Speech detection method, apparatus, device, and storage medium

Abstract

Embodiments of the present application provide a speech detection method, device and equipment and a storage medium, relating to the technical field of artificial intelligence. The speech detection method comprises: acquiring speech data; inputting the speech data into a trained synthesized speech detection model, performing feature extraction to obtain a speech feature map, then performing feature separation on the speech feature map to obtain a separated feature map, and then performing attention weighting on the separated feature map to obtain an output feature map, thereby obtaining a speech detection result. In the embodiment, the speech detection method uses an end-to-end synthesized speech detection model to perform synthesized speech detection, reduces the complexity of the detection model, introduces an attention mechanism, and through adaptive weighting combination of different size receptive fields, can more effectively capture distinguishing features for distinguishing authenticity from feature maps of different sizes, thereby improving the detection performance and generalization ability of the synthesized speech detection network, improving the accuracy of the speech detection result, and realizing authenticity discrimination of the speech data.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to speech detection methods, apparatus, devices, and storage media. Background Technology

[0002] In recent years, speech synthesis technology has been gradually applied to speech signal processing systems such as voice interaction, voice broadcasting, and personalized voice production. Effective use of speech synthesis technology can improve the user experience of voice interaction and has broad potential applications. With the development and maturation of speech synthesis and speech conversion technologies, synthesized speech is becoming increasingly natural, possessing acoustic characteristics similar to real speech. This makes the synthesized and converted speech increasingly realistic, to the point that it is difficult for the human ear to distinguish the authenticity of the speech.

[0003] In related technologies, synthetic speech detection systems generally consist of a front-end feature extractor and a back-end binary classifier. The features captured by the front-end are mostly spectral features and phase features. Although designing different network structures, loss functions, and training methods can improve the performance of synthetic speech detection models, the detection potential of the model essentially depends on the information captured in the initial features. However, spectral features and phase features lose some information, which largely contributes to the poor performance of the back-end network in detecting the authenticity of speech. Summary of the Invention

[0004] The main objective of this invention is to propose a speech detection method, apparatus, device, and storage medium, which utilizes an end-to-end synthetic speech detection approach to improve the accuracy of speech authenticity detection results.

[0005] To achieve the above objectives, a first aspect of the present invention provides a speech detection method, comprising:

[0006] Acquire speech data, which includes synthesized speech data or real speech data;

[0007] The speech data is input into a trained synthetic speech detection model. The synthetic speech detection model includes a feature extraction module, an attention module, and an output module. The feature extraction module is connected to the attention module, and the attention module is connected to the output module. The attention module includes an attention convolution module and a first pooling layer. The attention convolution module includes one or more input convolution layers and attention output layers.

[0008] The speech data is processed by the feature extraction module to obtain a speech feature map.

[0009] The attention convolution module uses one or more input convolutional layers to perform feature separation on the speech feature map, resulting in one or more separated feature maps.

[0010] The attention output layer performs attention weighting on one or more of the separated feature maps to obtain the output feature map;

[0011] The output feature map is pooled through the first pooling layer to obtain a pooled feature map.

[0012] The output module performs speech detection on the pooled feature map to obtain the speech detection result.

[0013] In some embodiments, the input convolutional layer includes a first convolutional layer and a dilated convolutional layer, and the separation feature map includes a first separation feature map and a dilated separation feature map;

[0014] The attention convolution module uses one or more input convolutional layers to perform feature separation on the speech feature map, resulting in one or more separated feature maps, including:

[0015] The attention convolution module uses at least one first convolutional layer to perform feature separation on the speech feature map, thereby obtaining at least one first separated feature map.

[0016] The attention convolution module uses at least one dilated convolutional layer to perform feature separation on the speech feature map, resulting in at least one dilated feature map.

[0017] In some embodiments, the attention output layer obtains an output feature map by performing attention weighting on one or more of the separated feature maps, including:

[0018] A fused feature map is obtained based on at least one of the first separation feature maps and at least one of the cavity separation feature maps;

[0019] A compact feature vector is obtained based on the fused feature map;

[0020] The output feature map is obtained by applying attention weights to the separated feature map using the compact feature vector.

[0021] In some embodiments, obtaining a compact feature vector based on the fused feature map includes:

[0022] The fused feature map is pooled to obtain a fused pooled feature map;

[0023] The fused pooled feature map is fully connected to obtain the compact feature vector.

[0024] In some embodiments, the step of using the compact feature vector to perform attention weighting on the separated feature map to obtain the output feature map includes:

[0025] Obtain a soft attention matrix for one or more of the separated feature maps;

[0026] The soft attention value is calculated using the compact feature vector and the soft attention matrix;

[0027] The output feature map is obtained by weighting the separated feature map using the soft attention value.

[0028] In some embodiments, the output module performs speech detection based on the pooled feature map to obtain a speech detection result, including:

[0029] The pooled feature map is fully connected at least once to obtain a fully connected output map;

[0030] The speech detection result is obtained by performing a softmax activation operation on the fully connected output graph.

[0031] In some embodiments, the synthetic speech detection model includes at least one attention module connected in sequence;

[0032] The first attention module is connected to the feature extraction module, and the last attention module is connected to the output module. The first attention module receives the speech feature map, and the other attention modules besides the first attention module receive the pooled feature map output by the previous attention module.

