algorithm of an auxiliary function, so as to perform blind source separation on the frequency domain observation signal in the frequency domain separation signal model to obtain a frequency domain separation signal; the geometric constraints for iteratively updating the preset separation matrix are constraints created based on the wave direction information; The time domain conversion module is used to convert the frequency domain separation signal into a time domain signal to obtain a time domain separation signal.

9. An electronic device, characterized in that: It comprises a processor and a memory; wherein, when the processor executes the computer program stored in the memory, the method for blind source separation of speech signals as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: Used to store computer programs; wherein, when the computer program is executed by a processor, the method for blind source separation of speech signals as described in any one of claims 1 to 7 is implemented.

Images