Anti-interference voice wake-up method, system and terminal device

Abstract

The application belongs to the technical field of speech signal processing, and provides an anti-interference voice wake-up method, which comprises the following steps: obtaining a voice signal, performing noise suppression on the voice signal through a signal processing sub-network to output an enhanced feature, outputting a first discrimination result and a scene recognition result based on the enhanced feature through a wake-up acoustic model and a wake-up decoder respectively, outputting a second discrimination result through the wake-up acoustic model and an end-to-end wake-up discrimination sub-network, and finally adaptively fusing the first and second discrimination results according to the scene recognition result to output a final wake-up decision. The application improves the wake-up robustness in a high-noise scene through integrated modeling and scene adaptive decision-making.

Description

Technical Field

[0001] This application belongs to the field of speech signal processing technology, and in particular relates to an anti-interference voice wake-up method, system and terminal device. Background Technology

[0002] Voice wake-up technology, as an entry point for human-computer interaction, is widely used in scenarios such as in-vehicle systems, smart homes, and wearable devices. Existing voice wake-up systems typically employ a two-stage architecture: first, the noisy speech is enhanced through a signal processing module, such as spectral subtraction, Wiener filtering, or neural network-based speech denoising; then, the enhanced speech input wake-up module performs keyword recognition, and the wake-up module is often implemented based on a hidden Markov model or a deep neural network.

[0003] However, current methods suffer from problems such as the independent deployment of signal processing and voice wake-up modules and inconsistent training objectives. When the type or intensity of noise changes, the voice enhancement results are difficult to match with the characteristics of the wake-up model, leading to a decrease in wake-up rate or an increase in false wake-up rate. Summary of the Invention

[0004] This application provides an anti-interference voice wake-up method, system, and terminal device, which can solve the problems of independent deployment of the signal processing module and voice wake-up module and inconsistent training objectives in the current method. When the noise type or intensity changes, the voice enhancement result is difficult to match with the wake-up model characteristics, resulting in a decrease in wake-up rate or an increase in false wake-up rate.

[0005] In a first aspect, embodiments of this application provide an anti-interference voice wake-up method, comprising: extracting features from a voice signal in a target application scenario to obtain voice features; performing noise suppression and feature enhancement on the voice features through a signal processing sub-network in an integrated voice processing model to obtain enhanced voice features and a scene recognition result; based on the enhanced voice features and the voice features, outputting a first wake-up discrimination result through a wake-up acoustic model and a wake-up decoder in the integrated voice processing model; based on the enhanced voice features and the voice features, outputting a second wake-up discrimination result through a wake-up acoustic model and an end-to-end wake-up decision sub-network in the integrated voice processing model; and adaptively fusing the first wake-up discrimination result and the second wake-up discrimination result according to the scene recognition result to output a final wake-up decision result.

[0006] In one possible implementation of the first aspect, the above-mentioned speech features are enhanced by performing noise suppression and feature enhancement on the speech features through the signal processing sub-network in the integrated speech processing model, including:

[0007] The speech features are input into the signal processing subnetwork, and noise suppression is performed on the speech features based on the time-frequency domain features;

[0008] Output enhanced speech features after noise suppression.

[0009] Optionally, in another possible implementation of the first aspect, the above-mentioned wake-up discrimination result is output based on enhanced speech features and speech features, through the wake-up acoustic model and wake-up decoder in the integrated speech processing model, including:

[0010] The enhanced speech features and speech features are input into the wake-up acoustic model for temporal modeling to obtain syllable-level posterior probability representations.

[0011] Based on syllable-level posterior probability representation, the preset keywords are decoded by the wake-up decoder to obtain the first wake-up discrimination result.

[0012] Optionally, in another possible implementation of the first aspect, the above-mentioned second wake-up discrimination result is output based on enhanced speech features and speech features, through the wake-up acoustic model and end-to-end wake-up decision sub-network in the integrated speech processing model, including:

[0013] The speech features and enhanced speech features are used as input features and fed into the wake-up acoustic model for temporal modeling, and the output is audio embedding.

[0014] The audio embedding is input into the end-to-end wake-up discriminant subnetwork, which models the entire network without relying on explicit wake-up word alignment information, and outputs the second wake-up discriminant result.