[0033] To achieve the above objectives, a second aspect of the present invention provides a speech detection device, comprising:

[0034] A synthesized speech acquisition unit is used to acquire speech data;

[0035] An input unit is used to input the speech data into a trained synthetic speech detection model to obtain a speech detection result. The synthetic speech detection model includes a feature extraction module, an attention module, and an output module. The feature extraction module is connected to the attention module, and the attention module is connected to the output module. The attention module includes an attention convolution module and a first pooling layer. The attention convolution module includes one or more input convolution layers and attention output layers.

[0036] A feature extraction unit is used to extract features from the speech data through the feature extraction module to obtain a speech feature map.

[0037] The feature separation unit is used to perform feature separation on the speech feature map through the attention convolution module using one or more input convolutional layers to obtain one or more separated feature maps.

[0038] An attention weighting unit is used to perform attention weighting on one or more of the separated feature maps through the attention output layer to obtain an output feature map;

[0039] A pooling unit is used to pool the output feature map through the first pooling layer to obtain a pooled feature map.

[0040] The detection unit is used to perform speech detection on the pooled feature map through the output module to obtain the speech detection result.

[0041] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0042] To achieve the above objectives, a fourth aspect of this application provides a storage medium, which is a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0043] The speech detection method, apparatus, device, and storage medium proposed in this invention involve acquiring speech data; inputting the speech data into a trained synthetic speech detection model; the synthetic speech detection model extracting features from the speech data to obtain speech feature maps; then separating the speech feature maps to obtain one or more separated feature maps; further, attention weighting is applied to these separated feature maps to obtain an output feature map; the output feature map is then pooled to obtain a pooled feature map; finally, the speech detection result is obtained based on the pooled feature map. In this embodiment, the speech detection method utilizes an end-to-end synthetic speech detection model for synthetic speech detection, reducing the complexity of the detection model while introducing an attention mechanism. Through adaptive weighted combination of receptive fields of different sizes, the size of the receptive field of neurons is adaptively adjusted, enabling more effective capture of discriminative features for distinguishing between true and false speech from feature maps of different sizes. This improves the detection performance and generalization ability of the synthetic speech detection network, increases the accuracy of speech detection results, and achieves the distinction between true and false speech data. Attached Figure Description

[0044] Figure 1 This is a flowchart of the speech detection method provided in an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the structure of the synthesized speech detection model provided in an embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the structure of a synthetic speech detection model provided in another embodiment of the present invention.

[0047] Figure 4 This is a schematic diagram of the attention module structure of a synthetic speech detection model provided in another embodiment of the present invention.

[0048] Figure 5 This is a schematic diagram of the feature extraction module structure of a synthetic speech detection model provided in another embodiment of the present invention.

[0049] Figure 6 yes Figure 1 The flowchart for step S140.

[0050] Figure 7 yes Figure 1 The flowchart for step S150.

[0051] Figure 8 yes Figure 7 The flowchart for step S152.

[0052] Figure 9 yes Figure 7 The flowchart for step S153.

[0053] Figure 10 This is a schematic diagram of the structure of a synthetic speech detection model provided in another embodiment of the present invention.

[0054] Figure 11 yes Figure 1 The flowchart for step S170.

[0055] Figure 12 This is a structural block diagram of the speech detection device provided in an embodiment of the present invention.

[0056] Figure 13 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0058] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart.

[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0060] First, let's clarify some of the terms used in this invention:

[0061] Synthetic speech, also known as Text-to-Speech (TTS) technology, can convert any text information into standard, fluent speech in real time. It involves multiple disciplines such as acoustics, linguistics, digital signal processing, and computer science, and is a cutting-edge technology in the field of Chinese information processing. Its main problem is how to convert text information into audible sound information, that is, to make machines speak like humans. Traditional sound playback devices (systems), such as tape recorders, achieve this by pre-recording sound and then playing it back. This method has significant limitations in terms of content, storage, transmission, convenience, and timeliness. Computer speech synthesis, on the other hand, can convert any text into highly natural-sounding speech at any time, truly enabling machines to "speak like humans."

[0062] Deep learning involves learning the inherent patterns and hierarchical representations of sample data. The information gained during this learning process greatly aids in interpreting data such as text, images, and sound. Its ultimate goal is to enable machines to possess analytical and learning capabilities similar to humans, allowing them to recognize data such as text, images, and sound. Deep learning is a complex machine learning algorithm that has achieved results in speech and image recognition far exceeding previous related technologies. Deep learning has also yielded significant advancements in search technology, data mining, machine learning, machine translation, natural language processing, multimedia learning, speech recognition, recommendation and personalization technologies, and other related fields. Deep learning enables machines to mimic human activities such as sight, hearing, and thought, solving many complex pattern recognition problems and significantly advancing artificial intelligence-related technologies.

[0063] Attention Mechanism: Research on the attention mechanism first appeared in psychology. When a person sees a scene, although they are faced with the entire scene, their attention is focused on only a certain point. In 2014, Bahdanau et al. first introduced attention units into machine translation, combining the attention mechanism with neural networks to improve the accuracy of machine translation. Subsequently, many studies have attempted to combine attention units with neural networks for different tasks, and all have achieved significant advantages over traditional methods to varying degrees.

[0064] Pooling is an important concept in convolutional neural networks (CNNs), essentially a form of downsampling. There are several common pooling functions, the most common being max pooling, which divides the input image into several rectangular regions and outputs the maximum value for each sub-region. Pooling layers typically act on each input feature individually, continuously reducing the spatial size of the data, thus decreasing the number of parameters and computational cost. This also helps control overfitting to some extent. Generally, pooling layers are periodically inserted between convolutional layers in CNNs.

