Method for designing signal of reference loop in echo cancellation AEC signal processing

Abstract

The invention relates to the technical field of acoustic echo cancellation methods, in particular to a method for designing a signal of a reference loop in echo cancellation AEC signal processing, which comprises a multi-layer recursion collaborative echo separation module used for receiving a mixed signal, separating an echo signal in the mixed signal from a target voice signal through a multi-layer recursion characteristic, and sending the separated echo signal to a multi-layer recursion collaborative echo separation module; the method comprises the following steps: carrying out feature separation on the basis of a mixed signal to obtain preliminarily separated target signal estimation; constructing an echo path fitting model to represent echo path distribution, and calculating a residual signal; and recursive feedback optimization is carried out based on the residual signal to obtain an updated path fitting model, so that accurate elimination of the echo signal is realized, and accurate separation of the target voice signal and the echo signal is realized through a layer-by-layer progressive structure of feature separation, path fitting and feedback optimization by adopting a multi-layer recursive collaborative echo separation algorithm.

Description

Technical Field

[0001] The invention relates to the technical field of acoustic echo cancellation methods, and in particular to a design method for a reference loop signal in echo cancellation AEC signal processing. Background Art

[0002] With the rapid development of voice-controlled devices such as smart speakers and smart assistants, users have increasing requirements for voice recognition and interactive experience of devices. However, in actual use scenarios, smart speakers usually have a more complex acoustic environment. The sound played by the speaker will not only be directly transmitted to the user's microphone, but also form echo interference through reflection, which significantly affects the accuracy of voice recognition. In order to improve the user experience, most existing technical solutions rely on acoustic echo cancellation (AEC) technology to eliminate the sound emitted by the speaker itself and the echo in the surrounding environment as much as possible.

[0003] Traditional AEC solutions usually deal with echo cancellation problems based on adaptive filters. In these solutions, a feedback reference signal is established, the echo path function is calculated, and the echo signal is attempted to be modeled and eliminated. These methods can solve basic echo interference to a certain extent, but there are still several significant technical difficulties, especially in complex multi-path reflection environments. For example, due to the fixed positions of speakers and microphones and the complexity of environmental sound reflections, traditional echo models often cannot accurately fit the actual acoustic propagation path. In addition, the construction of feedback reference signals is usually limited to electrical signals, ignoring the characteristics of the acoustic path, resulting in the model being unable to effectively describe the multiple reflected sounds in the environment, thereby affecting the accuracy and convergence speed of echo cancellation.

[0004] Another common problem in existing adaptive filtering methods is that since filter parameters usually need to rely on a fixed step size for updating, the convergence speed is slow in dynamic environments, and echo signals are difficult to suppress quickly and effectively. At the same time, traditional methods only process in the electrical signal domain, without combining the electrical signal and acoustic signal domains, so it is difficult to adapt to complex acoustic scenarios. In addition, current echo cancellation models mostly use single-layer filtering or simple multi-stage processing, lacking an effective recursive optimization structure, and it is difficult to effectively isolate and suppress environmental echoes in multipath environments, resulting in unsatisfactory voice signal extraction.

[0005] The existence of these technical problems makes it difficult for traditional AEC solutions to meet the high-precision requirements of smart speakers for echo cancellation in scenarios such as multiple echoes and complex acoustic environments. These defects indicate a key technical requirement, which is to design an AEC echo cancellation solution that can more efficiently adapt to complex multi-path echo propagation environments. The solution should be able to combine the characteristics of electrical and acoustic signals, and gradually improve the accuracy and convergence speed of echo suppression through an adaptive multi-layer recursive optimization mechanism. Summary of the invention

[0006] Based on the limitations of the prior art, the present invention proposes a novel design method for a reference loop signal in the signal processing of echo cancellation to solve the shortcomings of the traditional method in a multi-path complex echo environment.