[0015] Optionally, in another possible implementation of the first aspect, the first wake-up discrimination result and the second wake-up discrimination result are adaptively fused based on the scene recognition result to output the final wake-up decision result, including:

[0016] The fusion weights of the first wake-up discrimination result and the second wake-up discrimination result are determined based on the scene recognition results;

[0017] The first wake-up judgment result and the second wake-up judgment result are weighted and fused according to the fusion weight to obtain the comprehensive wake-up score;

[0018] The comprehensive wake-up score is compared with the wake-up threshold, and the final wake-up decision result is output based on the comparison result.

[0019] Optionally, in another possible implementation of the first aspect, the fusion weights for determining the first wake-up discrimination result and the second wake-up discrimination result based on the scene recognition result include:

[0020] When the scene recognition result indicates a high-noise scene, the fusion weight assigned to the second wake-up discrimination result is set to be higher than the fusion weight assigned to the first wake-up discrimination result;

[0021] When the scene recognition result indicates a low-noise scene, the fusion weight assigned to the first wake-up discrimination result is set to be higher than the fusion weight assigned to the second wake-up discrimination result.

[0022] Optionally, in another possible implementation of the first aspect, the aforementioned integrated speech processing model is obtained through end-to-end joint training, the training including:

[0023] Construct a training dataset, which includes noise-free speech samples, low-noise speech samples, and high-noise speech samples.

[0024] Based on the training dataset, a multi-loss function joint optimization strategy is used to train the integrated speech processing model. The multi-loss function includes at least signal processing loss, wake-up discrimination loss, and end-to-end wake-up discrimination loss.

[0025] Optionally, in another possible implementation of the first aspect, the above-mentioned training of the integrated speech processing model based on the training dataset and employing a joint optimization strategy using multiple loss functions includes:

[0026] Calculate the signal processing loss between the enhanced speech features output by the signal processing subnetwork and the corresponding noise-free speech;

[0027] Calculate the classification loss between the syllable confidence score output by the wake-up acoustic model and the true label;

[0028] Focal Loss is used to calculate the end-to-end wake-up discrimination loss between the second wake-up discrimination result output by the end-to-end wake-up discrimination subnetwork and the true label.

[0029] The total loss is obtained by weighted summation of the signal processing loss, classification loss, and end-to-end wake-up loss.

[0030] The parameters of the integrated speech processing model are updated based on the total loss;

[0031] The above-mentioned use of Focal Loss to calculate the end-to-end wake-up discrimination loss between the second wake-up discrimination result output by the end-to-end wake-up discrimination sub-network and the true label specifically includes:

[0032] For high-noise speech samples, the weighting coefficient of the focus loss function is increased, and the end-to-end wake-up discrimination loss between the second wake-up discrimination result and the real label is calculated based on the weighting coefficient using the focus loss function.

[0033] Secondly, embodiments of this application provide an anti-interference voice wake-up system, comprising: a feature extraction module for extracting features from a voice signal in a target application scenario to obtain voice features; a feature enhancement module for performing noise suppression and feature enhancement on the voice features through a signal processing sub-network in an integrated voice processing model to obtain enhanced voice features and scene recognition results; a first wake-up discrimination module for outputting a first wake-up discrimination result based on the enhanced voice features and the voice features, through a wake-up acoustic model and a wake-up decoder in the integrated voice processing model; a second wake-up discrimination module for outputting a second wake-up discrimination result based on the enhanced voice features and the voice features, through a wake-up acoustic model and an end-to-end wake-up decision sub-network in the integrated voice processing model; and an adaptive fusion decision module for adaptively fusing the first wake-up discrimination result and the second wake-up discrimination result according to the scene recognition result to output a final wake-up decision result.

[0034] Thirdly, embodiments of this application provide a terminal device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement an anti-interference voice wake-up method as described above.