[0065] Skip-layer connections: In neural networks, this refers to connections that combine the predictions from the last layer with the feature predictions from shallower layers. If only deconvolution is used to upsample the feature map of the last layer to the size of the original image, a lot of detail will be lost because the feature map of the last layer is too small. Therefore, the prediction information of the last layer, which contains global information, is combined with the prediction information of shallower layers, which contain more local details, through direct connections.

[0066] Receptive field: refers to the input area that neurons in a neural network "see". In a convolutional neural network, the calculation of a certain element in the feature operation is affected by a certain area on the input image. This area is the receptive field of that element.

[0067] In recent years, speech synthesis technology has been gradually applied to speech signal processing systems such as voice interaction, voice broadcasting, and personalized voice production, including voice assistants and online banking. Effective use of speech synthesis technology can improve the user experience of voice interaction and has broad potential applications. With the development and maturation of speech synthesis and conversion technologies, synthesized speech is becoming increasingly natural, possessing acoustic characteristics similar to real speech. This makes the synthesized and converted speech increasingly realistic, even to the point that the human ear can hardly distinguish it from real speech. However, this also brings serious security risks. If exploited by criminals, it will pose a security threat, posing a challenge to the detection of synthesized speech.

[0068] In related technologies, synthetic speech detection systems generally consist of a front-end feature extractor and a back-end binary classifier. The features captured by the front-end are typically handcrafted, such as spectral features and phase features. Although designing different network structures, loss functions, and training methods can improve the performance of synthetic speech detection models, the model's detection potential essentially depends on the information captured in the initial features. However, the creation of handcrafted features results in the loss of some information, which largely contributes to the poor performance of the back-end network in synthetic speech detection. Therefore, many teams have proposed end-to-end synthetic speech detection. Although synthetic speech detection networks Res-TSSDNet and Inc-TSSDNet have shown good detection capabilities, there is still room for improvement in detection performance, generalization, stability, and model complexity.

[0069] Based on this, embodiments of the present invention provide a speech detection method, apparatus, device, and storage medium. The speech detection method utilizes an end-to-end synthetic speech detection model to perform synthetic speech detection, reducing the complexity of the detection model while introducing an attention mechanism. Through adaptive weighted combination of receptive fields of different sizes, the size of the receptive field of neurons is adaptively adjusted, which can more effectively capture discriminative features for distinguishing between true and false speech from feature maps of different sizes, improve the detection performance and generalization ability of the synthetic speech detection network, improve the accuracy of speech detection results, and realize the distinction between true and false speech data.

[0070] The present invention provides a speech detection method, apparatus, device, and storage medium, which are specifically described through the following embodiments. First, the speech detection method in the embodiments of the present invention is described.

[0071] This invention can acquire and process relevant data based on artificial intelligence (AI) technology. AI is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new type of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0072] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0073] The speech detection method provided in this invention relates to the field of artificial intelligence technology, and particularly to the field of data mining technology. The speech detection method provided in this invention can be applied to a terminal, a server, or software running on either a terminal or a server. The terminal communicates with the server via a network. The speech detection method can be executed by the terminal or the server, or by both the terminal and the server collaboratively. The speech detection method involves acquiring speech data; inputting the speech data into a trained synthetic speech detection model; the synthetic speech detection model extracting features from the speech data to obtain a speech feature map; then separating the speech feature map to obtain one or more separated feature maps; then performing attention weighting based on one or more separated feature maps to obtain an output feature map; pooling the output feature map to obtain a pooled feature map; and finally obtaining the speech detection result based on the pooled feature map.

[0074] In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, or smartwatch, etc. Additionally, the terminal can also be a smart in-vehicle device. The server can be a standalone server, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms; it can also be a service node in a blockchain system, where the service nodes form a peer-to-peer (P2P) network. The P2P protocol is an application layer protocol running on top of the Transmission Control Protocol (TCP). A server-side component for a voice detection method can be installed on the server, allowing interaction with the terminal. This can be achieved by installing corresponding software on the server, such as an application implementing the voice detection method, but not limited to these forms. The terminal and server can connect via Bluetooth, USB (Universal Serial Bus), or a network, etc., and this embodiment does not impose any limitations on these connections.

[0075] This invention can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0076] Figure 1 This is an optional flowchart of the speech detection method provided in the embodiments of the present invention. Figure 1 The method may include, but is not limited to, steps S110 to S170.

[0077] Step S110: Acquire voice data.

[0078] Step S120: Input the speech data into the trained synthetic speech detection model.

[0079] Step S130: Extract features from the speech data using the feature extraction module to obtain a speech feature map.

[0080] Step S140: Use the attention convolution module to perform feature separation on the speech feature map using one or more input convolutional layers to obtain one or more separated feature maps.

[0081] Step S150: The attention output layer performs attention weighting based on one or more separate feature maps to obtain the output feature map.

[0082] Step S160: The output feature map is pooled through the first pooling layer to obtain a pooled feature map.

[0083] Step S170: Obtain the speech detection result through the output module based on the pooling feature map.

