Voice deception detection method and device

Abstract

Embodiments of the present invention disclose a method and device for detecting voice spoofing, which can be used in the financial field or other technical fields. The method includes: slicing the voice data to be tested to obtain a sequence of voice data segments; extracting Mel-frequency cepstral coefficient features from each voice data segment in the sequence of voice data segments; inputting the Mel-frequency cepstral coefficient features into a two-dimensional convolutional network to obtain a first feature vector output by the two-dimensional convolutional network; inputting the voice data segment into a one-dimensional convolutional network to obtain a second feature vector output by the one-dimensional convolutional network; performing feature fusion on the first and second feature vectors to obtain a fused feature vector; and inputting the fused feature vector sequence into a preset prediction model to obtain a voice spoofing detection result corresponding to the voice data to be tested, as output by the prediction model. The present invention helps improve the accuracy of voice spoofing detection.

Description

Technical Field

[0001] The present invention relates to the field of voice forgery, and in particular to a method and device for detecting voice deception. Background Art

[0002] Existing voice spoofing detection solutions typically use raw audio data as input, automatically extracting features during training without the influence of manual filtering. However, methods based on raw audio data are not ideal for combating replay attacks (i.e., using recordings to deceive the recipient). Improving the ability of voice spoofing detection to combat replay attacks is an urgent problem that existing technologies need to solve. Summary of the Invention

[0003] In order to solve at least one technical problem in the above-mentioned background technology, the present invention proposes a voice deception detection method and device.

[0004] To achieve the above object, according to one aspect of the present invention, a method for detecting voice deception is provided, the method comprising:

[0005] Segment the voice data to be tested to obtain a sequence of voice data segments;

[0006] Extracting Mel-frequency cepstral coefficient features for each of the speech data segments in the speech data segment sequence;

[0007] Inputting the Mel-frequency cepstral coefficient feature into a two-dimensional convolutional network to obtain a first feature vector output by the two-dimensional convolutional network;

[0008] Inputting the speech data segment into a one-dimensional convolutional network to obtain a second feature vector output by the one-dimensional convolutional network;

[0009] Performing feature fusion on the first feature vector and the second feature vector to obtain a fused feature vector;

[0010] The fused feature vector sequence is input into a preset prediction model to obtain a voice deception detection result corresponding to the voice data to be tested output by the prediction model.

[0011] Optionally, before the voice data to be tested is segmented, the method further includes:

[0012] The data format of the voice data to be tested is converted into a single-channel matrix.

[0013] Optionally, before inputting the fused feature vector sequence into a preset prediction model, the method further includes:

[0014] The fused feature vector sequence is adjusted to the same dimension as the speech data to be tested through a preset autoencoder.

[0015] Optionally, inputting the fused feature vector sequence into a preset prediction model includes:

[0016] Each time, n consecutive fused feature vectors in the fused feature vector sequence are simultaneously input into the prediction model, where n is an integer greater than 1.

[0017] Optionally, the two-dimensional convolutional network specifically includes: a two-dimensional convolutional layer, an activation layer and a global average pooling layer; the output node of the two-dimensional convolutional network is the global average pooling layer.

[0018] Optionally, the one-dimensional convolutional network includes: a one-dimensional convolutional layer, an activation layer and a global average pooling layer; the output node of the one-dimensional convolutional layer is the global average pooling layer.

[0019] Optionally, the performing feature fusion on the first feature vector and the second feature vector specifically includes:

[0020] Feature fusion is performed on the first feature vector and the second feature vector through a fully connected layer.

[0021] Optionally, the prediction model is obtained by training a long short-term memory network and / or a gated recurrent unit.

[0022] In order to achieve the above object, according to another aspect of the present invention, a voice deception detection device is provided, the device comprising:

[0023] A voice data segmentation unit is used to segment the voice data to be tested to obtain a sequence of voice data segments;

[0024] a Mel-frequency cepstral coefficient feature extraction unit, configured to extract a Mel-frequency cepstral coefficient feature from each of the speech data segments in the speech data segment sequence;

[0025] A first feature vector generating unit is configured to input the Mel-frequency cepstral coefficient feature into a two-dimensional convolutional network to obtain a first feature vector output by the two-dimensional convolutional network;

[0026] a second feature vector generating unit, configured to input the speech data segment into a one-dimensional convolutional network to obtain a second feature vector output by the one-dimensional convolutional network;

[0027] a feature fusion unit, configured to perform feature fusion on the first feature vector and the second feature vector to obtain a fused feature vector;

[0028] The voice deception detection result determination unit is used to input the fused feature vector sequence into a preset prediction model to obtain the voice deception detection result corresponding to the voice data to be tested output by the prediction model.