[0007] The present invention provides a design method for a reference loop signal in echo cancellation AEC signal processing, including a multi-layer recursive cooperative echo separation module, which is used to receive a mixed signal, separate an echo signal from a target voice signal in the mixed signal through a multi-layer recursive characteristic, and obtain an echo cancellation signal. The method includes:

[0008] Perform feature separation based on the mixed signal to obtain a preliminary separated target signal estimate;

[0009] constructing an echo path fitting model to represent the echo path distribution, and calculating a residual signal;

[0010] Recursive feedback optimization is performed based on the residual signal to obtain an updated path fitting model, thereby achieving accurate elimination of the echo signal.

[0011] Preferably, the feature separation comprises:

[0012] Receive a mixed signal M(t), where M(t) is the mixed signal at time t, including the target speech signal S(t) and the echo signal R(t), that is, M(t)=S(t)+R(t);

[0013] Through the feature separation matrix F = {f ij} Decompose the mixed signal to obtain a preliminary separated target signal estimate and echo characteristic signal R i (t), where:

[0014] f ij is the separation coefficient of the characteristic separation matrix; the signal separation process is implemented by the following signal separation equation: in is the initial separated target signal estimation, R i (t) is the characteristic signal of the i-th echo.

[0015] Preferably, the echo path fitting model construction includes:

[0016] The echo path distribution function P(t) is defined to represent the echo signal distribution characteristics of multipath reflection; where P(t) = {p i (t)},p i (t) is the echo distribution function of the ith path at time t;

[0017] The path residual ΔP(t)=R(t)-P(t) is calculated through the path fitting model to obtain the difference between the current echo signal and the path model.

[0018] Preferably, the recursive optimization of the echo path fitting model comprises:

[0019] Update the coefficient matrix of the path fitting model C = {c ij}, so that the echo path distribution function conforms to the actual signal distribution;

[0020] where c ij is the path matching coefficient, which represents the adjustment parameter of the path model. Through recursive optimization, the path distribution function satisfies

[0021]

[0022] The path fitting model is adjusted in feedback optimization to minimize the path residual.

[0023] Preferably, the feedback optimization comprises:

[0024] Define feedback error E(t) = ΔM(t) - P(t), where ΔM(t) is the updated residual signal, which is used to describe the error between the path model and the actual signal;

[0025] Setting a step size control parameter α, and updating the echo path distribution function based on the feedback error and the step size parameter so that P(t)=P(t)+α·E(t);

[0026] The feedback optimization process achieves recursive minimization of the error by adjusting the step size control parameter until E(t) converges to a preset threshold.

[0027] Preferably, the multi-layer recursive collaborative echo separation module includes the following data streams:

[0028] Obtain target speech signal estimation and echo signal features of the mixed signal by feature separation;

[0029] inputting the echo signal characteristics into an echo path fitting model to determine a path residual;

[0030] In the feedback optimization process, the path model is continuously updated according to the path residual to obtain the final echo cancellation signal.

[0031] Preferably, the data stream further comprises:

[0032] The signal acquisition module acquires and provides mixed signal input;

[0033] The synergistic effect of feature separation, path fitting and feedback optimization is achieved through recursive separation and distribution fitting;

[0034] The final signal output module provides an optimized echo-cancelled signal for further processing or output.

[0035] Preferably, M(t) is a mixed signal; F is a feature separation matrix; f ij is the coefficient of the signal separation matrix; S(t) is the target speech signal; R(t) is the echo signal; is the target signal estimation for preliminary separation; R i (t) is the characteristic signal of the i-th echo, P(t) is the echo path distribution function; i (t) is the echo distribution function of the ith path; C is the path fitting coefficient matrix; c ij is the path matching coefficient; E(t) is the feedback error signal; ΔM(t) is the updated residual signal; α is the step size control parameter.

