An artificial intelligence echo cancellation method, device, equipment, medium and product

Abstract

The application provides an artificial intelligence echo cancellation method, device, equipment, medium and product, and relates to the field of acoustic echo cancellation. The method comprises the following steps: preprocessing a training signal set to obtain preprocessed data; training network parameters according to a multi-loss function training strategy, a first loss function and a second loss function to obtain an optimized neural network model; detecting a to-be-tested signal set through a double-speaking detection strategy to obtain a target network structure; performing echo cancellation processing according to an echo cancellation strategy to obtain an echo-canceled signal complex spectrum; and performing speech synthesis on the echo-canceled signal complex spectrum through an overlap-add algorithm and an inverse Fourier transform algorithm to obtain an echo-canceled enhanced speech signal. The application solves the technical problem that the prior art is difficult to balance echo suppression capability and near-end speech protection capability, is prone to causing speech distortion or echo residue, and thus causes low artificial intelligence echo cancellation efficiency.

Description

Technical Field

[0001] This application relates to the field of acoustic echo cancellation, and more particularly to an artificial intelligence-based method, apparatus, device, medium, and product for echo cancellation. Background Technology

[0002] In voice communication systems, the far-end signal is transmitted via the network to the near-end, played by the near-end speaker, and then picked up again by the microphone and transmitted to the far-end, forming an acoustic echo. Acoustic echo cancellation technology typically uses adaptive filters for linear filtering to remove components in the echo that are linearly correlated with the far-end signal. Even after adaptive filtering, some echo may still remain, which can be processed using an Artificial Intelligence (AI) echo cancellation module.

[0003] Existing AI echo cancellation technology is a technique that uses artificial intelligence algorithms to eliminate echoes. It can analyze the characteristics of audio signals, identify echo signals, and remove them from the original signal, thereby improving communication quality and user experience.

[0004] However, existing technologies struggle to balance echo suppression and near-end speech protection, which can easily lead to speech distortion or echo residue, resulting in low efficiency of AI-based echo cancellation. Summary of the Invention

[0005] This application provides an artificial intelligence-based echo cancellation method, apparatus, device, medium, and product to address the problem that existing technologies struggle to balance echo suppression and near-end speech protection capabilities, easily leading to speech distortion or echo residue, thus resulting in low efficiency of artificial intelligence-based echo cancellation.

[0006] In a first aspect, this application provides an artificial intelligence-based echo cancellation method, comprising:

[0007] Obtain the set of signals to be trained and the network parameters;

[0008] The training signal set is preprocessed to obtain preprocessed data;

[0009] Based on the preprocessed data, the multi-loss function training strategy, and the first and second loss functions, the network parameters are trained to obtain an optimized neural network model; among them, the first loss function is mainly used to suppress echo, and the second loss function is used to preserve speech.

[0010] A set of signals to be tested is acquired, and a dual-talk detection strategy is used to process the set of signals to be tested in order to obtain the target network structure corresponding to the optimized neural network model; wherein, the target network structure is an echo suppression structure and / or a speech preservation structure.

[0011] Based on the optimized neural network model, target network structure, set of signals to be tested, and echo cancellation strategy, echo cancellation processing is performed to obtain the complex spectrum of the de-echo signal;

[0012] By using an overlap-addition algorithm and an inverse Fourier transform algorithm, speech synthesis is performed on the complex spectrum of the de-echo signal to obtain a de-echo enhanced speech signal.

[0013] In one possible design, the first loss function is dynamic compression loss and / or evaluation metric loss; whereby the evaluation metric loss is used to evaluate speech quality.

[0014] The second loss function is the negative signal-to-noise ratio loss and / or mean square error loss.

[0015] In one possible design, the training signal set includes a far-end reference signal, a main microphone signal, and a near-end speech signal;

[0016] The training signal set is preprocessed to obtain preprocessed data, including:

[0017] An adaptive filtering algorithm is used to remove linear echo from the far-end reference signal and the main microphone signal to obtain a signal with removed linear echo.

[0018] Signal processing is performed on the far-end reference signal and the linear echo removed signal to obtain a spliced ​​complex spectrum; wherein, the spliced ​​complex spectrum is obtained by splicing the complex spectra obtained from the signal processing of the far-end reference signal and the linear echo removed signal respectively;

[0019] Signal processing is performed on the near-end speech signal to obtain the target complex spectrum;

[0020] The spliced ​​complex spectrum is subjected to feature dimensionality reduction processing to obtain compressed features;

[0021] Deep feature extraction is performed on the compressed features to obtain deep features;

[0022] The depth features are decoded using the first decoding structure and the second decoding structure to obtain a set of complex spectral masks; wherein the set of complex spectral masks includes a first complex spectral mask obtained by the first decoding structure and a second complex spectral mask obtained by the second decoding structure.

[0023] In one possible design, the network parameters are trained based on preprocessed data, a multi-loss function training strategy, a first loss function, and a second loss function to obtain an optimized neural network model, including:

[0024] The first loss value of the target complex spectrum and the first complex spectrum mask is calculated using the first loss function;

[0025] The second loss value of the target complex spectrum and the second complex spectrum mask is calculated using the second loss function;

[0026] According to the multi-loss function training strategy, the first loss value and the second loss value are weighted and summed to obtain the final loss score;

[0027] Based on the loss score and loss multi-loss function training strategy, the network parameters are trained to obtain an optimized neural network model.

[0028] In one possible design, a set of signals to be tested is obtained, and a dual-talk detection strategy is used to process the set of signals to be tested in order to obtain the target network structure corresponding to the optimized neural network model, including:

[0029] An adaptive filtering algorithm is used to remove linear echoes from the set of signals to be tested in order to obtain the signal to be detected with the linear echo removed.