[0035] Beneficial Effects: This application constructs an integrated speech processing model, jointly training the signal processing subnetwork, wake-up acoustic model, and end-to-end wake-up decision subnetwork end-to-end. This solves the adaptability problem caused by inconsistent training objectives in traditional independent modules, significantly improving the system robustness in complex noisy scenarios. Furthermore, the end-to-end wake-up decision subnetwork assists in judging wake-up results in noisy scenarios without relying on explicit wake-up word alignment information. Combined with weighted optimization using a focus loss function for high-noise samples during the training phase, the wake-up accuracy in noisy scenarios is further improved. Simultaneously, the scene recognition subnetwork perceives the current acoustic environment and adaptively weights and fuses the first and second wake-up judgment results, enabling wake-up decisions to dynamically adapt to different scenarios and improve the user wake-up experience. Moreover, compared to traditional solutions that deploy signal processing and wake-up independently, this application effectively reduces memory usage and computational overhead, lowering resource consumption and power consumption of edge devices while maintaining or even improving performance. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart illustrating an anti-interference voice wake-up method provided in an embodiment of this application;

[0038] Figure 2 This is a flowchart of the wake-up module processing according to an embodiment of this application;

[0039] Figure 3 This is a schematic diagram of a model training scheme provided in an embodiment of this application;

[0040] Figure 4 This is a schematic diagram of adaptive decision-making based on an acoustic scene provided in an embodiment of this application;

[0041] Figure 5 This is a schematic diagram of the structure of an anti-interference voice wake-up system provided in one embodiment of this application;

[0042] Figure 6 This is a schematic diagram of the structure of the terminal device provided in the embodiments of this application. Detailed Implementation

[0043] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0044] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0045] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0046] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0047] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0048] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0049] The following is a detailed description of an anti-interference voice wake-up method, system, terminal device, and storage medium provided in this application, with reference to the accompanying drawings.

[0050] Figure 1 A flowchart illustrating an anti-interference voice wake-up method provided in an embodiment of this application is shown.

[0051] like Figure 1 As shown, the anti-interference voice wake-up method includes the following steps:

[0052] S101. Extract features from the speech signal in the target application scenario to obtain speech features;

[0053] In this application embodiment, the target application scenarios include, but are not limited to, voice interaction scenarios with environmental noise, such as in-vehicle environments, home environments, and environments where wearable devices are located. The acquired voice signal includes both clean voice data from quiet scenarios and noisy voice data with environmental noise, background music, and interference from multiple speakers. It should be noted that this application is not limited to these; the method provided in this application can be applied to any scenario with voice wake-up requirements and accompanying background interference. For example, in an in-vehicle driving scenario, the acquired voice signal can be superimposed with various types of interference components such as vehicle tire noise, wind noise, background music from in-vehicle entertainment devices, and conversations among passengers.

[0054] S102. Noise suppression and feature enhancement are performed on speech features through the signal processing sub-network in the integrated speech processing model to obtain enhanced speech features and scene recognition results.

[0055] As one possible implementation, the above-mentioned speech features are FBANK features or MFCC features. This application is not limited to these, and other types of speech acoustic features can also be selected according to the computing power and accuracy requirements of the actual application.

[0056] Furthermore, in this embodiment of the application, step S102 specifically includes:

[0057] The speech features are input into the signal processing subnetwork, and noise suppression is performed on the speech features based on the time-frequency domain features;

[0058] Output enhanced speech features after noise suppression.

[0059] It should be noted that the aforementioned integrated speech processing model is an end-to-end model architecture that integrates signal processing and speech wake-up. It is not a simple concatenation of the signal processing model and the speech wake-up model, but rather a collaborative modeling of speech enhancement, wake-up recognition, and scene discrimination tasks within the same feature space through a shared feature encoding structure. For example, the signal processing sub-network is as follows: Figure 2 As shown, its input is the frequency domain features of noisy speech, and the output includes denoised FBANK features (i.e., enhanced speech features), while simultaneously outputting basic information related to acoustic scene classification. It is important to note that the optimization objective of this signal processing sub-network is not solely based on speech reconstruction quality, but also considers wake-up discrimination requirements during end-to-end training, thereby fundamentally avoiding the problem of "enhanced results being unfriendly to the wake-up task" in traditional cascaded schemes.

[0060] S103. Based on enhanced speech features and speech features, the first wake-up discrimination result is output through the wake-up acoustic model and wake-up decoder in the integrated speech processing model;

[0061] Furthermore, in this embodiment, step S103 specifically includes:

[0062] The enhanced speech features and speech features are input into the wake-up acoustic model for temporal modeling to obtain syllable-level posterior probability representations.

[0063] Based on syllable-level posterior probability representation, the preset keywords are decoded by the wake-up decoder to obtain the first wake-up discrimination result.

