A speech feature extraction and detection method

Abstract

The application relates to a speech feature extraction and detection method. The application aims to solve the problems that an existing automatic speaker system is vulnerable to various cheating speech, the automatic speaker system cannot completely intercept, and false interception leads to low detection accuracy. The process is as follows: obtaining a training set and a verification set of preprocessed speech signals; extracting eCQSCC and FFV features; obtaining trained eCQSCC feature+Gaussian mixture model and FFV feature+Gaussian mixture model; outputting scores of the training set by the eCQSCC feature+Gaussian mixture model; outputting scores of the training set by the FFV feature+Gaussian mixture model; obtaining a pre-trained BosarisToolkit tool; obtaining a trained BosarisToolkit tool; and obtaining a fusion result of a to-be-detected speech signal. The application is used in the field of speech feature extraction and detection.

Description

Technical Field

[0001] This invention relates to methods for speech feature extraction and detection. Background Technology

[0002] Speaker recognition technology is a comprehensive technology that integrates knowledge from multiple fields. Because different people have different vocal cords, vocal tracts, and even lip shapes, as well as different vocal habits, the resulting voices will vary to varying degrees. These differences may be subtle, but after excellent feature extraction, these differences are gradually amplified, thus giving rise to the biometric feature of "voiceprint." Like fingerprints or iris features, voiceprint features offer good assurance in terms of reliability and uniqueness, thus meeting the prerequisites for biometric identification. Therefore, voiceprint recognition technology is now widely used in security fields such as financial security, social security, and communication security, as well as in smart homes. In addition, popular domestic payment software such as Alipay and WeChat have also implemented voiceprint locks for user information identification.

[0003] In recent years, the gradual development of voiceprint recognition technology has brought convenience to people in various fields of production and daily life. Simultaneously, with the increasing sophistication of machine learning, automatic speaker authentication systems have achieved higher recognition rates and better identification methods. However, at the same time, voiceprint synthesis technology is also gradually improving, which creates security risks. If someone maliciously uses various algorithms to synthesize a person's voiceprint, it will pose a significant threat to the security of automatic recognition systems. Furthermore, various portable small recording devices are constantly being developed, making it increasingly easy to secretly record someone's voiceprint. Recording and then playing back the audio can also impact users' information and property security. This kind of attack on automatic recognition systems using synthesized or played-back audio is collectively called spoofing. Automatic speaker systems are vulnerable to various spoofing attacks, so protecting automatic speaker authentication systems has become increasingly important, and the importance of spoofing detection (SD) is self-evident. Summary of the Invention

[0004] The purpose of this invention is to address the problems of existing automatic speaker systems being vulnerable to various deceptive speech attacks, incomplete or erroneous interception, and low detection accuracy, and to propose a speech feature extraction and detection method.

[0005] The specific process of a speech feature extraction and detection method is as follows:

[0006] Step 1: Obtain the audio database of the speech signal and divide the audio database of the speech signal into a training set and a validation set;

[0007] Step 2: Preprocess the training set and validation set of the speech signal in the audio database to obtain the preprocessed training set and validation set of the speech signal.

[0008] Step 3: Extract eCQSCC features from the training and validation sets of the preprocessed speech signals;

[0009] Step 4: Extract fundamental frequency variation (FFV) features from the training and validation sets of the preprocessed speech signals;

[0010] Step 5: Input the eCQSCC features of the training set of the preprocessed speech signal into the Gaussian mixture model for training until convergence, and obtain the trained eCQSCC features + Gaussian mixture model.

[0011] Step 6: Input the fundamental frequency variation (FFV) features of the training set of the preprocessed speech signal into the Gaussian mixture model for training until convergence, and obtain the trained FFV features + Gaussian mixture model.

[0012] Step 7: Input the training set into the trained eCQSCC feature + Gaussian mixture model, and the eCQSCC feature + Gaussian mixture model outputs the score of the training set.

[0013] Step 8: Input the training set into the trained FFV feature + Gaussian mixture model, and the FFV feature + Gaussian mixture model outputs the score of the training set;

[0014] Step 9: Use the BosarisToolkit tool to fuse the scores output in Step 7 and Step 8 into a score level, train the fusion process to obtain a pre-trained BosarisToolkit tool, and obtain the fusion result.

[0015] Step 10: Input the eCQSCC features of the validation set of the preprocessed speech signal into the trained eCQSCC feature + Gaussian mixture model, and the eCQSCC feature + Gaussian mixture model outputs the score of the validation set.