[0030] Signal processing is performed on the set of signals to be tested and the signal to be detected after removing linear echoes to obtain a set of complex spectra;

[0031] A dual-talk detection strategy is used to detect and process the complex spectrum set to obtain dual-talk detection tags; the dual-talk detection tags are used to determine whether the set of signals to be tested is a pure echo segment.

[0032] If the set of signals to be tested is a pure echo segment, then the target network structure is an echo-suppressed structure; wherein, the echo-suppressed structure completes the decoding process through the first decoding structure;

[0033] If the set of signals to be tested is a dual-segment network, then the target network structure is a protected speech structure; wherein, the protected speech structure is decoded by a second decoding structure.

[0034] In one possible design, the first decoding structure includes first network parameters, and the second decoding structure includes second network parameters;

[0035] The first network parameters are completely different from the second network parameters and / or only the last layer is different.

[0036] Secondly, this application provides an artificial intelligence-based echo cancellation device, comprising:

[0037] The acquisition module is used to acquire the set of signals to be trained and the network parameters;

[0038] The preprocessing module is used to preprocess the training signal set to obtain preprocessed data;

[0039] The training module is used to train the network parameters based on preprocessed data, a multi-loss function training strategy, a first loss function, and a second loss function to obtain an optimized neural network model; among them, the first loss function is mainly used to suppress echo, and the second loss function is used to preserve speech.

[0040] The detection module is used to acquire the set of signals to be tested and to process the set of signals to be tested using a dual-talk detection strategy to obtain the target network structure corresponding to the optimized neural network model; wherein, the target network structure is an echo suppression structure and / or a speech protection structure;

[0041] The echo cancellation module is used to perform echo cancellation processing based on the optimized neural network model, target network structure, set of signals to be tested, and echo cancellation strategy to obtain the complex spectrum of the de-echo signal;

[0042] The speech synthesis module is used to synthesize speech from the complex spectrum of the echo-de-echo signal using an overlap-addition algorithm and an inverse Fourier transform algorithm, in order to obtain an echo-de-echo enhanced speech signal.

[0043] Thirdly, this application provides an artificial intelligence echo cancellation device, including: a memory and a processor;

[0044] The memory stores the instructions that the computer executes;

[0045] The processor executes computer execution instructions stored in memory, causing the processor to perform the artificial intelligence echo cancellation method as described in the first aspect of the invention.

[0046] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the artificial intelligence echo cancellation method as described in the first aspect of the invention.

[0047] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the artificial intelligence echo cancellation method of the first aspect of the invention.

[0048] This application provides an artificial intelligence-based echo cancellation method, apparatus, device, medium, and product, comprising: acquiring a set of signals to be trained and network parameters; preprocessing the training signal set to obtain preprocessed data; training the network parameters according to the preprocessed data, a multi-loss function training strategy, a first loss function, and a second loss function to obtain an optimized neural network model; acquiring a set of signals to be tested and performing detection processing on the set of signals to be tested using a dual-talk detection strategy to obtain a target network structure corresponding to the optimized neural network model; performing echo cancellation processing according to the optimized neural network model, the target network structure, the set of signals to be tested, and the echo cancellation strategy to obtain a de-echo signal complex spectrum; and performing speech synthesis on the de-echo signal complex spectrum using an overlap-addition algorithm and an inverse Fourier transform algorithm to obtain a de-echo enhanced speech signal. Compared to existing technologies, it is difficult to balance echo suppression capability and near-end speech protection capability, which easily leads to speech distortion or echo residue, resulting in low efficiency of artificial intelligence echo cancellation. This application adopts a loss function that focuses on speech preservation and a loss function that focuses on echo suppression, shares the shallow parameters of the AI ​​network, and combines a two-talk detection module to further suppress the echo of the far-end single talk in the pure echo segment, thereby improving the performance of echo cancellation. At the same time, it increases the speech protection capability in the two-talk segment, thereby improving the efficiency of AI echo cancellation. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A schematic diagram of the system architecture of an artificial intelligence echo cancellation method provided in this application embodiment;

[0051] Figure 2 A flowchart illustrating an artificial intelligence-based echo cancellation method provided in this application embodiment. Figure 1 ;

[0052] Figure 3 A flowchart illustrating an artificial intelligence-based echo cancellation method provided in this application embodiment. Figure 2 ;

[0053] Figure 4 A flowchart illustrating an artificial intelligence-based echo cancellation method provided in this application embodiment. Figure 3 ;

[0054] Figure 5 A flowchart illustrating an artificial intelligence-based echo cancellation method provided in this application embodiment. Figure 4 ;

[0055] Figure 6 This is a schematic diagram of the operation process of an artificial intelligence echo cancellation system provided in an embodiment of this application;

[0056] Figure 7 This is a schematic diagram of the structure of the artificial intelligence echo cancellation device provided in the embodiments of this application;

[0057] Figure 8 This is a schematic diagram of the structure of an artificial intelligence echo cancellation device provided in an embodiment of this application. Detailed Implementation

[0058] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0059] In the embodiments of this application, the terms "first" and "second" are used to distinguish identical or similar items with substantially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, nor do they necessarily imply difference. It should be noted that in the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner. In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more.

[0060] It should be noted that the phrase "at...time" in the embodiments of this application can refer to the instant at which a certain situation occurs, or to a period of time after the occurrence of a certain situation; the embodiments of this application do not specifically limit this. Furthermore, the artificial intelligence echo cancellation method provided in the embodiments of this application is merely an example; artificial intelligence echo cancellation methods may also include more or less content.

[0061] In voice communication systems, the far-end signal is transmitted via the network to the near-end, played by the near-end speaker, and then picked up again by the microphone and transmitted to the far-end, forming an acoustic echo. Acoustic echo cancellation technology typically uses adaptive filters for linear filtering to remove components in the echo that are linearly correlated with the far-end signal. Even after adaptive filtering, some echo may still remain, which can be processed using an AI echo cancellation module.