[0064] For example, preset keywords could be XX voice assistant, XXX classmate, etc., but this application does not limit them.

[0065] In one embodiment of this application, as Figure 2The wake-up acoustic model shown can be implemented using wku_encoder, with its core function being to model the temporal information of speech and output the posterior probability of the corresponding syllable. The wake-up decoder can be implemented using wku_decoder, and the decoding process can employ a greedy search or bundle search strategy, ultimately outputting the first wake-up score, score1, which is the first wake-up discrimination result. This application is not limited to this; the specific network structure of the wake-up encoder and decoder can be adaptively adjusted according to the hardware computing power, wake-up response speed, and recognition accuracy requirements of the actual scenario.

[0066] S104. Based on enhanced speech features and speech features, the second wake-up discrimination result is output through the wake-up acoustic model and end-to-end wake-up decision sub-network in the integrated speech processing model.

[0067] Furthermore, in this embodiment of the application, step S104 specifically includes:

[0068] The speech features and enhanced speech features are used as input features and fed into the wake-up acoustic model for temporal modeling, and the output is audio embedding.

[0069] The audio embedding is input into the end-to-end wake-up discriminant subnetwork, which models the entire network without relying on explicit wake-up word alignment information, and outputs the second wake-up discriminant result.

[0070] In one embodiment of this application, as Figure 2 The end-to-end wake-up discriminant subnetwork shown can be implemented using wku_e2e, which can directly perform global modeling of the entire input speech feature and output wake-up confidence score2, i.e., the second wake-up discrimination result. For example, in high-noise scenarios, the syllable alignment information relied upon by the speech wake-up subnetwork may be severely damaged by noise, leading to a significant decrease in discrimination accuracy. However, the end-to-end wake-up discriminant subnetwork does not rely on fine syllable alignment information and performs wake-up discrimination through the global features of the entire speech sentence. This can effectively avoid the errors caused by the failure of alignment information and provide a more robust auxiliary basis for wake-up decisions in complex scenarios.

[0071] S105. Based on the scene recognition results, adaptively fuse the first wake-up discrimination result and the second wake-up discrimination result, and output the final wake-up decision result.

[0072] Furthermore, in this embodiment of the application, step S105 specifically includes:

[0073] The fusion weights of the first wake-up discrimination result and the second wake-up discrimination result are determined based on the scene recognition results;

[0074] The first wake-up judgment result and the second wake-up judgment result are weighted and fused according to the fusion weight to obtain the comprehensive wake-up score;

[0075] The comprehensive wake-up score is compared with the wake-up threshold, and the final wake-up decision result is output based on the comparison result.

[0076] Furthermore, in this embodiment of the application, the above-mentioned determination of the fusion weight of the first wake-up discrimination result and the second wake-up discrimination result based on the scene recognition result includes:

[0077] When the scene recognition result indicates a high-noise scene, the fusion weight assigned to the second wake-up discrimination result is set to be higher than the fusion weight assigned to the first wake-up discrimination result;

[0078] When the scene recognition result indicates a low-noise scene, the fusion weight assigned to the first wake-up discrimination result is set to be higher than the fusion weight assigned to the second wake-up discrimination result.

[0079] As one possible implementation, high-noise scenarios can provide... Low-noise environments can provide .

[0080] Note: score is the wake-up confidence score, ranging from [0,1], wake-up threshold th ranges from (0,1), score1 is the first wake-up discrimination result, score2 is the second wake-up discrimination result, th_noise1 is the wake-up threshold of the wake-up decoder, and th_noise2 is the wake-up threshold of the end-to-end wake-up discrimination sub-network.

[0081] For example, a high-noise scene can be a scene with a signal-to-noise ratio of less than 5dB, while a low-noise scene can be a scene with a signal-to-noise ratio of greater than or equal to 5dB.

[0082] In this embodiment of the application, when the scene recognition result indicates a high-noise scene, the fusion weight assigned to the second wake-up discrimination result is set to be higher than the fusion weight assigned to the first wake-up discrimination result; when the scene recognition result indicates a low-noise or quiet scene, the fusion weight assigned to the first wake-up discrimination result is set to be higher than the fusion weight assigned to the second wake-up discrimination result. Figure 4 The process of scene-adaptive wake-up decision-making is shown. This mechanism can reduce the probability of false wake-ups while ensuring the wake-up rate.