[0029] To achieve the above-mentioned objective, according to another aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned voice deception detection method when executing the computer program.

[0030] To achieve the above objectives, according to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the steps of the above voice spoofing detection method are implemented.

[0031] To achieve the above object, according to another aspect of the present invention, a computer program product is provided, comprising a computer program / instruction, which implements the steps of the above voice spoofing detection method when executed by a processor.

[0032] The beneficial effects of the present invention are:

[0033] The present invention simultaneously utilizes manual features (i.e., the first eigenvector) and original speech data features (i.e., the second eigenvector), allowing the model to utilize both manual features and lossless original speech signals. The present invention fuses the two features and then extracts the global features of the sequence through a prediction model, enabling the model to better analyze the local and global information of the speech and achieve accurate detection of speech forgery attacks. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0035] Figure 1 is a flow chart of a method for detecting voice deception according to an embodiment of the present invention;

[0036] Figure 2 It is a schematic diagram of the network structure of the present invention;

[0037] Figure 3 This is a first flow chart of voice deception detection according to the present invention;

[0038] Figure 4 is a second flow chart of voice deception detection according to the present invention;

[0039] Figure 5 This is a structural block diagram of a voice deception detection device according to an embodiment of the present invention;

[0040] Figure 6 Schematic diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0042] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0043] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or apparatuses.

[0044] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0045] It should be noted that the acquisition, storage, use, and processing of data in the technical solution of this application comply with the relevant provisions of national laws and regulations.

[0046] It should be noted that the voice deception detection method and device of the present invention can be used in the financial field, and can also be used in any field other than the financial field. The application field of the voice deception detection method and device of the present invention is not limited.

[0047] This paper proposes an end-to-end voice spoofing detection method that integrates multiple features, including a deep neural network model that combines artificial features with raw data. To effectively extract voice features, this paper proposes a voice spoofing detection method based on deep feature fusion—a multi-scale voice model. This method is used to distinguish between speech conversion, text-to-speech, and replay attacks, providing security for scenarios such as voiceprint recognition and voice authorization.

[0048] Figure 1 This is a first flow chart of the voice deception detection method according to an embodiment of the present invention. Figure 1 As shown, in one embodiment of the present invention, the voice deception detection method of the present invention includes steps S101 to S106.

[0049] Step S101 : Segment the voice data to be tested to obtain a sequence of voice data segments.

[0050] In one embodiment of the present invention, before the voice data to be tested is segmented in step S101, the method of the present invention further includes:

[0051] The data format of the voice data to be tested is converted into a single-channel matrix.

[0052] The present invention first obtains the target voice data to be tested. The present invention uses the target voice data to determine whether the voice source is directly from a real human or forged data trained by algorithms such as adversarial learning and recorded and played back.

[0053] After obtaining the voice data to be tested, the present invention first converts the bit rate and format of the voice data to be tested, and the data format is unified into a single-channel matrix (1×L dimensions, L is the length of the data) to facilitate subsequent model processing.

[0054] Furthermore, the present invention segments the format-adjusted voice data to obtain voice data segments of equal length. For example, each small segment of the voice signal to be tested (e.g., 17,747 sampling points) is intercepted. Taking 16 kHz as an example, the original voice data contains 16,000 samples per second, and the dimension is 1×16,000.

[0055] Step S102: extracting Mel-frequency cepstral coefficient features from each of the speech data segments in the speech data segment sequence.

[0056] In one embodiment of the present invention, this step extracts Mel-frequency cepstral coefficient (MFCC) features for each speech data segment. The Mel-frequency cepstral coefficient feature extraction step includes: pre-emphasis, framing, windowing, calculating the power spectrum for each frame, passing through a set of triangular filters (the distribution of the filters in the frequency coordinates conforms to the Mel scale, which is constructed by simulating the nonlinear perception of the human ear to sound, for example, sensitive to low frequencies and insensitive to high frequencies), and obtaining the energy value of the frame in the filter through Fourier transform. The filter bank is then decorrelated using discrete cosine transform (DCT). Finally, the signal-to-noise ratio is improved by mean normalization.