[0062] In practical applications, AI echo cancellation modules have difficulty balancing echo suppression and near-end speech protection capabilities. If the echo suppression capability is strong, in dual-talk segments where both echo and near-end speech exist, the near-end speech is suppressed, leading to speech distortion. If the speech protection capability is strong, the echo remaining after adaptive filtering cannot be completely eliminated in pure echo segments, resulting in echo residue and poor listening experience during calls.

[0063] The existing technical solutions have the following limitations:

[0064] Optionally, limitations of a single loss function: Existing AI models are typically trained using a single loss function (such as mean squared error or cross-entropy), which makes it difficult to achieve a dynamic balance between pure echo segments (requiring strong suppression) and bilingual segments (requiring speech preservation). For example, if the loss function focuses on suppressing echoes, it may over-suppress near-end speech; conversely, if it focuses on speech preservation, it may not be able to effectively eliminate residual echoes.

[0065] Optionally, there is a lack of dynamic strategy adjustment: existing AI models use fixed network parameters during both training and testing phases, making it impossible to adjust processing strategies based on real-time scenarios (such as two-way speech detection results). For example, in two-way speech, it is necessary to preserve near-end speech, while in pure echo speech, it is necessary to further suppress residual echo, but existing technologies cannot achieve adaptive switching using the same model.

[0066] Optionally, the model's generalization ability is insufficient: existing AI models usually rely on a single task (such as echo cancellation or speech enhancement) as a training objective and lack a multi-task collaborative optimization mechanism, resulting in poor robustness of the model in complex scenarios (such as noise interference and multipath echo).

[0067] To address the aforementioned issues, the inventors, during their research on the low efficiency of AI-based echo cancellation, discovered that existing technologies struggle to balance echo suppression and near-end speech protection, easily leading to speech distortion or echo residue, thus resulting in low AI-based echo cancellation efficiency. Therefore, the inventors considered employing a loss function that emphasizes speech preservation and a loss function that emphasizes echo suppression, sharing shallow parameters of the AI ​​network, and combining this with a dual-talk detection module to further suppress echoes from far-end single-talk segments in pure echo segments, thereby improving echo cancellation performance, while simultaneously enhancing speech protection capabilities in dual-talk segments. Based on this, embodiments of this application provide an AI-based echo cancellation method, apparatus, device, medium, and product, applicable to the field of acoustic echo cancellation, aiming to solve the problem of low efficiency in existing AI-based echo cancellation technologies.

[0068] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0069] Figure 1 This is a schematic diagram of the system architecture for an artificial intelligence echo cancellation method provided in an embodiment of this application. The artificial intelligence echo cancellation system is a computer device. Figure 1 In the above architecture, at least one of data acquisition device 101, processing device 102 and display device 103 is included.

[0070] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the processing system architecture of the artificial intelligence echo cancellation method. In other feasible embodiments of this application, the above architecture may include more or fewer components than illustrated, or combine some components, or divide some components, or arrange different components, which can be determined according to the actual application scenario and is not limited here. Figure 1 The components shown can be implemented in hardware, software, or a combination of both.

[0071] In the specific implementation process, the data acquisition device 101 may include an input / output interface or a communication interface. The data acquisition device 101 can connect to the processing device through the input / output interface or the communication interface to acquire relevant data.

[0072] The processing device 102 can obtain an optimized neural network model and a de-echo-enhanced speech signal through training and testing based on relevant data.

[0073] The display device 103 can also be a touch screen or the screen of a terminal device, used to receive user commands while displaying the above-mentioned content, so as to realize interaction with the user.

[0074] It should be understood that the aforementioned processing device can be implemented by a processor reading instructions from memory and executing those instructions, or it can be implemented by a chip circuit.

[0075] Furthermore, the network architecture and business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0076] The technical solution of this application will be described in detail below with reference to specific embodiments:

[0077] Figure 2 A flowchart illustrating an artificial intelligence-based echo cancellation method provided in this application embodiment. Figure 1 ,like Figure 2 As shown, the method includes:

[0078] S201. Obtain the set of signals to be trained and the network parameters.

[0079] S202. Preprocess the training signal set to obtain preprocessed data.

[0080] S203. Based on the preprocessed data, the multi-loss function training strategy, the first loss function, and the second loss function, the network parameters are trained to obtain an optimized neural network model.

[0081] The first loss function is mainly used to suppress echoes, while the second loss function is used to preserve speech.

[0082] Optionally, the first loss function is dynamic compression loss and / or evaluation metric loss.

[0083] Among them, the evaluation index loss is used to evaluate speech quality.

[0084] Optionally, the second loss function is negative signal-to-noise ratio loss and / or mean square error loss.

[0085] Among them, the first loss function focuses on suppression, while the second loss function focuses on protection, and both can reflect the gap between the network output and the real label.

[0086] During the training phase, a loss function that emphasizes suppression and a loss function that emphasizes speech preservation are used for joint training. The shallow features of the network are shared, and deep features with different parameters are obtained through training. This ensures that the network speech enhancement style is consistent. Furthermore, since a large number of shallow features are shared, the strength of network suppression can be selected with only a small increase in the total number of network parameters.

[0087] S204. Obtain the set of signals to be tested, and use a dual-talk detection strategy to detect and process the set of signals to be tested in order to obtain the target network structure corresponding to the optimized neural network model.

[0088] The target network structure is an echo-suppressing structure and / or a speech-preserving structure.

[0089] In one possible implementation, the two-talk detection strategy combines traditional signal processing methods (such as energy-based two-talk detection) with deep learning methods (such as LSTM-based temporal classification models) to generate more reliable two-talk labels through a multimodal fusion strategy. For example, traditional methods provide preliminary two-talk probabilities, while deep learning models capture the dynamic changes in speech activity through temporal modeling, and finally generate a decision result through a weighted fusion strategy.