[0083] In this embodiment of the application, the above-mentioned integrated speech processing model is obtained through end-to-end joint training, which includes:

[0084] Construct a training dataset, which includes noise-free speech samples, low-noise speech samples, and high-noise speech samples.

[0085] Based on the training dataset, a multi-loss function joint optimization strategy is used to train the integrated speech processing model. The multi-loss function includes at least signal processing loss, wake-up discrimination loss, and end-to-end wake-up discrimination loss.

[0086] It should be noted that noisy speech (high noise and low noise) is generated by probabilistically superimposing noise with a signal-to-noise ratio of -5dB to 20dB onto noisy speech, and the superposition probability follows a uniform distribution.

[0087] For example, low-noise speech samples and high-noise speech samples can be speech samples with a signal-to-noise ratio (SNR) in the range of 10dB to 20dB and speech samples with an SNR in the range of -5dB to 5dB, respectively.

[0088] Furthermore, in this embodiment, the above-mentioned training of the integrated speech processing model based on the training dataset using a multi-loss function joint optimization strategy includes:

[0089] Calculate the signal processing loss between the enhanced speech features output by the signal processing subnetwork and the corresponding noise-free speech;

[0090] Calculate the classification loss between the syllable confidence score output by the wake-up acoustic model and the true label;

[0091] Focal Loss is used to calculate the end-to-end wake-up discrimination loss between the second wake-up discrimination result output by the end-to-end wake-up discrimination subnetwork and the true label.

[0092] The total loss is obtained by weighted summation of the signal processing loss, classification loss, and end-to-end wake-up loss.

[0093] The parameters of the integrated speech processing model are updated based on the total loss.

[0094] Furthermore, in this embodiment of the application, the above-mentioned calculation of the end-to-end wake-up discrimination loss between the second wake-up discrimination result output by the end-to-end wake-up discrimination sub-network and the real label using Focal Loss specifically includes:

[0095] For high-noise speech samples, the weighting coefficient of the focus loss function is increased, and the end-to-end wake-up discrimination loss between the second wake-up discrimination result and the real label is calculated based on the weighting coefficient using the focus loss function.

[0096] In one embodiment of this application, the signal processing loss can be the minimum mean square error loss (MMSE) or SI-SNR loss, primarily used to constrain the difference between the enhanced speech output by the signal processing branch and the corresponding clean speech; the classification and discrimination loss is the cross-entropy loss (CE) or a sequence discrimination loss based on temporal alignment (such as CTC loss), primarily used to constrain the accuracy of the speech wake-up branch in recognizing the wake-up word; the end-to-end wake-up discrimination loss can be the cross-entropy loss, primarily used to constrain the wake-up discrimination capability of the end-to-end wake-up auxiliary branch for the entire speech segment. For example, refer to... Figure 3 The end-to-end joint training logic shown can be used to calculate the total loss L using a weighted formula, where the weight coefficients of each loss term can be preset and adjusted according to the training objectives and scenario requirements. It should be noted that by introducing multiple loss functions simultaneously during the same training process, the model can optimize speech enhancement quality while explicitly constraining the effectiveness of the enhancement results for the wake-up discrimination task. Compared to traditional methods that only use speech reconstruction error as the signal processing objective, this joint optimization strategy guides the model to learn speech feature representations that are more conducive to wake-up discrimination, thereby achieving more stable wake-up performance in high-noise scenarios.

[0097] The following will combine Figure 2 This paper describes an implementation example of the overall architecture of an integrated speech processing model.

[0098] First, the input audio signal is acquired and its features are extracted to obtain the speech features fbank. The speech features are then input into a signal processing neural network to generate enhanced speech features fbank-sgp and output an acoustic scene classification scene to represent the acoustic scene in which the current speech occurs. The enhanced speech features are then input into a wake-up acoustic model to model the speech temporal information to obtain a syllable-level posterior probability representation post-psyllable. Based on the syllable-level posterior probability representation, a wake-up decoder decodes preset keywords to obtain a first wake-up score score1. The enhanced speech features are then input into an end-to-end wake-up decision subnetwork to obtain a second wake-up score score2. A joint decision is made based on the first wake-up score score1, the second wake-up score score2, and the acoustic scene classification scene scene to generate a comprehensive wake-up score score. The comprehensive wake-up score is compared with a wake-up threshold th. If the comprehensive wake-up score score is greater than the wake-up threshold th, speech wake-up is determined to have occurred; otherwise, it is determined to be a rejection.