[0057] The present invention generally uses 26 dimensions, representing 26 triangular filters obtained through Fourier transform, and 13 dimensions representing the cepstral features retained after the DCT operation, for a total of 39 dimensions. The length of the cepstral features represents the number of frames and is related to the frame size, frame shift, overlap length, and speech segment size set during framing. Mel-frequency cepstral coefficients are artificial features, and their calculation method artificially incorporates the human auditory perception of sound. This allows for a robust representation of audio data.

[0058] Step S103: input the Mel-frequency cepstral coefficient feature into a two-dimensional convolutional network to obtain a first feature vector output by the two-dimensional convolutional network.

[0059] In one embodiment of the present invention, the two-dimensional convolutional network specifically includes: a two-dimensional convolutional layer, an activation layer and a global average pooling layer; the output node of the two-dimensional convolutional network is the global average pooling layer.

[0060] like Figure 2 and Figure 3 As shown, in one embodiment of the present invention, the present invention uses the Mel-frequency cepstral coefficient feature (the two-dimensional matrix after MFCC extraction) as the input of the two-dimensional convolutional network, and the output node is the global average pooling layer to obtain a feature vector that matches the speech data segment. The two-dimensional convolutional network part includes three convolutional layers, an activation layer, and a global average pooling layer. The global average pooling layer is used to convert the two-dimensional features into one-dimensional features for further analysis of the MFCC features. This part automatically transforms and extracts features from the input in a deep learning manner. The MFCC data is introduced into this deep learning model to facilitate fusion with the deep extraction features.

[0061] Step S104: input the speech data segment into a one-dimensional convolutional network to obtain a second feature vector output by the one-dimensional convolutional network.

[0062] In one embodiment of the present invention, the one-dimensional convolutional network includes: a one-dimensional convolutional layer, an activation layer and a global average pooling layer; the output node of the one-dimensional convolutional layer is the global average pooling layer.

[0063] like Figure 2 and Figure 3 As shown, in one embodiment of the present invention, the present invention uses the speech data before the extraction of the Mel-frequency cepstral coefficient features, that is, the original speech data segment, as the input of the one-dimensional convolutional network, and the output node is the hidden layer of the one-dimensional convolutional network to obtain a feature vector that matches the speech data segment. The one-dimensional convolutional network comprises five one-dimensional convolutional layers, an activation layer, and a global average pooling layer. Compared with full connection, the global average pooling layer can effectively reduce the number of parameters of the neural network while compressing the feature dimension to speed up the convergence speed. The role of this step is to connect the speech data that has not been transformed (such as Fourier transform, Mel-frequency cepstral coefficients, etc.) to the neural network, to remove the subjectivity of artificial influence to the greatest extent, and to fully mine and utilize the speech data for subsequent combination and classification with the above-mentioned output.

[0064] Step S105 : performing feature fusion on the first feature vector and the second feature vector to obtain a fused feature vector.

[0065] In one embodiment of the present invention, the two feature extractions in step S103 and step S104 are performed simultaneously.

[0066] In one embodiment of the present invention, the step S105 of performing feature fusion on the first feature vector and the second feature vector specifically includes:

[0067] Feature fusion is performed on the first feature vector and the second feature vector through a fully connected layer.

[0068] For example, the two feature extraction networks described above output 32×1 dimensional data and 64×1 dimensional data, respectively. These two dimensions are the output dimensions of the two networks. The fully connected layer's input is 96×1, and the output can be reduced to 64 or 72. There is no fixed size here; it just needs to be smaller than the previous two dimensions. This concludes the feature extraction and fusion phase of the neural network of the present invention.

[0069] Step S106: input the fused feature vector sequence into a preset prediction model to obtain a voice deception detection result corresponding to the voice data to be tested output by the prediction model.

[0070] In one embodiment of the present invention, the prediction model is obtained by training a long short-term memory network and / or a gated recurrent unit.

[0071] In the present invention, after obtaining the fused feature vector corresponding to each speech data segment, the present invention constructs a fused feature vector sequence, and then inputs the fused feature vector sequence into a preset prediction model.

[0072] The present invention inputs the fused feature vector sequence into a prediction model to determine whether the voice data being tested is forged. The prediction model takes zt as input, and depending on the sample content, it can be designed to use either a single neuron output (binary classification, 0 for authenticity, 1 for forgery) to determine whether the target voice is a real human voice, or a multi-neuron output (outputting one-hot codes of several categories, each representing whether it is a corresponding category) to determine the specific source and attack method of the voice.