[0090] Multimodal fusion strategies can overcome the limitations of single detection methods. For example, in low signal-to-noise ratio scenarios, traditional methods may lead to misjudgments due to noise interference, while deep learning models can reduce the misjudgment rate by utilizing temporal information. Conversely, in scenarios with significant differences in the strength of the two-way signals, traditional methods can respond quickly, avoiding the latency of deep learning models. By combining the advantages of both, the accuracy and real-time performance of two-way detection are improved, thus providing a more reliable basis for dynamic parameter selection.

[0091] S205. Based on the optimized neural network model, target network structure, set of signals to be tested, and echo cancellation strategy, perform echo cancellation processing to obtain the complex spectrum of the de-echo signal.

[0092] Specifically, based on the target network structure, the complex spectrum of the signal set to be tested is subjected to echo cancellation processing, and the complex spectrum mask is multiplied with the complex spectrum to obtain the network-estimated complex spectrum of the de-echo signal.

[0093] S206. Using the overlap addition algorithm and the inverse Fourier transform algorithm, the complex spectrum of the echo-de-echo signal is synthesized into speech to obtain the echo-de-echo enhanced speech signal.

[0094] Specifically, the estimated time-domain echo-cancelled and enhanced speech signal is obtained by using a speech synthesis module that includes an Overlap-Add (OLA) algorithm and an inverse Fourier transform algorithm.

[0095] During the testing phase, in conjunction with dual-talk detection, if the segment is determined to be a pure echo segment, AI speech enhancement is performed using network parameters trained with a strong suppression loss function to suppress the echo; if the segment is determined to be a dual-talk segment or a pure near-end speech segment, AI speech enhancement is performed using network parameters trained with a loss function that emphasizes speech protection to protect the speech.

[0096] In one possible implementation, an online learning mechanism is introduced during the testing phase to dynamically adjust the weights of the loss function based on real-time feedback. For example, by monitoring the echo cancellation effect of the current frame (such as residual echo intensity and speech distortion level), the weight ratio of the suppression-guided and protection-guided loss functions is dynamically adjusted to adaptively optimize the network output.

[0097] The online learning mechanism allows the model to further adapt to scene changes during the testing phase, avoiding performance limitations caused by fixed loss weights during training. For example, in the early stages of a two-speech segment, if slight suppression of near-end speech is detected, the weight of the protection guidance loss is dynamically increased to enhance speech protection capabilities; while in the later stages of a pure echo segment, if residual echoes are not completely eliminated, the weight of the suppression guidance loss is increased to further optimize echo suppression. By dynamically adjusting the loss function, the model's robustness and adaptability in complex scenarios are significantly improved.

[0098] This embodiment provides an artificial intelligence echo cancellation method, comprising: acquiring a set of signals to be trained and network parameters; preprocessing the training signal set to obtain preprocessed data; training the network parameters according to the preprocessed data, a multi-loss function training strategy, a first loss function, and a second loss function to obtain an optimized neural network model; acquiring a set of signals to be tested and performing detection processing on the set of signals to be tested using a dual-talk detection strategy to obtain a target network structure corresponding to the optimized neural network model; performing echo cancellation processing according to the optimized neural network model, the target network structure, the set of signals to be tested, and the echo cancellation strategy to obtain a de-echo signal complex spectrum; and performing speech synthesis on the de-echo signal complex spectrum using an overlap-addition algorithm and an inverse Fourier transform algorithm to obtain a de-echo enhanced speech signal. Compared to existing technologies, which struggle to balance echo suppression and near-end speech protection capabilities, leading to speech distortion or echo residue, this method results in lower efficiency for artificial intelligence echo cancellation. This application adopts a loss function that focuses on speech preservation and a loss function that focuses on echo suppression, shares the shallow parameters of the AI ​​network, and combines a two-talk detection module to further suppress the echo of the far-end single talk in the pure echo segment, thereby improving the performance of echo cancellation. At the same time, it increases the speech protection capability in the two-talk segment, thereby improving the efficiency of AI echo cancellation.

[0099] Figure 3 A flowchart illustrating an artificial intelligence-based echo cancellation method provided in this application embodiment. Figure 2 ,like Figure 3 As shown, the training signal set includes a far-end reference signal, a main microphone signal, and a near-end speech signal. Step S202 specifically includes:

[0100] S301. Using an adaptive filtering algorithm, the far-end reference signal and the main microphone signal are processed to remove linear echo, so as to obtain a signal with removed linear echo.

[0101] In this embodiment, the far-end reference signal and the main microphone signal are passed through an adaptive filter (AF) to initially remove linear echoes from the signals, resulting in a linear echo-removed signal.

[0102] The AF can be updated using algorithms such as Normalized Least Mean Squares (NLMS) and Partitioned Block Frequency-Domain Adaptive Filter (PBFDAF).

[0103] S302. Perform signal processing on the far-end reference signal and the signal with the linear echo removed to obtain the spliced ​​complex spectrum.

[0104] The spliced ​​complex spectrum is obtained by splicing the complex spectra of the far-end reference signal and the signal after removing the linear echo, which are obtained through signal processing respectively.

[0105] In this embodiment, the far-end reference signal is processed to obtain a complex spectrum, and then the signal after removing the linear echo is processed to obtain a complex spectrum. The obtained complex spectra are then spliced ​​together to obtain a complex spectrum.

[0106] S303. Perform signal processing on the near-end speech signal to obtain the target complex spectrum.

[0107] In this embodiment, the complex spectrum of the clean near-end speech after signal processing is... As the target complex spectrum.

[0108] Where k is the frequency index and λ is the frame index.

[0109] S304. Perform feature dimensionality reduction on the spliced ​​complex spectrum to obtain compressed features.