[0084] In one embodiment, the speech detection method acquires speech data; inputs the speech data into a trained synthetic speech detection model, which extracts features from the speech data to obtain a speech feature map, then performs feature separation on the speech feature map to obtain one or more separated feature maps, and then performs attention weighting on the one or more separated feature maps to obtain an output feature map, which is then pooled to obtain a pooled feature map, and finally the speech detection result is obtained based on the pooled feature map. By using an end-to-end synthetic speech detection model for synthetic speech detection, the complexity of the detection model is reduced while introducing an attention mechanism, improving the accuracy of speech detection results and enabling the determination of the authenticity of speech data.

[0085] In one embodiment, in step S110, the speech data can be normal speech data or synthesized speech data. The synthesized speech data is synthesized speech obtained by converting text information into speech information using a speech synthesis method. This embodiment does not specifically limit the method of obtaining synthesized speech. Its purpose is to detect whether the input speech data is synthesized speech and to determine the authenticity of the speech data. If it is synthesized speech, the output result indicates that the speech data is fake speech; otherwise, it indicates that the speech data is real speech.

[0086] In one embodiment, to facilitate data processing, the speech data is converted into a corresponding spectrogram for subsequent processing. That is, the input to the synthesized speech detection model is a spectrogram or Mel spectrogram of the speech data. The type of spectrogram is not specifically limited here.

[0087] In one embodiment, after obtaining the speech data, step S120 inputs the speech data into the trained synthetic speech detection model for speech detection in subsequent steps.

[0088] In one embodiment, a synthetic speech detection model is trained using deep learning pre-training. This trained model is then ported to a target device for speech detection, yielding a speech detection result. Based on the result, it is determined whether the speech data is synthetic speech. The result can be either real speech or synthetic speech. The target device can be a terminal for different speech application scenarios, and is not specifically limited here.

[0089] In one embodiment, reference is made to Figure 2 The synthesized speech detection model 100 includes an input module 200, a feature extraction module 300, an attention module 400, and an output module 500. The input module 200 receives the speech data used for detection and sends it to the feature extraction module 300. The feature extraction module 300 is connected to the attention module 400, and the attention module 400 is connected to the output module 500. The output module 500 outputs the authenticity detection result of the speech data.

[0090] In one embodiment, reference is made to Figure 3 The synthesized speech detection model 100 includes at least one attention module 400 connected in sequence; three are illustrated in the figure. The first attention module 400 is connected to the feature extraction module 300, and the last attention module 400 is connected to the output module 500. The first attention module 400 receives the speech feature map output by the feature extraction module 300, and the other attention modules 400 receive the pooled feature map output by the previous attention module 400. The output module 500 obtains the speech detection result based on the pooled feature map output by the last attention module 400.

[0091] In one embodiment, reference is made to Figure 4 Each attention module 400 includes: an attention convolution module 410 and a first pooling layer 420. The attention convolution module 410 includes: a second convolution layer 411, one or more input convolution layers, an attention output layer 414, a third convolution layer 415, and a skip convolution layer 416. The input convolution layer includes: a first convolution layer 412 and a dilated convolution layer 413. The figure uses one first convolution layer 412 and one dilated convolution layer 413 as an example for illustration. The number of different types of input convolution layers is not limited here.

[0092] It is understood that all types of convolutional layers in the embodiments of this application include: grouped or depthwise convolution, batch normalization, and activation functions. The figures do not explicitly illustrate batch normalization and activation functions, but this does not mean that the convolutional layers in this application do not contain these components. In one embodiment, the activation function is the ReLU activation function.

[0093] In one embodiment, step S130 utilizes the feature extraction module 300 to extract features from the speech data, obtaining a speech feature map. (Refer to...) Figure 5 The feature extraction module 300 includes a feature extraction convolutional layer 310 and a feature extraction pooling layer 320.

[0094] In this embodiment, the feature extraction convolutional layer 310 extracts features from the input speech data (e.g., a spectrogram), that is, it performs convolution processing on the graph to obtain feature information. Then, the feature extraction pooling layer 320 downsamples the extracted information to obtain a speech feature map.

[0095] After obtaining the speech feature map, the speech feature map is input into the attention module for attention calculation. In one embodiment, referring to... Figure 4 The speech feature map first enters the attention convolution module 410 of the attention module 400 for attention operation, and then enters the first pooling layer 420 for pooling operation.

[0096] In one embodiment, the second convolutional layer 411 of the attention convolution module 410 is a convolutional layer with a 1x1 convolutional kernel. First, the speech feature map is convolved once, and then the processed speech feature map is input into one or more input convolutional layers and skip convolutional layers 416 respectively. The skip convolutional layers 416 mainly perform subsequent skip connection operations.

[0097] In one embodiment, since the input convolutional layer includes at least one first convolutional layer 412 and at least one dilated convolutional layer 413, the attention convolutional module in step S140 uses more than one input convolutional layer to perform feature separation on the speech feature map to obtain more than one separated feature map, and the corresponding separated feature map includes a first separated feature map and a dilated separated feature map.

[0098] In this embodiment, dilated convolutional layers are used to increase the receptive field while reducing computational cost, enabling the detection of large-scale targets and improving resolution. In addition, the dilated convolution operation can be configured with a scaling factor to capture multi-scale contextual information, thereby improving subsequent detection accuracy.

[0099] In one embodiment, reference is made to Figure 6 Step S140 includes, but is not limited to, steps S141 to S142.

[0100] Step S141: The speech feature map is separated by using at least one first convolutional layer through the attention convolution module to obtain at least one first separated feature map.

[0101] Step S142: The speech feature map is separated by using at least one dilated convolutional layer through the attention convolution module to obtain at least one dilated separation feature map.