[0036] Preferably, the recursive optimization and feedback optimization of the path fitting model are implemented through a multi-layer recursive relationship, wherein:

[0037] The target speech signal estimate and echo path model output obtained initially are used in the subsequent feedback optimization process;

[0038] The feedback optimization result is further used to iteratively adjust the coefficient matrix of feature separation and path fitting, so that the recursive optimization continuously enhances the signal separation effect and finally obtains a high-precision echo cancellation signal.

[0039] Preferably, the method is used in an echo cancellation system of an intelligent speaker device, and efficient separation of signal features and echo cancellation are achieved in mixed signal processing through the multi-layer recursive collaborative echo separation module.

[0040] The beneficial effects of the present invention are:

[0041] The present invention adopts a multi-layer recursive collaborative echo separation algorithm, and realizes the accurate separation of the target speech signal and the echo signal through the progressive structure of feature separation, path fitting and feedback optimization. Specifically, in the feature separation process, the present invention extracts the preliminary target signal and echo feature signal in the mixed signal through the feature separation matrix, providing an accurate data basis for subsequent path fitting. In the path fitting layer, the echo path distribution model is further constructed, and the path residual is calculated. The fitting accuracy is adjusted through recursive optimization, thereby solving the technical problem that the traditional echo path model cannot fully fit the multiple reflection paths.

[0042] The algorithm mechanism of the present invention also fully considers the synergistic relationship between electrical signals and acoustic signals. On the basis of echo signal path fitting, the accurate matching of the path model is achieved through the dynamic update of feedback error. Compared with the prior art, the multi-layer recursive collaborative structure of the present invention can not only improve the convergence speed of echo cancellation, but also significantly enhance the adaptability of the system to complex acoustic environments, making the echo cancellation effect more significant.

[0043] Through such a progressive algorithm design, the present invention can achieve accurate speech signal extraction in a multi-path complex echo environment, overcoming the shortcomings of traditional solutions in terms of convergence speed and echo cancellation effect. The various layers of algorithms achieve complementary and superimposed effects in structure. The synergistic effect of the path fitting layer and the feedback optimization layer enhances the anti-interference ability of the system, while the multi-layer progressive mechanism of feedback optimization further improves the accuracy of signal separation. The present invention effectively solves the incoordination problem of traditional technologies in electrical and acoustic signal processing, realizes a more accurate and efficient echo cancellation function, and provides a new technical solution for the application of smart speakers and other devices in complex acoustic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is the overall system flow chart of the present invention. DETAILED DESCRIPTION

[0045] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0046] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0047] The technical solution of the present invention is described in detail below in conjunction with specific embodiments.

[0048] Please refer to Figure 1 The present invention relates to a design method for a reference loop signal in the signal processing of echo cancellation, which processes a mixed signal through a multi-layer recursive collaborative echo separation algorithm to obtain effective echo cancellation. In the specific implementation, the present invention adopts a hierarchical progressive structure, from feature separation, echo path fitting to feedback optimization, which is progressive layer by layer to ensure efficient separation and accurate elimination of echo signals.

[0049] The present invention provides a signal processing method including a multi-layer recursive cooperative echo separation module. The module receives a mixed signal, applies a recursive cooperative algorithm, separates the echo signal and the target speech signal in the mixed signal, and thus obtains an echo cancellation signal with high precision. The steps included in this method include: obtaining a preliminary separated target signal estimate based on the feature separation of the mixed signal; then describing the distribution characteristics of the echo path by constructing an echo path fitting model, and further calculating the residual signal; finally, obtaining an updated path fitting model through a feedback optimization process based on the residual signal to achieve accurate echo signal cancellation.

[0050] Preferably, the mixed signal is an acoustic input signal collected by a microphone. In one embodiment of the present invention, the collection frequency can be set to 16kHz to ensure the accuracy and integrity of the collected signal. The mixed signal includes a target speech signal and an environmental echo signal. The preliminary separated target signal is obtained by a feature separation method, which is helpful for the subsequent accurate extraction and separation of the echo signal.