[0110] In this embodiment, the spliced ​​complex spectrum is fed into the encoder in the network for feature dimensionality reduction processing to obtain the depth features of the signal, i.e., compressed features.

[0111] The encoder can employ networks such as Convolutional Neural Networks (CNN), Conformer, and Recurrent Neural Networks (RNN).

[0112] There are no specific requirements for the selected network type and construction order. The encoder is used to reduce the dimensionality of the complex spectrum features of the signal to a low dimension for feature construction.

[0113] In one possible implementation, a self-attention mechanism is introduced into the encoder module to dynamically adjust the weights of different frequency bands or time-domain segments by calculating the global dependencies between features. For example, in each layer of the encoder, attention weights are used to weight the complex spectral features of the input, enhancing features that are highly correlated with echoes and suppressing noise or non-critical information.

[0114] The self-attention mechanism enables the network to focus more on key features for echo cancellation (such as the spectral distribution of far-end signals) while weakening task-irrelevant noise or interference. For example, in two-speech segments, attention weights suppress echo components of far-end signals and enhance the spectral features of near-end speech; in pure echo segments, they prioritize enhancing the suppression capability of echo bands. By dynamically adjusting feature weights, the model's adaptability to complex scenarios is further improved, reducing the imbalance between suppression and protection capabilities caused by feature aliasing.

[0115] S305. Perform deep feature extraction processing on the compressed features to obtain deep features.

[0116] In this embodiment, the obtained compressed features are input into a feature extractor for deep feature extraction.

[0117] The feature extractor can be a gated recurrent unit (GRU), a temporal convolutional network (TCN), or a recurrent neural network (RNN), and there are no specific requirements for the type of network selected or the order in which they are built.

[0118] S306. The depth features are decoded using the first decoding structure and the second decoding structure to obtain a set of complex spectral masks.

[0119] In this embodiment, the depth features are recovered into complex spectral masks through the first decoding structure and the second decoding structure to obtain a set of complex spectral masks.

[0120] The complex spectrum mask set includes a first complex spectrum mask obtained through a first decoding structure and a second complex spectrum mask obtained through a second decoding structure.

[0121] Optionally, the first decoding structure includes first network parameters, and the second decoding structure includes second network parameters.

[0122] Optionally, the first network parameters are completely different from the second network parameters and / or only the last layer is different.

[0123] Specifically, both the first and second decoding structures are decoders. Similar to encoders, decoder 1 and decoder 2 can adopt network structures with the same modules but different parameters, or they can adopt network structures with the same modules but different parameters in the last layer.

[0124] In one possible implementation, a multi-head parallel decoding structure is introduced into the decoder module. Separate decoder branches are designed for suppression-guided and protection-guided tasks, respectively. Each branch shares shallow features from the encoder but is optimized using different network parameters (such as activation functions and inter-layer connections). For example, the suppression-guided decoder is trained using a dynamic compression loss function, while the protection-guided decoder is trained using a negative signal-to-noise ratio loss function. Finally, a gating network dynamically selects the output path.

[0125] The multi-head parallel architecture allows for more independent optimization objectives for different tasks, avoiding performance trade-offs between suppression and protection inherent in a single decoder. For example, in pure echo segments, the suppression-guided decoder can focus on eliminating residual echoes, while the output of the protection-guided decoder is suppressed; in dual-talk segments, a gating mechanism prioritizes the output of the protection-guided decoder. Furthermore, sharing shallow features ensures that the outputs of different decoders maintain consistency in speech enhancement style, avoiding speech incoherence issues caused by parameter differences.

[0126] In this embodiment, an adaptive filtering algorithm effectively removes the linear echo between the far-end reference signal and the main microphone signal. Combined with a concatenated complex spectrum constructed through signal processing, the dual-channel information is integrated. Efficient information compression and deep pattern mining are achieved through feature dimensionality reduction and deep feature extraction. Finally, a differentiated set of complex spectrum masks is generated through a dual-decoding structure, achieving high-precision separation and enhancement of the near-end speech signal. This improves speech clarity while reducing computational complexity. Furthermore, the differentiated design of the dual-decoding structure parameters enhances the robustness and diversity of feature extraction, effectively addressing the dual requirements of linear echo remnant suppression and target speech recovery. This improves the efficiency of AI-based echo cancellation.

[0127] Figure 4 A flowchart illustrating an artificial intelligence-based echo cancellation method provided in this application embodiment. Figure 3 ,like Figure 4 As shown, the specific implementation steps of S203 above include:

[0128] S401. Calculate the first loss value between the target complex spectrum and the first complex spectrum mask using the first loss function.

[0129] S402. Calculate the second loss value of the target complex spectrum and the second complex spectrum mask using the second loss function.

[0130] Specifically, the loss function is calculated for the target complex spectrum and the complex spectrum mask output by the decoder.

[0131] Specifically, a first loss function with strong suppression capability is used to calculate the first loss value of the target complex spectrum and the first complex spectrum mask output by decoder 1; a second loss function with strong speech protection capability is used to calculate the second loss value of the target complex spectrum and the second complex spectrum mask output by decoder 2.

[0132] S403. According to the multi-loss function training strategy, the first loss value and the second loss value are weighted and summed to obtain the final loss score.

[0133] S404. Based on the loss score and the loss multi-loss function training strategy, the network parameters are trained to obtain an optimized neural network model.

[0134] In this embodiment, the difference between the target complex spectrum and the double mask is quantified by a first loss function and a second loss function, respectively. A weighted summation of the loss values, combined with a multi-loss function training strategy, forms the final loss score. This achieves refined control and multi-dimensional optimization of the model training process, avoiding the limitations of a single loss function and balancing the importance of different optimization objectives through weight allocation. Finally, the network parameters are optimized through backpropagation, significantly improving the echo suppression accuracy, target speech recovery capability, and overall robustness of the neural network model in complex acoustic scenarios. This enhances the efficiency of AI-based echo cancellation.