[0102] In one embodiment, reference is made to Figure 4 This example uses two branches: a first convolutional layer 412 and a dilated convolutional layer 413. However, three or more branches can be set as needed, and no specific limitation is made here. In this embodiment, the kernel of the first convolutional layer 412 is 1x1, the kernel of the dilated convolutional layer 413 is 1x3, and the expansion rate is 2.

[0103] The input speech feature map is represented as: X∈R (L×C) Where L and C represent the number of rows and columns, respectively. The first convolutional layer has 412 pairs of speech features. Figure X Performing a convolution operation yields the corresponding first separation feature map. Represented as: 413 pairs of speech features from dilated convolutional layers Figure X Feature separation is performed to obtain the corresponding hole separation feature map. Represented as: in, This indicates that the first convolutional layer handles speech features. Figure X The operations performed include convolution, batch normalization, and ReLU activation. This indicates that the first convolutional layer handles speech features. Figure X The operations performed include convolution, batch normalization, and ReLU activation.

[0104] The first separation feature map mentioned above is obtained. Separation feature map of voids Then, the attention output layer 414 uses step S150 to perform attention weighting on the corresponding elements based on one or more separate feature maps to obtain the output feature map.

[0105] In one embodiment, the weighting operations of the attention output layer 414 include: a fusion operation, a compact feature computation operation, and a selection operation. (Refer to...) Figure 7 Step S150 includes, but is not limited to, steps S151 to S153.

[0106] Step S151: Obtain a fused feature map based on at least one first separation feature map and at least one hole separation feature map.

[0107] Step S152: Obtain compact feature vectors based on the fused feature maps.

[0108] Step S153: Attention weighting is applied to the separated feature map using compact feature vectors to obtain the output feature map.

[0109] In one embodiment, the fusion operation step S151 performs a fusion operation on the first separated feature map. Separation feature map of voids By adding corresponding elements, the separate feature maps of different branches are fused to obtain the fused feature map U, represented as:

[0110] After obtaining the fused feature map U of multiple branches, step S152 performs a compact feature calculation operation, referring to... Figure 8 Step S152 includes, but is not limited to, steps S1521 to S1522.

[0111] Step S1521: Pool the fused feature map to obtain a fused pooled feature map.

[0112] Step S1522: Perform a fully connected operation on the fused pooled feature map to obtain a compact feature vector.

[0113] In one embodiment, in step S1521, global average pooling is performed on the fused feature map U to obtain a fused pooled feature map S. Global average pooling refers to calculating an average value for all pixels of the feature map of each output channel. After global average pooling, the fused pooled feature map S is obtained, which is a feature vector with dimension C.

[0114] This embodiment uses global average pooling to take the average value of each feature map. The vector composed of the average values ​​of these feature maps is used as the fused pooling feature map S. Since there are no parameters to optimize in the global average pooling operation, the problem of overfitting can be avoided.

[0115] In one embodiment, the pooled feature map S∈R is fused. (C×1) The c-th element is calculated as follows:

[0116]

[0117] Among them, U c F represents the element in the c-th column of the fused feature map U. gp () denotes global average pooling operation, U c (i) represents the element in the i-th row and c-th column of the fused feature map U.

[0118] In one embodiment, step S1522 utilizes a fully connected layer to fully connect the fused pooled feature map S, obtaining a compact feature vector z∈R. (d×1) , where d is the dimension of the compact feature vector z. In this embodiment, the compact feature vector z is represented as:

[0119] z = F fc (S)=δ(B′(WS))

[0120] Among them, F fc () denotes a fully connected operation, δ() denotes the ReLU activation function, B′() denotes the batch normalization operation, and W∈R (d×C) This represents the coefficient matrix.

[0121] In one embodiment, the dimension d of the compact feature vector z is expressed as:

[0122] d = max(C / r, N)

[0123] Where r represents the reduction rate, which can be set according to actual needs or prior knowledge, and N represents the minimum value of d.

[0124] After obtaining the compact feature vector z, step S153 uses a selection operation to apply attention weighting to the separated feature map using the compact feature vector to obtain the output feature map V.

[0125] In one embodiment, reference is made to Figure 9 Step S153 includes, but is not limited to, steps S1531 to S1533.

[0126] Step S1531: Obtain a soft attention matrix for one or more separate feature maps.

[0127] Step S1532: Calculate the soft attention value using the compact feature vector and the soft attention matrix.

[0128] Step S1533: The separated feature map is weighted using soft attention values ​​to obtain the output feature map.

[0129] In one embodiment, the soft attention mechanism calculates weights for each feature map, with feature maps in the focal region receiving higher weights and feature maps in other regions receiving lower weights. Instead of selecting the highest probability from N pieces of information as the output, it calculates a weighted average of the N pieces of information as the output. In this embodiment, the soft attention matrix of the separated feature maps represents the weights of each dimension of information from different receptive fields, and the corresponding soft attention matrix is ​​calculated using the softmax function.

[0130] Reference Figure 4 To explain, the attention convolutional module 410 contains two branches: a first convolutional layer 412 and a dilated convolutional layer 413. Different branches correspond to different separation feature maps. The first separation feature map... The corresponding soft attention matrix is ​​A∈R (C×d) Void separation feature map The corresponding soft attention matrix is ​​B∈R (C×d) In one embodiment, soft attention matrices A and B are initialized before training. The soft attention matrices are continuously distributed in the range [0,1] and are a deterministic attention. The value is differentiable. The gradient is calculated through the corresponding network and the learning is optimized through forward propagation and backward feedback. After the learning process is completed, the soft attention matrix can be obtained.