[0051] In a preferred embodiment of the present invention, the present invention further includes a feature separation process. The process mainly decomposes the mixed signal based on the feature separation matrix to obtain a preliminary separated target signal estimate. The feature separation matrix F = {f ij} contains the signal separation factor f ij , which is used to control the separation strength between different signals. Through matrix operation, the target speech signal S(t) and the multipath echo signal R(t) can be extracted from the mixed signal M(t) to achieve the initial separation of the signals. The specific equation of this separation process is in is the target signal estimation after preliminary separation, R i (t) is the characteristic signal of the i-th echo. Preferably, the matrix coefficient f ij The initial value of can be set between 0.1 and 0.9 to adapt to different echo environments and improve separation accuracy.

[0052] In practical applications, such as in the scenario of smart speakers, by setting different separation coefficients, it is possible to effectively separate the sound emitted by the speaker and other interfering sounds in the environment, thereby improving the quality of the voice signal and enhancing the user experience.

[0053] In a preferred embodiment of the present invention, the present invention includes a step of constructing an echo path fitting model to further accurately model the echo signal. In the model, the echo path distribution function P(t)={p i (t)} represents the propagation and reflection characteristics of the echo signal on different paths. i(t) represents the echo distribution function of the ith path at time t. Through this model, the actual echo signal can be better fitted in the multiple reflection path.

[0054] Furthermore, the calculation of the echo path residual ΔP(t)=R(t)-P(t) is used to determine the fitting error, providing a basis for subsequent feedback optimization. Preferably, the residual threshold used in the model can be set to 0.01 to balance the fitting accuracy and computational complexity of the model.

[0055] In the application scenario of smart speakers, the echo path fitting model can effectively describe the propagation characteristics of the sound waves emitted by the speaker through the reflection paths such as the room walls and ceiling, so that the final echo signal is closer to the actual situation, which is conducive to achieving accurate echo cancellation.

[0056] In a preferred embodiment of the present invention, the present invention further includes a recursive optimization of the echo path fitting model. The recursive optimization process updates the coefficient matrix C of the path fitting model by ij} Implement adaptive adjustment of the echo path model to ensure the matching accuracy between the model and the actual echo signal distribution. The adjustment formula of the path fitting model is: where c ij is the path matching coefficient, which represents the control of the fitting accuracy of each path.

[0057] Preferably, in actual implementation, the path matching coefficient c ij The initial value of can be set to 0.5 to ensure the stability of the model, and with the recursive iteration of feedback optimization, the coefficients are gradually optimized to reduce the path residual. The number of iterations of this recursive process can be set between 5 and 10 times, depending on the complexity of the actual application environment and the requirements for echo cancellation accuracy.

[0058] This optimization process is of great significance in smart speakers. By recursively updating the coefficient matrix, the model can be continuously adaptively adjusted to improve the fitting accuracy of the echo signal emitted by the speaker, thereby achieving higher quality speech separation.

[0059] In a preferred embodiment of the present invention, the present invention further describes the details of the feedback optimization process. The feedback optimization judges the deviation between the current model and the actual echo path by introducing the feedback error E(t)=ΔM(t)-P(t), where ΔM(t) is the updated residual signal, which is used to describe the fitting error of the path model. The feedback optimization sets the step control parameter α and adjusts the echo path distribution function P(t) based on the feedback error to achieve the purpose of minimizing the path residual. The update process follows the formula P(t)=P(t)+α·E(t), where the value of α is preferably between 0.05 and 0.2, and the specific value is adjusted according to different noise environments and signal complexity to ensure the stability of the feedback optimization process.

[0060] Furthermore, in order to achieve rapid convergence, the convergence threshold of the feedback error is preferably set to 0.001, which ensures the optimal fitting accuracy under a reasonable amount of calculation. This optimization can significantly improve the adaptability to the echo path in a complex environment in the smart speaker scenario, making the elimination of the echo signal more accurate and improving the performance of the overall system.