[0135] Figure 5 A flowchart illustrating an artificial intelligence-based echo cancellation method provided in this application embodiment. Figure 4 ,like Figure 5 As shown, step S204 specifically includes:

[0136] S501. Using an adaptive filtering algorithm, the set of signals to be tested is processed to remove linear echoes in order to obtain the signal to be detected with the linear echo removed.

[0137] The set of signals to be tested may include a far-end reference signal, a main microphone signal, and a near-end speech signal. The composition of the signals in the set of signals to be tested is determined by a dual-talk detection strategy.

[0138] S502. Perform signal processing on the set of signals to be tested and the signal to be detected after removing linear echoes to obtain a set of complex spectra.

[0139] S503. Using a dual-talk detection strategy, the complex spectrum set is processed to obtain dual-talk detection tags.

[0140] Among them, the dual-talk detection tag is used to determine whether the set of signals to be tested is a pure echo segment.

[0141] Specifically, the complex spectrum set is fed into the Double-Talk Detect (DTD) decision module to output the predicted DTD label.

[0142] Specifically, the dual-lecture detection strategy can adopt traditional schemes based on correlation detection or energy / power detection, or deep learning methods, without any restrictions.

[0143] S504. If the set of signals to be tested is a pure echo segment, then the target network structure is an echo-suppressing structure.

[0144] The echo suppression structure completes the decoding process through the first decoding structure.

[0145] S505. If the set of signals to be tested is a dual-speech segment, then the target network structure is a protected voice structure.

[0146] The protection of the speech structure is achieved through the decoding process of the second decoding structure.

[0147] Specifically, the DTD tags are used to determine whether the set of signals to be tested is a pure echo segment.

[0148] Optionally, if the set of signals to be tested is predicted to be a pure echo segment, then decoder 1 with stronger suppression capability is selected in terms of network structure.

[0149] Optionally, if the set of signals to be tested is predicted to be a two-segment signal, then decoder 2 with stronger voice protection capability is selected in terms of network structure.

[0150] The dual-speech segment consists of signals that include both far-end echo and near-end speech.

[0151] In this embodiment, an adaptive filtering algorithm effectively removes linear echoes and generates a complex spectrum set. Combined with a dual-talk detection strategy, it accurately identifies pure echo segments and dual-talk segments, achieving intelligent adaptation to different acoustic scenarios. In pure echo segments, echo suppression capability is enhanced by suppressing the echo structure (first decoding structure), while in dual-talk segments, the integrity of the target speech is ensured by protecting the speech structure (second decoding structure). Ultimately, while improving echo cancellation accuracy, it effectively avoids speech damage in dual-talk scenarios, enhancing the system's adaptability and robustness in complex acoustic environments. This improves the efficiency of AI-based echo cancellation.

[0152] This application also provides a possible embodiment. Figure 6This is a schematic diagram of the operation process of the artificial intelligence echo cancellation system provided in the embodiments of this application, such as... Figure 6 As shown, during the training phase, the far-end reference signal and the main microphone signal are initially processed by an adaptive filter (using algorithms such as NLMS / PBFDAF) to remove linear echoes, and then the output complex spectrum signal is sent to the signal processing module. This signal, along with the near-end speech signal, is processed by the signal processing module and then input into the AI ​​echo cancellation network module. The decoder output and the target signal are input into loss function 1 and loss function 2, respectively, to calculate the error. After being weighted by the weighted processing module, the loss values ​​are fused according to preset weights and fed back to the AI ​​network. The network parameters are updated through optimization algorithms such as gradient descent, forming a closed-loop training until the loss converges. During the testing phase, the trained network is used directly: the input signal is processed by the AF module to remove linear echoes, and then passed to the AI ​​network by the signal processing module. The DTD decision result selects the corresponding decoder output complex spectrum mask value, which is combined with the far-end reference signal to generate an enhanced speech signal through speech synthesis and signal processing modules, and finally outputs the echo-cancelled speech.

[0153] The core of the system is a two-stage processing approach: linear echo suppression and nonlinear AI cancellation. The adaptive filter minimizes linear echo by dynamically adjusting the filter coefficients, and outputs a complex spectrum signal as the input basis for the AI ​​network. The AI ​​echo cancellation network compresses features through the encoder, mines deep speech characteristics through the feature extractor, and optimizes network parameters through loss function calculation and weighting module by fusing multiple loss functions (such as spectral loss and temporal loss) to ensure the generalization ability of different echo scenarios during the training phase. During the testing phase, the decoding path is dynamically switched through DTD decision to improve the robustness of real-time processing.

[0154] In this embodiment, the voice quality is significantly improved through a collaborative suppression mechanism of linear and nonlinear echoes: the adaptive filter AF effectively removes linear echoes, reducing the burden on subsequent AI processing; the AI ​​echo cancellation network, through a multi-decoder architecture and a DTD decision module, achieves adaptive processing of pure echo and non-pure echo scenarios, avoiding voice distortion in two-way speech; the multi-loss function weighted training mechanism enhances the network's generalization ability to complex echo environments, improving the accuracy of echo cancellation; the final output enhanced voice signal retains more original voice details in both the time and frequency domains, achieving a more natural voice interaction experience, suitable for high-quality voice communication needs in various scenarios such as conference systems and intelligent voice assistants.

[0155] Figure 7 This is a schematic diagram of the structure of the artificial intelligence echo cancellation device provided in the embodiments of this application, as shown below. Figure 7 As shown, the device includes: an acquisition module 71, a preprocessing module 72, a training module 73, a detection module 74, an echo cancellation module 75, and a speech synthesis module 76.