[0131] The process of calculating the corresponding soft attention value using the soft attention matrix is ​​represented as follows:

[0132]

[0133]

[0134] Among them, A c It is the c-th row of A, a c Represents the first separation feature map The soft attention value, B c It is the c-th element of B, b c Characteristic map of cavity separation The soft attention value.

[0135] In one embodiment, the sum of the soft attention values ​​corresponding to different branches in the attention convolution module 410 is 1. Different weights are set for the branches, resulting in different convolution kernel sizes for different branches. This allows the network to select the appropriate convolution kernel according to its needs, applying the attention mechanism to the receptive field. This enables the network to adaptively change the receptive field for targets of different sizes, improving the model's detection performance and generalization ability. For example... Figure 4 The diagram shows two branches, therefore we have a c +b c =1, which allows the network to be expanded to multiple branches as needed.

[0136] In one embodiment, step S1533 utilizes a soft attention value a c and b c The separated feature maps are weighted to obtain the output feature map V∈R. (L×C) , represented as:

[0137]

[0138] Where V = [V1, V2, ..., V C ], V c ∈R L V c This represents the c-th value in the output feature map.

[0139] In one embodiment, after the attention output layer 414 obtains the output feature map V, it inputs the output feature map V into the third convolutional layer 415 to obtain the corresponding convolutional output. Simultaneously, it superimposes the output of the skip-layer convolutional layer 416 to obtain the output feature map V, thus achieving skip-layer connection. This skip-layer connection method combines the parallel structure of convolutions with different receptive fields in the attention convolutional module 410, combining the prediction results of the last layer with the feature predictions of shallower layers. This allows the prediction information of the last layer with global information to be combined with the shallower layers with more local details through direct connection, without adding extra parameters or computational load to the network. Simultaneously, it can increase the training speed of the model, improve training performance, and effectively solve the degradation problem.

[0140] After obtaining the output feature map V, step S160 involves the first pooling layer pooling the output feature map to obtain the pooled feature map.

[0141] In one embodiment, reference is made to Figure 10The synthesized speech detection model 100 includes at least one attention module 400 connected in sequence. The first attention module 400 is connected to the feature extraction module 300, and the last attention module 400 is connected to the output module 500. The first attention module 400 receives the speech feature map output by the feature extraction module 300, and the other attention modules 400 receive the pooled feature map output by the previous attention module 400. In this embodiment, the feature extraction pooling layer 320 of the feature extraction module 300 uses max pooling, only the first pooling layer 420 corresponding to the last attention module 400 uses global average pooling, and the first pooling layers 420 corresponding to the other attention modules 400 use max pooling. This approach is designed to reduce computational load, improve model prediction speed, enhance model performance, and reduce overfitting.

[0142] After obtaining the pooled feature map output by the last attention module 400 through the above steps, step S170 is executed, and the output module obtains the speech detection result based on the pooled feature map.

[0143] In one embodiment, reference is made to Figure 11 Step S170 includes, but is not limited to, steps S171 to S172.

[0144] Step S171: Perform at least one fully connected operation on the pooled feature map to obtain a fully connected output map.

[0145] Step S172: Perform softmax activation on the fully connected output graph to obtain the speech detection result.

[0146] In one embodiment, reference is made to Figure 10 The output module 500 includes three fully connected layers 510 and a softmax layer 520. The diagram is for illustrative purposes only and does not imply that the number of fully connected layers 510 in the output module 500 of this embodiment can only be three. This embodiment utilizes the fully connected layers 510 and the softmax layer 520 for activation operations to obtain the speech detection result of the input speech data. Based on the speech detection result, it is determined whether the speech data belongs to synthesized speech. The speech detection result can be either real speech or synthesized speech.

[0147] The speech detection method provided in this embodiment of the invention acquires speech data; inputs the speech data into a trained synthetic speech detection model; the synthetic speech detection model extracts features from the speech data to obtain a speech feature map; then performs feature separation on the speech feature map to obtain one or more separated feature maps; then performs attention weighting based on one or more separated feature maps to obtain an output feature map; then performs pooling on the output feature map to obtain a pooled feature map; finally, the speech detection result is obtained based on the pooled feature map.

[0148] In this embodiment, the speech detection method utilizes an end-to-end synthetic speech detection model to perform synthetic speech detection. This reduces the complexity of the detection model while introducing an attention mechanism. By adaptively weighting and combining receptive fields of different sizes, the size of the receptive field of neurons is adaptively adjusted. This enables more effective capture of discriminative features for distinguishing between true and false speech from feature maps of different sizes, thereby improving the detection performance and generalization ability of the synthetic speech detection network, increasing the accuracy of speech detection results, and achieving the distinction between true and false speech data.

[0149] In this embodiment, the speech detection method utilizes an end-to-end synthetic speech detection model to detect speech data. The synthetic speech detection model employed is a lightweight model whose attention module combines parallel structures of convolutions with different receptive fields using skip connections. Simultaneously, an attention mechanism is introduced, adaptively adjusting the receptive field size of neurons through adaptive weighted combinations of receptive fields of different sizes. This effectively captures discriminative features for identifying the authenticity of speech data from feature maps of different sizes, improving the detection performance and generalization ability of the synthetic speech detection process. This enhances its ability to resist security attacks involving forged speech, safeguarding public safety and property, and maintaining social stability.