[0016] The FFV features of the validation set of the preprocessed speech signal are input into the trained FFV feature + Gaussian mixture model, and the FFV feature + Gaussian mixture model outputs the score of the validation set.

[0017] The pre-trained BosarisToolkit is used to fuse the scores of the validation set output by the eCQSCC feature + Gaussian mixture model and the scores of the validation set output by the FFV feature + Gaussian mixture model to obtain the fusion result. When the result meets the requirements, the trained BosarisToolkit is obtained. When the result does not meet the requirements, continue to step nine.

[0018] Step 11: Preprocess the speech signal to be tested to obtain the preprocessed speech signal to be tested;

[0019] eCQSCC features are extracted from the preprocessed speech signal to be tested, and the eCQSCC features of the preprocessed speech signal to be tested are obtained.

[0020] The fundamental frequency variation (FFV) feature is extracted from the preprocessed speech signal to obtain the fundamental frequency variation (FFV) feature of the preprocessed speech signal to be tested.

[0021] Step 12: Input the preprocessed eCQSCC features of the speech signal to be tested into the trained eCQSCC feature + Gaussian mixture model for detection. The eCQSCC feature + Gaussian mixture model outputs the score of the speech signal to be tested.

[0022] Step 13: Input the fundamental frequency change (FFV) feature of the preprocessed speech signal to be tested into the trained FFV feature + Gaussian mixture model for detection. The FFV feature + Gaussian mixture model outputs the score of the speech signal to be tested.

[0023] Step 14: Using the trained BosarisToolkit, the scores of the test speech signal output by the eCQSCC feature + Gaussian mixture model and the scores of the test speech signal output by the FFV feature + Gaussian mixture model are fused to obtain the fusion result of the test speech signal.

[0024] The specific process of a speech feature extraction and detection method is as follows:

[0025] Step 1: Obtain the audio database of the speech signal and divide the audio database of the speech signal into a training set and a validation set;

[0026] Step 2: Preprocess the training set and validation set of the speech signal in the audio database to obtain the preprocessed training set and validation set of the speech signal.

[0027] Step 3: Extract eCQSCC features from the training and validation sets of the preprocessed speech signals;

[0028] Step 4: Input the eCQSCC features of the training set of the preprocessed speech signal into the attention mechanism model for training;

[0029] Step 5: Input the eCQSCC features of the validation set of the preprocessed speech signal into the attention mechanism model for validation. If the result meets the requirements, the trained attention mechanism model is obtained; if the result does not meet the requirements, continue to step 4.

[0030] Step 6: Acquire the speech signal to be tested in the LA scene, preprocess the speech signal to be tested, and obtain the preprocessed speech signal to be tested.

[0031] The LA refers to logical access;

[0032] Step 7: Extract eCQSCC features from the preprocessed speech signal to be tested, and obtain the eCQSCC features of the preprocessed speech signal to be tested.

[0033] Step 8: Input the preprocessed eCQSCC features of the speech signal to be tested into the trained attention mechanism model for detection, and obtain the detection result of the speech signal to be tested.

[0034] The beneficial effects of this invention are as follows:

[0035] The purpose of this invention is to detect deceptive speech by focusing on two attack methods: playback deceptive speech attack and synthesis deceptive speech attack. It proposes a speech feature extraction and detection method, and fuses the features with prosodic features at a scoring level, inputting them into a deep residual neural network based on an attention mechanism to complete the deceptive speech detection.

[0036] Firstly, the extended Constant-Q Symmetric-subband Coefficients (eCQSCC) feature extraction method based on the phase symbol amplitude-phase spectrum adds linear information to the nonlinear information, achieving better performance. Simultaneously, the concept of feature fusion is introduced, fusing the eCQSCC features and Fundamental Frequency Variation (FFV) features of this invention at a score level, further improving the deception detection performance of the features. Using a Gaussian mixture model for detection, an EER of 6.78% and a t-DCF of 0.133 are obtained in the PA scenario, and an EER of 4.48% and a t-DCF of 0.124 are obtained in the LA scenario.

[0037] Secondly, based on the proposed novel features, a neural network with a residual structure is built. This network is used for feature learning, and channel attention and spatial attention mechanisms are added to further enhance the system's ability to detect spoofed speech. When using the network with the added attention mechanisms for detection, an EER of 0.04% and a t-DCF of 0.001 are achieved in the LA scenario. This surpasses the first place winner in the LA scenario of the ASVspoof2019 challenge. Attached Figure Description

[0038] Figure 1 Feature extraction diagram for eCQSCC;

[0039] Figure 2a For FFT frequency plot; Figure 2b CQT frequency diagram;

[0040] Figure 3a The graph shows the eCQSCC spoofing detection performance of EER in the PA scenario; Figure 3b The performance graph of eCQSCC spoofing detection for EER in LA scenario; Figure 3c Figure 3d shows the eCQSCC spoofing detection performance of t-DCF in the PA scenario; Figure 3d shows the eCQSCC spoofing detection performance of t-DCF in the LA scenario.