[0156] The acquisition module 71 is used to acquire the set of signals to be trained and the network parameters;

[0157] The preprocessing module 72 is used to preprocess the training signal set to obtain preprocessed data;

[0158] Training module 73 is used to train network parameters based on preprocessed data, a multi-loss function training strategy, a first loss function, and a second loss function to obtain an optimized neural network model; wherein, the first loss function is mainly used to suppress echo, and the second loss function is used to preserve speech;

[0159] The detection module 74 is used to acquire the set of signals to be tested and to perform detection processing on the set of signals to be tested using a dual-talk detection strategy to obtain the target network structure corresponding to the optimized neural network model; wherein, the target network structure is an echo suppression structure and / or a speech protection structure;

[0160] The echo cancellation module 75 is used to perform echo cancellation processing based on the optimized neural network model, the target network structure, the set of signals to be tested and the echo cancellation strategy, so as to obtain the complex spectrum of the de-echo signal;

[0161] The speech synthesis module 76 is used to synthesize speech from the complex spectrum of the de-echo signal using an overlap-addition algorithm and an inverse Fourier transform algorithm, so as to obtain a de-echo enhanced speech signal.

[0162] In one possible design, the first loss function is dynamic compression loss and / or evaluation metric loss; whereby the evaluation metric loss is used to evaluate speech quality.

[0163] The second loss function is the negative signal-to-noise ratio loss and / or mean square error loss.

[0164] In one possible design, the training signal set includes a far-end reference signal, a main microphone signal, and a near-end speech signal;

[0165] The training signal set is preprocessed to obtain preprocessed data, including:

[0166] The preprocessing module 72 is also used to perform linear echo removal processing on the far-end reference signal and the main microphone signal through an adaptive filtering algorithm to obtain a linear echo removed signal;

[0167] Signal processing is performed on the far-end reference signal and the linear echo removed signal to obtain a spliced ​​complex spectrum; wherein, the spliced ​​complex spectrum is obtained by splicing the complex spectra obtained from the signal processing of the far-end reference signal and the linear echo removed signal respectively;

[0168] Signal processing is performed on the near-end speech signal to obtain the target complex spectrum;

[0169] The spliced ​​complex spectrum is subjected to feature dimensionality reduction processing to obtain compressed features;

[0170] Deep feature extraction is performed on the compressed features to obtain deep features;

[0171] The depth features are decoded using the first decoding structure and the second decoding structure to obtain a set of complex spectral masks; wherein the set of complex spectral masks includes a first complex spectral mask obtained by the first decoding structure and a second complex spectral mask obtained by the second decoding structure.

[0172] In one possible design, the network parameters are trained based on preprocessed data, a multi-loss function training strategy, a first loss function, and a second loss function to obtain an optimized neural network model, including:

[0173] Training module 73 is also used to calculate the first loss value between the target complex spectrum and the first complex spectrum mask using the first loss function;

[0174] The second loss value of the target complex spectrum and the second complex spectrum mask is calculated using the second loss function;

[0175] According to the multi-loss function training strategy, the first loss value and the second loss value are weighted and summed to obtain the final loss score;

[0176] Based on the loss score and loss multi-loss function training strategy, the network parameters are trained to obtain an optimized neural network model.

[0177] In one possible design, a set of signals to be tested is obtained, and a dual-talk detection strategy is used to process the set of signals to be tested in order to obtain the target network structure corresponding to the optimized neural network model, including:

[0178] The detection module 74 is also used to perform linear echo removal processing on the set of signals to be tested using an adaptive filtering algorithm to obtain the linear echo removed signal to be detected.

[0179] Signal processing is performed on the set of signals to be tested and the signal to be detected after removing linear echoes to obtain a set of complex spectra;

[0180] A dual-talk detection strategy is used to detect and process the complex spectrum set to obtain dual-talk detection tags; the dual-talk detection tags are used to determine whether the set of signals to be tested is a pure echo segment.

[0181] If the set of signals to be tested is a pure echo segment, then the target network structure is an echo-suppressed structure; wherein, the echo-suppressed structure completes the decoding process through the first decoding structure;

[0182] If the set of signals to be tested is a dual-segment network, then the target network structure is a protected speech structure; wherein, the protected speech structure is decoded by a second decoding structure.

[0183] In one possible design, the first decoding structure includes first network parameters, and the second decoding structure includes second network parameters;

[0184] The first network parameters are completely different from the second network parameters and / or only the last layer is different.

[0185] This embodiment provides an artificial intelligence echo cancellation device that can execute an artificial intelligence echo cancellation method described in the above embodiment. Its implementation principle and technical effect are similar, and will not be repeated here.

[0186] In the aforementioned specific implementation of an artificial intelligence echo cancellation method, each module can be implemented as a processor. The processor can execute computer execution instructions stored in the memory, thereby enabling the processor to execute the aforementioned artificial intelligence echo cancellation method.

[0187] Figure 8 This is a schematic diagram of the structure of an artificial intelligence echo cancellation device provided in an embodiment of this application. Figure 8 As shown, the AI-powered echo cancellation device 80 includes at least one processor 81 and a memory 82. The AI-powered echo cancellation device 80 also includes a communication component 83. The processor 81, memory 82, and communication component 83 are connected via a bus 84.

[0188] In the specific implementation process, at least one processor 81 executes computer execution instructions stored in memory 82, causing at least one processor 81 to execute a method in the field of acoustic echo cancellation as executed by the artificial intelligence echo cancellation device as described above.

[0189] The specific implementation process of processor 81 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0190] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0191] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage.

[0192] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0193] The above description of the functions implemented by the AI-powered echo cancellation device and its main control device illustrates the solutions provided by the embodiments of the present invention. It is understood that, to achieve the aforementioned functions, the AI-powered echo cancellation device or main control device includes corresponding hardware structures and / or software modules for executing each function. By combining the units and algorithm steps of the various examples described in the embodiments of the present invention, the embodiments of the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the technical solutions of the embodiments of the present invention.