[0061] Through the above description, the present invention demonstrates efficient signal separation and optimization characteristics in the echo cancellation application scenario of the smart speaker, is suitable for multi-path echo propagation environment, can effectively improve the clarity and accuracy of voice interaction, and further optimize the user experience.

[0062] In a preferred embodiment of the present invention, the echo cancellation method of the present invention includes a data flow design of a multi-layer recursive collaborative echo separation module. The data flow forms a closed loop between signal acquisition, feature separation, path fitting and feedback optimization, ensuring that the signal is gradually optimized during the processing process to achieve the purpose of efficient separation and echo cancellation.

[0063] Specifically, the starting point of the data stream is the signal acquisition module 1, which obtains the initial mixed signal M(t) through a microphone and other devices. The acquisition frequency of the mixed signal is preferably 16kHz to ensure that the signal contains sufficient frequency domain information. After receiving the mixed signal, the data stream enters the feature separation module 1, which performs a preliminary separation of the target speech signal S(t) and the echo signal R(t) in the signal through the feature separation matrix F to obtain the estimated target signal

[0064] Subsequently, the echo characteristic signal R(t) is used as the input of the path fitting model and is passed to the echo path fitting module 3 to construct the echo path distribution function P(t), and calculate the path residual ΔP(t) to judge the model fitting accuracy. Finally, the residual signal ΔM(t) is fed back to the feedback optimization module 4, which recursively optimizes the path distribution function by controlling the step size α, adjusts and reduces the residual until the echo signal is effectively removed.

[0065] Through such a data flow structure, the signal flows in sequence between the processing modules and feeds back to each other, so that the system can gradually converge in the recursive optimization, significantly improving the final signal separation accuracy and achieving a higher echo cancellation effect.

[0066] In a preferred embodiment of the present invention, the present invention further optimizes the process and role of data flow in the multi-layer recursive collaborative echo separation module, ensuring the collaborative relationship between signal acquisition, feature separation, path fitting and feedback optimization. The signal acquisition module 1 is responsible for capturing and providing mixed signal input, and the feature separation module 2 realizes the preliminary separation of the signal through matrix decomposition. On this basis, the path fitting module 3 applies the fitting model, calculates the echo path residual, and forms a collaborative closed loop with the feedback optimization module 4.

[0067] In the feedback optimization process, after each recursive optimization, the signal is sent back to the feature separation module for further separation, so that the separation effect of the model is gradually enhanced, and finally accurate echo cancellation is achieved in multiple recursions. This data flow design achieves a higher synergy between the modules through feedback control, ensuring that the signal processing results are more accurate. Preferably, the number of recursions of feedback optimization is set between 5 and 15 times, and the specific value can be adjusted according to the complexity of the environment around the speaker to meet the echo cancellation requirements of different usage scenarios.

[0068] In a preferred embodiment of the present invention, the present invention defines in detail all symbols involved in the echo cancellation algorithm and their specific meanings to ensure the accuracy and consistency of each calculation formula in the application. Specifically, the following are the definitions of each symbol:

[0069] M(t) represents the mixed signal, which is the original input signal collected by the microphone at time t; S(t) is the target speech signal, which refers to the speech component that needs to be extracted from the mixed signal; R(t) is the echo signal, which represents the interference signal generated by the environmental reflection; is the target signal estimation for preliminary separation, F is the feature separation matrix, which is used to separate the target signal and echo signal in the mixed signal, and f ij is the coefficient of the signal separation matrix F, which represents the weight R of each component in the separation process i (t) represents the characteristic signal of the i-th echo; P(t) is the echo path distribution function, which is used to represent the echo distribution of different reflection paths; pi (t) is the echo distribution function of the ith path at time t; C is the path fitting coefficient matrix, which is used to control the accuracy of the path fitting model; c ij is the path matching coefficient, which describes the influence weight of each path on the echo distribution; E(t) is the feedback error signal, which indicates the error between the path model and the actual signal; ΔM(t) is the updated residual signal, which is used to describe the residual of the current fitting model; α is the step size control parameter, which is used for signal adjustment in feedback optimization to ensure convergence.