[0099] It should be noted that the integrated model provided in this application has three core innovative advantages: First, the signal processing module and the wake-up module are jointly trained end-to-end, which makes them more adaptable and significantly improves noise robustness; Second, it can adaptively adopt different decision weights and schemes according to the acoustic scene classification results to achieve refined wake-up control of scene perception; Third, the introduction of an end-to-end wake-up branch for auxiliary discrimination can effectively improve the wake-up success rate in complex scenes such as high noise.

[0100] This application provides an anti-interference voice wake-up method. First, it extracts features from the speech signal in the target application scenario to obtain speech features. Then, it uses a signal processing subnetwork in an integrated speech processing model to perform noise suppression and feature enhancement on the speech features, obtaining enhanced speech features and scene recognition results. Based on the enhanced speech features and the speech features, it outputs a first wake-up discrimination result through a wake-up acoustic model and a wake-up decoder in the integrated speech processing model. Next, based on the enhanced speech features and the speech features, it outputs a second wake-up discrimination result through a wake-up acoustic model and an end-to-end wake-up decision subnetwork in the integrated speech processing model. Finally, based on the scene recognition result, it adaptively fuses the first and second wake-up discrimination results to output the final wake-up decision result. This application can significantly improve the wake-up success rate in complex noisy environments and effectively reduce the false wake-up rate, enhancing the robustness and adaptability of the voice wake-up system in practical applications.

[0101] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0102] Corresponding to the anti-interference voice wake-up method in the above embodiment, Figure 5 The diagram shows a structural block diagram of an anti-interference voice wake-up system provided in an embodiment of this application. For ease of explanation, only the parts related to the embodiment of this application are shown.

[0103] Reference Figure 5 The system 500 includes:

[0104] The feature extraction module 501 is used to extract features from the speech signal in the target application scenario to obtain speech features;

[0105] The feature enhancement module 502 is used to perform noise suppression and feature enhancement on speech features through the signal processing sub-network in the integrated speech processing model to obtain enhanced speech features and scene recognition results.

[0106] The first wake-up discrimination module 503 is used to output the first wake-up discrimination result based on the enhanced speech features and speech features, through the wake-up acoustic model and wake-up decoder in the integrated speech processing model;

[0107] The second wake-up discrimination module 504 is used to output the second wake-up discrimination result based on enhanced speech features and speech features, through the wake-up acoustic model and end-to-end wake-up decision sub-network in the integrated speech processing model;

[0108] The adaptive fusion decision module 505 is used to adaptively fuse the first wake-up discrimination result and the second wake-up discrimination result based on the scene recognition result, and output the final wake-up decision result.

[0109] In practical use, the anti-interference voice wake-up system provided in this application embodiment can be configured in any terminal device to execute the aforementioned anti-interference voice wake-up method.

[0110] This application provides an anti-interference voice wake-up system. First, it extracts features from the voice signal in the target application scenario to obtain voice features. Then, it uses a signal processing subnetwork in an integrated voice processing model to perform noise suppression and feature enhancement on the voice features, resulting in enhanced voice features and scene recognition results. Based on the enhanced voice features and voice features, it outputs a first wake-up discrimination result through a wake-up acoustic model and wake-up decoder in the integrated voice processing model. Next, based on the enhanced voice features and voice features, it outputs a second wake-up discrimination result through a wake-up acoustic model and an end-to-end wake-up decision subnetwork in the integrated voice processing model. Finally, based on the scene recognition result, it adaptively fuses the first and second wake-up discrimination results to output the final wake-up decision result. This application can significantly improve the wake-up success rate in complex noisy environments and effectively reduce the false wake-up rate, enhancing the robustness and adaptability of the voice wake-up system in practical applications.

[0111] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0113] To implement the above embodiments, this application also proposes a terminal device.

[0114] Figure 6 This is a schematic diagram of the structure of a terminal device according to an embodiment of this application.

[0115] like Figure 6 As shown, the terminal device 200 includes:

[0116] The system includes a memory 210 and at least one processor 220, and a bus 230 connecting different components (including the memory 210 and the processor 220). The memory 210 stores a computer program, and when the processor 220 executes the program, it implements an anti-interference voice wake-up method as described in the embodiments of this application.