[0150] This invention also provides a voice detection device that can implement the above-described voice detection method, see reference. Figure 12 The device includes:

[0151] The synthesized speech acquisition unit 1210 is used to acquire speech data.

[0152] In one embodiment, the speech data can be normal speech data or synthesized speech data. The synthesized speech data is synthesized speech obtained by converting text information into speech information using a speech synthesis method. This embodiment does not specifically limit the method of obtaining synthesized speech. Its purpose is to detect whether the input speech data is synthesized speech and to determine the authenticity of the speech data.

[0153] In one embodiment, to facilitate data processing, the speech data is converted into a corresponding spectrogram for subsequent processing. That is, the input to the synthesized speech detection model is a spectrogram or Mel spectrogram of the speech data. The type of spectrogram is not specifically limited here.

[0154] The input unit 1220 is used to input speech data into the trained synthetic speech detection model to obtain speech detection results.

[0155] The feature extraction unit 1230 is used to extract features from the speech data through the feature extraction module to obtain a speech feature map;

[0156] The feature separation unit 1240 is used to perform feature separation on the speech feature map through the attention convolution module using one or more input convolutional layers to obtain one or more separated feature maps.

[0157] The attention weighting unit 1250 is used to perform attention weighting on one or more separate feature maps through the attention output layer to obtain an output feature map;

[0158] Pooling unit 1260 is used to pool the output feature map through the first pooling layer to obtain a pooled feature map;

[0159] The detection unit 1270 is used to perform speech detection on the pooled feature map through the output module to obtain the speech detection result.

[0160] In one embodiment, the synthetic speech detection model includes an input module, a feature extraction module, an attention module, and an output module. The input module receives speech data for detection and sends it to the feature extraction module. The feature extraction module is connected to the attention module, the attention module is connected to the output module, and the output module outputs the authenticity detection result of the speech data.

[0161] In one embodiment, the synthetic speech detection model includes at least one attention module connected in sequence. The first attention module is connected to the feature extraction module, and the last attention module is connected to the output module. The first attention module receives the speech feature map output by the feature extraction module, and the other attention modules receive the pooling feature map output by the previous attention module. The output module obtains the speech detection result based on the pooling feature map output by the last attention module.

[0162] In one embodiment, the attention module includes an attention convolution module and a first pooling layer. The attention convolution module includes a second convolution layer, one or more input convolution layers, an attention output layer, a third convolution layer, and a skip convolution layer. The input convolution layer includes a first convolution layer and a dilated convolution layer. The number of different types of input convolution layers is not limited here.

[0163] It is understood that all types of convolutional layers in the embodiments of this application include: grouped or depthwise convolution, batch normalization, and activation functions. The figures do not explicitly illustrate batch normalization and activation functions, but this does not mean that the convolutional layers in this application do not contain these components. In one embodiment, the activation function is the ReLU activation function.

[0164] In one embodiment, the synthetic speech detection model extracts features from speech data to obtain a speech feature map. Then, it separates the speech feature map into one or more separate feature maps. Next, it performs attention weighting on these separate feature maps to obtain an output feature map. The output feature map is then pooled to obtain a pooled feature map. Finally, the speech detection result is obtained based on the pooled feature map. This approach reduces the complexity of the detection model while introducing an attention mechanism. Through adaptive weighting combinations of receptive fields of different sizes, it adaptively adjusts the receptive field size of neurons, enabling it to more effectively capture discriminative features for distinguishing true from false speech from feature maps of different sizes. This improves the detection performance and generalization ability of the synthetic speech detection network, thereby increasing the accuracy of speech detection results.

[0165] The specific implementation of the speech detection device in this embodiment is basically the same as the specific implementation of the speech detection method described above, and will not be repeated here.

[0166] This invention also provides an electronic device, comprising:

[0167] At least one memory;

[0168] At least one processor;

[0169] At least one program;

[0170] The program is stored in a memory, and the processor executes the at least one program to implement the speech detection method described above. The electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0171] Please see Figure 13 , Figure 13 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0172] The processor 1301 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention.

[0173] The memory 1302 can be implemented in the form of ROM (Read-Only Memory), static storage device, dynamic storage device, or RAM (Random Access Memory). The memory 1302 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1302 and is called and executed by the processor 1301 to execute the voice detection method of the embodiments of this invention.

[0174] The input / output interface 1303 is used to implement information input and output;

[0175] The communication interface 1304 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0176] Bus 1305 transmits information between various components of the device (e.g., processor 1301, memory 1302, input / output interface 1303, and communication interface 1304);

[0177] The processor 1301, memory 1302, input / output interface 1303 and communication interface 1304 are connected to each other within the device via bus 1305.

[0178] This invention also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the above-described speech detection method.

[0179] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0180] The speech detection method, speech detection device, electronic device, and storage medium proposed in this invention acquire speech data; input the speech data into a trained synthetic speech detection model; the synthetic speech detection model extracts features from the speech data to obtain speech feature maps; then, it separates the speech feature maps to obtain one or more separated feature maps; attention weighting is then applied to these separated feature maps to obtain an output feature map; pooling is then applied to the output feature map to obtain a pooled feature map; finally, the speech detection result is obtained based on the pooled feature map. In this embodiment, the speech detection method utilizes an end-to-end synthetic speech detection model for synthetic speech detection, reducing the complexity of the detection model while introducing an attention mechanism. Through adaptive weighted combination of receptive fields of different sizes, the size of the receptive field of neurons is adaptively adjusted, enabling more effective capture of discriminative features for distinguishing true from false speech from feature maps of different sizes. This improves the detection performance and generalization ability of the synthetic speech detection network, increases the accuracy of speech detection results, and achieves the distinction between true and false speech data.