[0070] By defining these symbols, the present invention ensures the consistency of algorithm parameters in multiple recursive iterations, and ensures that the data flow in the path fitting and error feedback process is accurate and reasonable.

[0071] In a preferred embodiment of the present invention, a multi-layer progressive relationship is established in the recursive optimization and feedback optimization process of the path fitting model. First, the target speech signal estimation is initially obtained. The echo path model output P(t) will enter the feedback optimization module, and the path model will be recursively adjusted through the feedback error E(t), so that the path residual is gradually reduced.

[0072] The progressive relationship ensures that the cyclic iterations of feature separation, path fitting and feedback optimization reinforce each other. Specifically, the output signal of each feedback optimization will return to the feature separation module to further adjust the separation coefficient matrix F, and the path fitting module will use the updated path matching coefficient matrix C for further optimization, thereby achieving higher signal separation accuracy in multiple recursions.

[0073] Preferably, the convergence criterion of feedback optimization is set to an error threshold of 0.001 to ensure the best separation effect within a reasonable computing time. This progressive relationship further ensures that multipath echo signals in complex environments can be accurately separated and eliminated, providing an efficient signal processing solution for echo-sensitive applications such as smart speakers.

[0074] In a preferred embodiment of the present invention, the present invention applies the signal processing method of the multi-layer recursive collaborative echo separation module to the echo cancellation system of the smart speaker device. The method combines feature separation, path fitting and feedback optimization technology, and realizes efficient separation of signal features and echo cancellation in the smart speaker device through the multi-layer recursive collaborative algorithm. Since the distance between the user and the smart speaker is relatively fixed, and there are interferences such as direct sound, reflected sound, and environmental noise emitted by the speaker, the algorithm structure of the present invention has obvious advantages in this specific scenario.

[0075] Specifically, the microphone in the smart speaker collects the mixed signal M(t), which can effectively separate the echo signal R(t) emitted by the speaker itself and the user voice signal S(t). Utilizing the adaptive characteristics of the feedback optimization module, the system can adapt to changes in the environment around the smart speaker, and can operate stably and provide high-quality voice interaction experience whether in an open space or a closed room.

[0076] The present invention not only improves the voice recognition accuracy of smart speakers in complex acoustic environments, but also reduces the interference of the system in different usage scenarios and optimizes the user experience. Therefore, through the multi-layer recursive collaborative echo separation method of the present invention, the smart speaker can efficiently separate echo signals in any environment and improve the overall performance and reliability of the system.

[0077] It should be noted that 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, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for designing a reference loop signal in signal processing for echo cancellation, characterized in that: The method comprises a multi-layer recursive cooperative echo separation module, which is used to receive a mixed signal, separate the echo signal from the target speech signal in the mixed signal through the multi-layer recursive characteristics, and obtain an echo cancellation signal. The method comprises: Perform feature separation based on the mixed signal to obtain a preliminary separated target signal estimate; constructing an echo path fitting model to represent the echo path distribution, and calculating a residual signal; Recursive feedback optimization is performed based on the residual signal to obtain an updated path fitting model, thereby achieving accurate elimination of the echo signal.

2. The method for designing a reference loop signal in the signal processing of echo cancellation according to claim 1, characterized in that: The feature separation includes: Receive a mixed signal M(t), where M(t) is the mixed signal at time t, including the target speech signal S(t) and the echo signal R(t), that is, M(t)=S(t)+R(t); Through the feature separation matrix F = {f ij } Decompose the mixed signal to obtain a preliminary separated target signal estimate and echo characteristic signal R i (t), where: f ij is the separation coefficient of the characteristic separation matrix; the signal separation process is implemented by the following signal separation equation: in is the initial separated target signal estimation, R i (t) is the characteristic signal of the i-th echo.