[0041] Figure 4 FFV feature extraction flowchart; Figure 5 Add a windowed view to the FFV; Figure 6 For FFV feature maps; Figure 7 Diagram of FFV filter;

[0042] Figure 8a The graph shows the FFV spoofing detection performance of EER in the PA scenario; Figure 8b The performance graph of FFV spoofing detection for EER in LA scenario; Figure 8c The graph shows the FFV spoofing detection performance of t-DCF in the PA scenario; Figure 8d The graph shows the FFV spoofing detection performance of t-DCF in the LA scenario;

[0043] Figure 9 Here is a block diagram of the feature fusion system;

[0044] Figure 10a The image shows the deception detection performance after EER score fusion in the PA scenario. Figure 10b The deception detection performance graph after EER score fusion in the LA scenario; Figure 10c The deception detection performance graph of t-DCF after score-level fusion in PA scenario; Figure 10d The deception detection performance of t-DCF after score-level fusion in LA scenarios is shown in the graph.

[0045] Figure 11 Here is a diagram of the SE module structure; Figure 12 Schematic diagram of CBAM; Figure 13 This is a diagram of the residual block structure. Figure 14 To add a network structure diagram for the attention mechanism;

[0046] Figure 15a Figure 15b shows the performance of residual network eCQSCC spoofing detection under the attention mechanism of EER in PA scenario; Figure 15b shows the performance of residual network eCQSCC spoofing detection under the attention mechanism of EER in LA scenario. Figure 15cThe graph shows the performance of residual network eCQSCC spoofing detection under the attention mechanism of t-DCF in PA scenarios. Figure 15d The graph shows the performance of residual network eCQSCC in spoofing detection under the attention mechanism of t-DCF in LA scenarios. Detailed Implementation

[0047] Specific Implementation Method 1: The specific process of a speech feature extraction and detection method in this implementation method is as follows:

[0048] Step 1: Obtain the audio database of the speech signal and divide the audio database of the speech signal into a training set and a validation set;

[0049] Step 2: Preprocess the training set and validation set of the speech signal in the audio database to obtain the preprocessed training set and validation set of the speech signal.

[0050] Step 3: Extract eCQSCC features from the training and validation sets of the preprocessed speech signals;

[0051] Step 4: Extract fundamental frequency variation (FFV) features from the training and validation sets of the preprocessed speech signals;

[0052] Step 5: Input the eCQSCC features of the training set of the preprocessed speech signal into the Gaussian mixture model for training until convergence, and obtain the trained eCQSCC features + Gaussian mixture model.

[0053] Step 6: Input the fundamental frequency variation (FFV) features of the training set of the preprocessed speech signal into the Gaussian mixture model for training until convergence, and obtain the trained FFV features + Gaussian mixture model.

[0054] Step 7: Input the training set into the trained eCQSCC feature + Gaussian mixture model, and the eCQSCC feature + Gaussian mixture model outputs the score of the training set.

[0055] Step 8: Input the training set into the trained FFV feature + Gaussian mixture model, and the FFV feature + Gaussian mixture model outputs the score of the training set;

[0056] Step 9: Use the BosarisToolkit tool to fuse the scores output in Step 7 and Step 8 into a score level, train the fusion process to obtain a pre-trained BosarisToolkit tool, and obtain the fusion result.

[0057] Step 10: Input the eCQSCC features of the validation set of the preprocessed speech signal into the trained eCQSCC feature + Gaussian mixture model, and the eCQSCC feature + Gaussian mixture model outputs the score of the validation set.

[0058] The FFV features of the validation set of the preprocessed speech signal are input into the trained FFV feature + Gaussian mixture model, and the FFV feature + Gaussian mixture model outputs the score of the validation set.

[0059] The pre-trained BosarisToolkit is used to fuse the scores of the validation set output by the eCQSCC feature + Gaussian mixture model and the scores of the validation set output by the FFV feature + Gaussian mixture model to obtain the fusion result. When the result meets the requirements, the trained BosarisToolkit is obtained. When the result does not meet the requirements, continue to step nine.