[0117] Bus 230 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0118] Terminal device 200 typically includes various electronically readable media. These media can be any available media that can be accessed by terminal device 200, including volatile and non-volatile media, removable and non-removable media.

[0119] Memory 210 may also include computer system readable media in the form of volatile memory, such as RAM 240 and / or cache 250. Terminal device 200 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 260 may be used to read and write non-removable, non-volatile magnetic media (… Figure 6 Not shown; usually referred to as a "hard drive"). Although Figure 6 As not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 230 via one or more data media interfaces. Memory 210 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of this application.

[0120] A program / utility 280 having a set (at least one) of program modules 270 may be stored in, for example, memory 210. Such program modules 270 include—but are not limited to—an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 270 typically perform the functions and / or methods described in the embodiments of this application.

[0121] Terminal device 200 can also communicate with one or more external devices 290 (e.g., keyboard, pointing device, display 291, etc.), and with one or more devices that enable a user to interact with terminal device 200, and / or with any device that enables terminal device 200 to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed via input / output (I / O) interface 292. Furthermore, terminal device 200 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 293. As shown, network adapter 293 communicates with other modules of terminal device 200 via bus 230. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with terminal device 200, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0122] The processor 220 performs various functional applications and data processing by running programs stored in the memory 210.

[0123] It should be noted that the implementation process and technical principles of the terminal device in this embodiment are explained in the foregoing description of an anti-interference voice wake-up method according to an embodiment of this application, and will not be repeated here.

[0124] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0125] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0126] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some regions, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0127] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0128] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software 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 this application.

[0129] In the embodiments provided in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0130] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0131] The above-described 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An anti-interference voice wake-up method, characterized in that, include: Speech features are obtained by extracting features from speech signals in the target application scenario; The speech features are subjected to noise suppression and feature enhancement by the signal processing sub-network in the integrated speech processing model, resulting in enhanced speech features and scene recognition results. Based on the enhanced speech features and the speech features, the first wake-up discrimination result is output through the wake-up acoustic model and wake-up decoder in the integrated speech processing model; Based on the enhanced speech features and the speech features, a second wake-up discrimination result is output through the wake-up acoustic model and the end-to-end wake-up decision sub-network in the integrated speech processing model; Based on the scene recognition results, the first wake-up discrimination result and the second wake-up discrimination result are adaptively fused to output the final wake-up decision result; The fusion weights of the first wake-up discrimination result and the second wake-up discrimination result are determined based on the scene recognition result. The first wake-up judgment result and the second wake-up judgment result are weighted and fused according to the fusion weight to obtain a comprehensive wake-up score; The comprehensive wake-up score is compared with the wake-up threshold, and the final wake-up decision result is output based on the comparison result. When the scene recognition result indicates a high-noise scene, the fusion weight assigned to the second wake-up discrimination result is set to be higher than the fusion weight assigned to the first wake-up discrimination result; when the scene recognition result indicates a low-noise scene, the fusion weight assigned to the first wake-up discrimination result is set to be higher than the fusion weight assigned to the second wake-up discrimination result. Construct a training dataset, which includes noise-free speech samples, low-noise speech samples, and high-noise speech samples; Based on the training dataset, the integrated speech processing model is trained using a multi-loss function joint optimization strategy. The multi-loss function includes at least signal processing loss, wake-up discrimination loss, and end-to-end wake-up discrimination loss. Calculate the signal processing loss between the enhanced speech features output by the signal processing sub-network and the corresponding noise-free speech; Calculate the classification loss between the syllable confidence score output by the wake-up acoustic model and the true label; The Focal Loss method is used to calculate the end-to-end wake-up discrimination loss between the second wake-up discrimination result output by the end-to-end wake-up discrimination subnetwork and the real label. The total loss is obtained by weighted summing of the signal processing loss, the classification loss, and the end-to-end wake-up loss. The parameters of the integrated speech processing model are updated based on the total loss; The calculation of the end-to-end wake-up discrimination loss between the second wake-up discrimination result output by the end-to-end wake-up discrimination sub-network and the real label using Focal Loss specifically includes: For high-noise speech samples, the weighting coefficient of the focus loss function is increased, and the end-to-end wake-up discrimination loss between the second wake-up discrimination result and the real label is calculated based on the weighting coefficient using the focus loss function.