[0194] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method in the field of acoustic echo cancellation as described above.

[0195] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0196] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in an AI echo cancellation device or a master control device.

[0197] This application also provides a computer program product, comprising: a computer program stored in a readable storage medium, at least one processor of the artificial intelligence echo cancellation device being able to read the computer program from the readable storage medium, and at least one processor executing the computer program causing the artificial intelligence echo cancellation device to perform the scheme provided in any of the above embodiments.

[0198] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disk, or optical disk.

[0199] The technical solutions of this application have been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it is readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. The above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. An artificial intelligence-based echo cancellation method, characterized in that, include: Obtain the set of signals to be trained and the network parameters; The training signal set is preprocessed to obtain preprocessed data; Based on the preprocessed data, the multi-loss function training strategy, the first loss function, and the second loss function, the network parameters are trained to obtain an optimized neural network model; wherein, the first loss function is mainly used to suppress echo, and the second loss function is used to preserve speech; A set of signals to be tested is obtained, and the set of signals to be tested is processed by a dual-talk detection strategy to obtain the target network structure corresponding to the optimized neural network model; wherein, the target network structure is an echo suppression structure and / or a speech protection structure; Based on the optimized neural network model, the target network structure, the set of signals to be tested, and the echo cancellation strategy, echo cancellation processing is performed to obtain the complex spectrum of the de-echo signal; The complex spectrum of the de-echo signal is synthesized into a speech signal by using an overlap-addition algorithm and an inverse Fourier transform algorithm.

2. The method according to claim 1, characterized in that, The first loss function is dynamic compression loss and / or evaluation index loss; wherein the evaluation index loss is used to evaluate speech quality; The second loss function is the negative signal-to-noise ratio loss and / or mean square error loss.

3. The method according to claim 2, characterized in that, The training signal set includes a far-end reference signal, a main microphone signal, and a near-end speech signal; The preprocessing of the training signal set to obtain preprocessed data includes: An adaptive filtering algorithm is used to remove linear echo from the far-end reference signal and the main microphone signal to obtain a signal with removed linear echo. The far-end reference signal and the linearly removed echo signal are subjected to signal processing to obtain a spliced ​​complex spectrum; wherein, the spliced ​​complex spectrum is obtained by splicing the complex spectra obtained by the signal processing of the far-end reference signal and the linearly removed echo signal respectively; The near-end speech signal is processed to obtain the target complex spectrum; The spliced ​​complex spectrum is subjected to feature dimensionality reduction processing to obtain compressed features; The compressed features are subjected to depth feature extraction processing to obtain depth features; The depth features are decoded using a first decoding structure and a second decoding structure to obtain a set of complex spectral masks; wherein the set of complex spectral masks includes a first complex spectral mask obtained by the first decoding structure and a second complex spectral mask obtained by the second decoding structure.

4. The method according to claim 3, characterized in that, The step of training the network parameters based on the preprocessed data, the multi-loss function training strategy, the first loss function, and the second loss function to obtain an optimized neural network model includes: The first loss value between the target complex spectrum and the first complex spectrum mask is calculated using the first loss function. The second loss value of the target complex spectrum and the second complex spectrum mask is calculated using the second loss function. According to the multi-loss function training strategy, the first loss value and the second loss value are weighted and summed to obtain the final loss score. The network parameters are trained according to the loss score and the multiple loss function training strategy to obtain the optimized neural network model.

5. The method according to claim 4, characterized in that, The process of acquiring the set of signals to be tested and performing detection processing on the set of signals to be tested using a dual-talk detection strategy to obtain the target network structure corresponding to the optimized neural network model includes: The adaptive filtering algorithm is used to perform the linear echo removal process on the set of signals to be tested to obtain the linear echo removed signal to be detected. The signal processing is performed on the set of signals to be tested and the signal to be detected after removing linear echoes to obtain a set of complex spectra; The dual-talk detection strategy is used to perform the detection processing on the complex spectrum set to obtain dual-talk detection tags; wherein, the dual-talk detection tags are used to determine whether the set of signals to be tested is a pure echo segment; If the set of signals to be tested is the pure echo segment, then the target network structure is the echo-suppressed structure; wherein the echo-suppressed structure completes the decoding process through the first decoding structure; If the set of signals to be tested is a dual-segment, then the target network structure is the protected speech structure; wherein the protected speech structure completes the decoding process through the second decoding structure.

6. The method according to claim 5, characterized in that, The first decoding structure includes first network parameters, and the second decoding structure includes second network parameters; The first network parameters are completely different from the second network parameters and / or only the last layer is different.

7. An artificial intelligence-based echo cancellation device, characterized in that, include: The acquisition module is used to acquire the set of signals to be trained and the network parameters; A preprocessing module is used to preprocess the training signal set to obtain preprocessed data; The training module is used to train the network parameters based on the preprocessed data, the multi-loss function training strategy, the first loss function, and the second loss function to obtain an optimized neural network model; wherein, the first loss function is mainly used to suppress echo, and the second loss function is used to preserve speech. The detection module is used to acquire a set of signals to be tested, and to perform detection processing on the set of signals to be tested using a dual-talk detection strategy to obtain the target network structure corresponding to the optimized neural network model; wherein, the target network structure is an echo suppression structure and / or a speech protection structure; The echo cancellation module is used to perform echo cancellation processing based on the optimized neural network model, the target network structure, the set of signals to be tested, and the echo cancellation strategy, so as to obtain the complex spectrum of the de-echo signal; The speech synthesis module is used to synthesize speech from the complex spectrum of the de-echo signal using an overlap-addition algorithm and an inverse Fourier transform algorithm, so as to obtain a de-echo enhanced speech signal.

8. An artificial intelligence-based echo cancellation device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-6.

Images