[0073] In one embodiment of the present invention, before inputting the fused feature vector sequence into a preset prediction model in step S106, the following steps are further included:

[0074] The fused feature vector sequence is adjusted to the same dimension as the speech data to be tested through a preset autoencoder.

[0075] like Figure 4 As shown, in one embodiment of the present invention, the present invention provides another feature fusion method. Before the above fused feature vector sequence is connected to the prediction model and trained, a single-layer autoencoder is constructed so that its output has the same dimension as the original data (the voice data to be tested after format adjustment) (for example, 16,000 dimensions) and the original voice data is reconstructed to minimize the reconstruction error. After convergence, the hidden layer representation is connected to the prediction model. Different from the previous fusion method (feature extraction and prediction are directly fully connected), in order to further fuse features, the use of an autoencoder is to restore the data to the original data after dimensionality reduction to obtain a better dimensionality reduction feature representation fusion effect. The unsupervised nature of the autoencoder can reduce the amount of data required and increase the convergence speed of subsequent training.

[0076] An autoencoder is a specialized network structure that compresses data from high dimensions to low dimensions and then restores it to high dimensions, obtaining an output identical to the input. Because the network can restore data from low dimensions to high dimensions after training, it is generally assumed that the low-dimensional representation of the data already contains all the required information. Its specific structure is also variable; for example, the decoder can use a single layer, multiple layers, or even a stacked layer. Therefore, after training, the low-dimensional to high-dimensional portion is removed, much like retaining only the top half of an hourglass, to achieve dimensionality reduction. In this paper, it is also used to fuse features in deep learning.

[0077] In one embodiment of the present invention, the step S106 of inputting the fused feature vector sequence into a preset prediction model specifically includes:

[0078] Each time, n consecutive fused feature vectors in the fused feature vector sequence are simultaneously input into the prediction model, where n is an integer greater than 1.

[0079] Optionally, the present invention provides another voice deception detection process. Since the prediction model can mine information from ordered sequences, during training and prediction, unlike the above two processes, the present invention can input the fused feature vectors of multiple consecutive groups of segments of the same voice at a time instead of the fused feature vectors of one voice segment at a time. Here, m represents the total number of truncation of a voice, and the output zt of the fully connected layer represents the feature vector of the t-th voice segment in the voice T. The input of the prediction model can be (z1, z2, z3), (z4, z5, z6)...(zm-2, zm-1, zm). In this way, longer-scale information can be mined. The network output is a label for training whether the feature vector sequence is forged.

[0080] In the present invention, the entire process from extracting features from the data to obtaining the judgment result is the feature fusion-based deep voice forgery detection model. Its parameters are trained using known deep forgery data. To ensure the uniformity of the model input dimensions, the training data is sliced ​​according to the input dimensions of the above model. The real human voice can be marked as 1 and the forged voice can be marked as 0 (if there are multiple categories, the labels are represented as a one-hot code). The data pairs of voice segments and whether they are forged are constructed to train the network.

[0081] In the present invention, the forgery judgment of the entire speech can be obtained by calculating the proportion of speech segments classified as fake. Let R be the detection result of the entire speech, and r be the detection result of the speech segment or speech segment sequence. When the number of speech segment or speech segment sequence detection results is greater than the threshold, R can be assigned to the corresponding category.

[0082] In one embodiment of the present invention, the three parts of the network (two feature extraction blocks and one prediction block) are trained together. Therefore, this solution provides an end-to-end supervised classification model. The current mainstream dataset is the ASVsPROOF dataset. Research generally takes the performance on this dataset as the basis, and this solution can be trained with this dataset. You can also construct your own dataset. The data in the self-constructed dataset is divided into two categories: real data and forged data generated by a variety of forgery algorithms. Forged voice data can be obtained in the following ways: 1) text-to-speech algorithm, voice conversion algorithm generation (logical access); 2) most of the real data, but some content is modified through the above steps (voice deep fake); 3) real voice attack records (physical access).

[0083] The network training method of the present invention can be to perform binary classification for each of the three tasks described above, with serial judgment, creating a separate network instance for each task. Alternatively, all fake data and real data can be combined for multi-classification, which only requires a single network instance, but is less effective than the first method. Another method only distinguishes whether the data is real or not, and all fake data are combined without further subclassification.