3. The method for designing a reference loop signal in the signal processing of echo cancellation according to claim 2, characterized in that: The echo path fitting model construction includes: The echo path distribution function P(t) is defined to represent the echo signal distribution characteristics of multipath reflection; where P(t) = {p i (t)},p i (t) is the echo distribution function of the ith path at time t; The path residual ΔP(t)=R(t)-P(t) is calculated through the path fitting model to obtain the difference between the current echo signal and the path model.

4. The method for designing a reference loop signal in the signal processing of echo cancellation according to claim 3, characterized in that: The recursive optimization of the echo path fitting model includes: Update the coefficient matrix of the path fitting model C = {c ij }, so that the echo path distribution function conforms to the actual signal distribution; where c ij is the path matching coefficient, which represents the adjustment parameter of the path model. Through recursive optimization, the path distribution function satisfies The path fitting model is adjusted in feedback optimization to minimize the path residual.

5. The method for designing a reference loop signal in the signal processing of echo cancellation according to claim 4, characterized in that: The feedback optimization includes: Define feedback error E(t) = ΔM(t) - P(t), where ΔM(t) is the updated residual signal, which is used to describe the error between the path model and the actual signal; Setting a step size control parameter α, and updating the echo path distribution function based on the feedback error and the step size parameter so that P(t)=P(t)+α·E(t); The feedback optimization process achieves recursive minimization of the error by adjusting the step size control parameter until E(t) converges to a preset threshold.

6. The method for designing a reference loop signal in the signal processing of echo cancellation according to claim 5, characterized in that: The multi-layer recursive co-echo separation module includes the following data streams: Obtain target speech signal estimation and echo signal features of the mixed signal by feature separation; inputting the echo signal characteristics into an echo path fitting model to determine a path residual; In the feedback optimization process, the path model is continuously updated according to the path residual to obtain the final echo cancellation signal.

7. The method for designing a reference loop signal in the signal processing of echo cancellation according to claim 6, characterized in that: The data stream further includes: The signal acquisition module acquires and provides mixed signal input; The synergistic effect of feature separation, path fitting and feedback optimization is achieved through recursive separation and distribution fitting; The final signal output module provides an optimized echo-cancelled signal for further processing or output.

8. The method for designing a reference loop signal in the signal processing of echo cancellation according to claim 1, characterized in that: M(t) is the mixed signal; F is the feature separation matrix; f ij is the coefficient of the signal separation matrix; S(t) is the target speech signal; R(t) is the echo signal; is the target signal estimation for preliminary separation; R i 9t) is the characteristic signal of the i-th echo, P(t) is the echo path distribution function; p i (t) is the echo distribution function of the ith path; C is the path fitting coefficient matrix; c ij is the path matching coefficient; E(t is the feedback error signal; ΔM(t) is the updated residual signal; α is the step size control parameter.

9. The method for designing a reference loop signal in the signal processing of echo cancellation according to claim 8, characterized in that: The recursive optimization and feedback optimization of the path fitting model are achieved through a multi-layer recursive relationship, where: The target speech signal estimate and echo path model output obtained initially are used in the subsequent feedback optimization process; The feedback optimization result is further used to iteratively adjust the coefficient matrix of feature separation and path fitting, so that the recursive optimization continuously enhances the signal separation effect and finally obtains a high-precision echo cancellation signal.

10. The method for designing a reference loop signal in the signal processing of echo cancellation according to any one of claims 1 to 9, characterized in that: The method is used in an echo cancellation system of an intelligent speaker device, and the multi-layer recursive collaborative echo separation module is used to realize efficient separation of signal features and echo cancellation in mixed signal processing.

Images