2. The method according to claim 1, characterized in that, The enhanced speech features are obtained by performing noise suppression and feature enhancement on the speech features through the signal processing sub-network in the integrated speech processing model, including: The speech features are input into the signal processing subnetwork, and noise suppression processing is performed on the speech features based on the time-frequency domain features; Output enhanced speech features after noise suppression.

3. The method according to claim 2, characterized in that, Based on the enhanced speech features and the speech features, the first wake-up discrimination result is output through the wake-up acoustic model and wake-up decoder in the integrated speech processing model, including: The enhanced speech features and the speech features are input into the wake-up acoustic model for temporal modeling to obtain a syllable-level posterior probability representation; Based on the syllable-level posterior probability representation, the preset keywords are decoded by the wake-up decoder to obtain the first wake-up discrimination result.

4. The method according to claim 3, characterized in that, Based on the enhanced speech features and the speech features, the second wake-up discrimination result is output through the wake-up acoustic model and end-to-end wake-up decision sub-network in the integrated speech processing model, including: The speech features and the enhanced speech features are used as input features and fed into the wake-up acoustic model for temporal modeling, and the audio embedding is output. The audio embedding is input into the end-to-end wake-up discriminant subnetwork, and the network is modeled as a whole without relying on explicit wake-up word alignment information, and the second wake-up discriminant result is output.

5. An anti-interference voice wake-up system, characterized in that, include: The feature extraction module is used to extract features from the speech signal in the target application scenario to obtain speech features; The feature enhancement module is used to perform noise suppression and feature enhancement on the speech features through the signal processing sub-network in the integrated speech processing model, so as to obtain enhanced speech features and scene recognition results. The first wake-up discrimination module is used to output a first wake-up discrimination result based on the enhanced speech features and the speech features, through the wake-up acoustic model and wake-up decoder in the integrated speech processing model; The second wake-up discrimination module is used to output a second wake-up discrimination result based on the enhanced speech features and the speech features, through the wake-up acoustic model and the end-to-end wake-up decision sub-network in the integrated speech processing model; An adaptive fusion decision module is used to adaptively fuse the first wake-up discrimination result and the second wake-up discrimination result based on the scene recognition result, and output the final wake-up decision result; The fusion weights of the first wake-up discrimination result and the second wake-up discrimination result are determined based on the scene recognition result. The first wake-up judgment result and the second wake-up judgment result are weighted and fused according to the fusion weight to obtain a comprehensive wake-up score; The comprehensive wake-up score is compared with the wake-up threshold, and the final wake-up decision result is output based on the comparison result. When the scene recognition result indicates a high-noise scene, the fusion weight assigned to the second wake-up discrimination result is set to be higher than the fusion weight assigned to the first wake-up discrimination result; when the scene recognition result indicates a low-noise scene, the fusion weight assigned to the first wake-up discrimination result is set to be higher than the fusion weight assigned to the second wake-up discrimination result. Construct a training dataset, which includes noise-free speech samples, low-noise speech samples, and high-noise speech samples; Based on the training dataset, the integrated speech processing model is trained using a multi-loss function joint optimization strategy. The multi-loss function includes at least signal processing loss, wake-up discrimination loss, and end-to-end wake-up discrimination loss. Calculate the signal processing loss between the enhanced speech features output by the signal processing sub-network and the corresponding noise-free speech; Calculate the classification loss between the syllable confidence score output by the wake-up acoustic model and the true label; The Focal Loss method is used to calculate the end-to-end wake-up discrimination loss between the second wake-up discrimination result output by the end-to-end wake-up discrimination subnetwork and the real label. The total loss is obtained by weighted summing of the signal processing loss, the classification loss, and the end-to-end wake-up loss. The parameters of the integrated speech processing model are updated based on the total loss; The calculation of the end-to-end wake-up discrimination loss between the second wake-up discrimination result output by the end-to-end wake-up discrimination sub-network and the real label using Focal Loss specifically includes: For high-noise speech samples, the weighting coefficient of the focus loss function is increased, and the end-to-end wake-up discrimination loss between the second wake-up discrimination result and the real label is calculated based on the weighting coefficient using the focus loss function.

6. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 4.

Images