[0084] The present invention trains the network by feeding speech data and corresponding categories into batches until the loss function stabilizes. For autoencoders, the autoencoder portion needs to be trained first. Once training is complete, the decoder portion is removed, leaving the intermediate feature fusion layer connected back to the sequence network, which can then be trained with labeled data.

[0085] Table 1 below is a specific network structure example of a one-dimensional convolutional network according to an embodiment of the present invention.

[0086]

[0087] Table 1 Table 2 Table 2 is a specific network structure example of a two-dimensional convolutional network according to an embodiment of the present invention.

[0088] In summary, the present invention utilizes both manual features and original speech data, allowing the model to utilize both manual features and lossless original speech signals. Unlike existing fusion methods, the present invention dynamically fuses the two features in a deep learning model and automatically calculates the feature fusion results, rather than simply screening existing feature extraction methods. Local features of the speech sequence are extracted through convolutional neural networks and feature fusion, and then global features of the sequence are extracted through a sequence prediction network, such as a long short-term memory network, so that the model can better analyze the local and global information of the speech. Training on deep fake data can detect speech forgery attacks.

[0089] As can be seen from the above embodiments, the solution of the present invention can enable the model to introduce manual features, that is, additional artificial knowledge, while retaining and utilizing the original speech data. The two features are dynamically fused in the deep learning model to better integrate information. The fusion solution based on the autoencoder can also generate feature fusion modules in an unsupervised manner without relying on data labels. By extracting the features of the speech segment and then inputting them into the long short-term memory network as an element in the sample sequence, it is possible to pay attention to the local features of the speech and the global contextual relationships at the same time, and to mine information more comprehensively. It can effectively detect speech conversion, text-to-speech, playback, etc.

[0090] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0091] Based on the same inventive concept, an embodiment of the present invention also provides a voice deception detection device, which can be used to implement the voice deception detection method described in the above embodiment, as described in the following embodiment. Since the principle of solving the problem by the voice deception detection device is similar to that of the voice deception detection method, the embodiment of the voice deception detection device can refer to the embodiment of the voice deception detection method, and the repeated parts will not be repeated. As used below, the term "unit" or "module" can be a combination of software and / or hardware that implements the predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceived.

[0092] Figure 5 This is a first structural block diagram of a voice deception detection device according to an embodiment of the present invention. Figure 5 As shown, in one embodiment of the present invention, the voice deception detection device of the present invention includes:

[0093] The voice data segmentation unit 1 is used to segment the voice data to be tested to obtain a sequence of voice data segments;

[0094] Mel-frequency cepstral coefficient feature extraction unit 2, used for extracting Mel-frequency cepstral coefficient features for each of the speech data segments in the speech data segment sequence;

[0095] A first feature vector generating unit 3 is configured to input the Mel-frequency cepstral coefficient feature into a two-dimensional convolutional network to obtain a first feature vector output by the two-dimensional convolutional network;

[0096] A second feature vector generating unit 4 is configured to input the speech data segment into a one-dimensional convolutional network to obtain a second feature vector output by the one-dimensional convolutional network;

[0097] A feature fusion unit 5 is configured to perform feature fusion on the first feature vector and the second feature vector to obtain a fused feature vector;

[0098] The voice deception detection result determination unit 6 is used to input the fused feature vector sequence into a preset prediction model to obtain the voice deception detection result corresponding to the voice data to be tested output by the prediction model.

[0099] In one embodiment of the present invention, the voice deception detection device of the present invention further includes:

[0100] The voice data format conversion unit is used to convert the data format of the voice data to be tested into a single-channel matrix.

[0101] In one embodiment of the present invention, the voice deception detection device of the present invention further includes:

[0102] The autoencoder unit is used to adjust the fused feature vector sequence to the same dimension as the speech data to be tested through a preset autoencoder.

[0103] In one embodiment of the present invention, the voice deception detection result determination unit 6 specifically inputs n consecutive fused feature vectors in the fused feature vector sequence into the prediction model simultaneously each time, where n is an integer greater than 1.

[0104] In one embodiment of the present invention, the feature fusion unit 5 specifically performs feature fusion on the first feature vector and the second feature vector through a fully connected layer.

[0105] In order to achieve the above object, according to another aspect of the present application, a computer device is also provided. Figure 6 As shown, the computer device includes a memory, a processor, a communication interface and a communication bus. The memory stores a computer program that can be run on the processor. When the processor executes the computer program, the steps in the above embodiment method are implemented.