[0181] The embodiments described in this invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of this invention, and do not constitute a limitation on the technical solutions provided by the embodiments of this invention. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this invention are also applicable to similar technical problems.

[0182] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present invention, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0183] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0184] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0185] The terms "first," "second," "third," "fourth," etc. (if present) in the specification and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0186] It should be understood that in this invention, "at least one (item)" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0187] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of 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 coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0188] 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.

[0189] Furthermore, the functional units in the various embodiments of the present invention 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.

[0190] 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, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0191] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the claims of the present invention.

Claims

1. A speech detection method, characterized in that, include: Acquire speech data, which includes synthesized speech data or real speech data; The speech data is input into a trained synthetic speech detection model. The synthetic speech detection model includes a feature extraction module, an attention module, and an output module. The feature extraction module is connected to the attention module, and the attention module is connected to the output module. The attention module includes an attention convolution module and a first pooling layer. The attention convolution module includes one or more input convolution layers and attention output layers. The input convolution layer includes a first convolution layer and a dilated convolution layer. The speech data is processed by the feature extraction module to obtain a speech feature map. The attention convolution module uses one or more input convolutional layers to perform feature separation on the speech feature map to obtain one or more separated feature maps, the separated feature maps including: a first separated feature map and a dilated separated feature map; The attention output layer performs attention weighting on one or more of the separated feature maps to obtain an output feature map; The output feature map is pooled through the first pooling layer to obtain a pooled feature map. The output module performs speech detection on the pooled feature map to obtain the speech detection result; Specifically, the attention convolution module uses one or more input convolutional layers to perform feature separation on the speech feature map, resulting in one or more separated feature maps, including: The attention convolution module uses at least one first convolutional layer to perform feature separation on the speech feature map, thereby obtaining at least one first separated feature map. The attention convolution module uses at least one dilated convolutional layer to perform feature separation on the speech feature map, resulting in at least one dilated feature map.

2. The speech detection method according to claim 1, characterized in that, The attention output layer performs attention weighting on one or more of the separated feature maps to obtain an output feature map, including: A fused feature map is obtained based on at least one of the first separation feature maps and at least one of the cavity separation feature maps; A compact feature vector is obtained based on the fused feature map; The output feature map is obtained by applying attention weights to the separated feature map using the compact feature vector.

3. The speech detection method according to claim 2, characterized in that, The compact feature vector obtained from the fused feature map includes: The fused feature map is pooled to obtain a fused pooled feature map; The fused pooled feature map is fully connected to obtain the compact feature vector.

4. The speech detection method according to claim 2, characterized in that, The step of applying attention weights to the separated feature map using the compact feature vector to obtain the output feature map includes: Obtain a soft attention matrix for one or more of the separated feature maps; The soft attention value is calculated using the compact feature vector and the soft attention matrix; The output feature map is obtained by weighting the separated feature map using the soft attention value.

5. The speech detection method according to any one of claims 1 to 4, characterized in that, The output module performs speech detection on the pooled feature map to obtain speech detection results, including: The pooled feature map is fully connected at least once to obtain a fully connected output map; The speech detection result is obtained by performing a softmax activation operation on the fully connected output graph.

6. The speech detection method according to any one of claims 1 to 4, characterized in that, The synthetic speech detection model includes: at least one attention module connected in sequence; The first attention module is connected to the feature extraction module, and the last attention module is connected to the output module. The first attention module receives the speech feature map, and the attention modules other than the first attention module receive the pooled feature map output by the previous attention module.

7. A voice detection device, characterized in that, include: A synthesized speech acquisition unit is used to acquire speech data; An input unit is used to input the speech data into a trained synthetic speech detection model. The synthetic speech detection model includes a feature extraction module, an attention module, and an output module. The feature extraction module is connected to the attention module, and the attention module is connected to the output module. The attention module includes an attention convolution module and a first pooling layer. The attention convolution module includes one or more input convolution layers and attention output layers. The input convolution layer includes a first convolution layer and a dilated convolution layer. A feature extraction unit is used to extract features from the speech data through the feature extraction module to obtain a speech feature map. The feature separation unit is used to perform feature separation on the speech feature map through the attention convolution module using one or more input convolutional layers to obtain one or more separated feature maps, wherein the separated feature map includes: a first separated feature map and a dilated separated feature map; An attention weighting unit is used to perform attention weighting on one or more of the separated feature maps through the attention output layer to obtain an output feature map; A pooling unit is used to pool the output feature map through the first pooling layer to obtain a pooled feature map. The detection unit is used to perform speech detection on the pooled feature map through the output module to obtain the speech detection result; Specifically, the attention convolution module uses one or more input convolutional layers to perform feature separation on the speech feature map, resulting in one or more separated feature maps, including: The attention convolution module uses at least one first convolutional layer to perform feature separation on the speech feature map, thereby obtaining at least one first separated feature map. The attention convolution module uses at least one dilated convolutional layer to perform feature separation on the speech feature map, resulting in at least one dilated feature map.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the speech detection method according to any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the speech detection method according to any one of claims 1 to 6.

Images