[0106] The processor may be a central processing unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0107] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and units, such as the corresponding program units in the above-described method embodiments of the present invention. The processor executes the non-transitory software programs, instructions, and modules stored in memory to perform various processor functions and work data processing, thereby implementing the methods in the above-described method embodiments.

[0108] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory may optionally include a memory remotely located relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0109] The one or more units are stored in the memory, and when executed by the processor, perform the method in the above embodiment.

[0110] The specific details of the above-mentioned computer device can be understood by referring to the corresponding descriptions and effects in the above-mentioned embodiments, and will not be repeated here.

[0111] In order to achieve the above-mentioned purpose, according to another aspect of the present application, a computer-readable storage medium is further provided, wherein the computer-readable storage medium stores a computer program, and the computer program implements the steps in the above-mentioned voice deception detection method when executed in a computer processor. It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiment method can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory (Flash Memory), a hard disk drive (HDD) or a solid-state drive (SSD), etc.; the storage medium can also include a combination of the above-mentioned types of memory.

[0112] In order to achieve the above-mentioned purpose, according to another aspect of the present application, a computer program product is further provided, comprising a computer program / instruction, which implements the steps of the above-mentioned voice deception detection method when executed by a processor.

[0113] Obviously, those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computing device. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0114] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for detecting voice deception, characterized in that: include: Segment the voice data to be tested to obtain a sequence of voice data segments; Extracting Mel-frequency cepstral coefficient features for each of the speech data segments in the speech data segment sequence; Inputting the Mel-frequency cepstral coefficient feature into a two-dimensional convolutional network to obtain a first feature vector output by the two-dimensional convolutional network; Inputting the speech data segment into a one-dimensional convolutional network to obtain a second feature vector output by the one-dimensional convolutional network; Performing feature fusion on the first feature vector and the second feature vector to obtain a fused feature vector; The fused feature vector sequence is input into a preset prediction model to obtain a voice deception detection result corresponding to the voice data to be tested output by the prediction model.

2. The voice deception detection method according to claim 1, characterized in that: Before the voice data to be tested is segmented, the method further includes: The data format of the voice data to be tested is converted into a single-channel matrix.

3. The voice deception detection method according to claim 1 or 2, characterized in that: Before inputting the fused feature vector sequence into a preset prediction model, the method further includes: The fused feature vector sequence is adjusted to the same dimension as the speech data to be tested through a preset autoencoder.

4. The voice deception detection method according to claim 1 or 2, characterized in that: The step of inputting the fused feature vector sequence into a preset prediction model includes: Each time, n consecutive fused feature vectors in the fused feature vector sequence are simultaneously input into the prediction model, where n is an integer greater than 1.

5. The voice deception detection method according to claim 1, wherein: The two-dimensional convolutional network specifically includes: a two-dimensional convolutional layer, an activation layer and a global average pooling layer; the output node of the two-dimensional convolutional network is the global average pooling layer.

6. The voice deception detection method according to claim 1, wherein: The one-dimensional convolutional network includes: a one-dimensional convolutional layer, an activation layer and a global average pooling layer; the output node of the one-dimensional convolutional layer is the global average pooling layer.

7. The voice deception detection method according to claim 1, wherein: The performing feature fusion on the first feature vector and the second feature vector specifically includes: Feature fusion is performed on the first feature vector and the second feature vector through a fully connected layer.

8. The voice deception detection method according to claim 1, wherein: The prediction model is obtained by training a long short-term memory network and / or a gated recurrent unit.

9. A voice deception detection device, characterized in that: include: A voice data segmentation unit is used to segment the voice data to be tested to obtain a sequence of voice data segments; a Mel-frequency cepstral coefficient feature extraction unit, configured to extract a Mel-frequency cepstral coefficient feature from each of the speech data segments in the speech data segment sequence; A first feature vector generating unit is configured to input the Mel-frequency cepstral coefficient feature into a two-dimensional convolutional network to obtain a first feature vector output by the two-dimensional convolutional network; a second feature vector generating unit, configured to input the speech data segment into a one-dimensional convolutional network to obtain a second feature vector output by the one-dimensional convolutional network; a feature fusion unit, configured to perform feature fusion on the first feature vector and the second feature vector to obtain a fused feature vector; The voice deception detection result determination unit is used to input the fused feature vector sequence into a preset prediction model to obtain the voice deception detection result corresponding to the voice data to be tested output by the prediction model.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

11. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

12. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images