[0060] Step 11: Preprocess the speech signal to be tested to obtain the preprocessed speech signal to be tested;

[0061] eCQSCC features are extracted from the preprocessed speech signal to be tested, and the eCQSCC features of the preprocessed speech signal to be tested are obtained.

[0062] The fundamental frequency variation (FFV) feature is extracted from the preprocessed speech signal to obtain the fundamental frequency variation (FFV) feature of the preprocessed speech signal to be tested.

[0063] Step 12: Input the preprocessed eCQSCC features of the speech signal to be tested into the trained eCQSCC feature + Gaussian mixture model for detection. The eCQSCC feature + Gaussian mixture model outputs the score of the speech signal to be tested.

[0064] Step 13: Input the fundamental frequency change (FFV) feature of the preprocessed speech signal to be tested into the trained FFV feature + Gaussian mixture model for detection. The FFV feature + Gaussian mixture model outputs the score of the speech signal to be tested.

[0065] Step 14: Using the trained BosarisToolkit, the scores of the test speech signal output by the eCQSCC feature + Gaussian mixture model and the scores of the test speech signal output by the FFV feature + Gaussian mixture model are fused to obtain the fusion result of the test speech signal.

[0066] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that, in step one, the audio database of the speech signal is obtained, and the audio database of the speech signal is divided into a training set and a validation set; the specific process is as follows:

[0067] A portion of the replay audio data from the ASVspoof 2019 event database was selected as the replay audio dataset in the audio database;

[0068] Synthetic speech data from the ASVspoof 2019 competition database was selected as the synthetic speech dataset in the audio database;

[0069] The playback audio dataset consists of two audio sets: a training set and a validation set.

[0070] The synthesized speech dataset consists of two audio sets: a training set and a validation set.

[0071] The audio in the ASVspoof 2019 competition database includes both Physical Access (PA) spoofing detection and Logical Access (LA) spoofing detection. Physical Access is mainly for various playback speech spoofing scenarios, while Logical Access is for synthesized speech spoofing scenarios.

[0072] This project utilizes the ASVspoof 2019 competition database for screening. The audio from ASVspoof 2019 includes both Physical Access (PA) and Logical Access (LA) spoof detection. Physical Access primarily targets various playback speech spoofing scenarios, while Logical Access targets synthetic speech spoofing scenarios. Both datasets are developed based on the VCTK database and consist of three audio sets: a training set, a validation set, and an evaluation set.

[0073] For replaying audio, the data volume of the ASVspoof 2019PA portion is larger than that of the ASVspoof 2019LA portion. This is because the ASVspoof 2019 audio replay spoofing attack dataset has a relatively realistic recording scenario, including distance, room size, indoor and outdoor noise levels, etc. However, due to the limitations of the computer's performance, insufficient computer memory may occur during training. Therefore, a portion of the PA data is selected as samples to reduce the data volume while better evaluating the proposed handcrafted features. Since the selection only reduces the size of the training set, if better anti-spoofing performance is achieved with fewer training samples, it demonstrates that this feature extraction algorithm has certain advantages in SD (Speed ​​Deception). Based on different distances, room sizes, and indoor and outdoor noise levels, PA can be divided into approximately 243 different combinations. Here, each combination of audio samples is reduced to 100; audio samples with fewer than 100 samples are retained in their original quantity.

[0074] The other steps and parameters are the same as in Specific Implementation Method 1.

[0075] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that, in step two, the training set and validation set of the audio database of the speech signal are preprocessed to obtain the preprocessed training set and validation set of the speech signal; the specific process is as follows:

[0076] The transmission loss of high-frequency components of speech signals in the air is higher than that of low-frequency components. In order to compensate for the loss of high-frequency components and protect the information of the vocal tract, it is necessary to pre-emphasize the speech signal to achieve compensation for the high-frequency components.

[0077] A first-order FIR high-pass filter is used for speech pre-emphasis, with the transfer function H(z) = 1 - az. -1 'a' is the pre-emphasis coefficient, which ranges from 0.9 to 1. Here, we let a = 0.98.

[0078] Speech signals are non-stationary, and their characteristic parameters change over time. Macroscopically, speech signals are generated by the continuous movement of the oral cavity, leading to changes in the vocal tract. Changes in the vocal tract result in corresponding changes in the emitted speech signal. However, within a short timeframe, the changes in the vocal tract are very slow relative to the frequency of the speech signal; that is, the speech signal is short-term stationary. Therefore, signal analysis is required after the speech signal is framed. Generally, a frame length of less than 50ms can be considered as a stationary speech signal within the frame; this paper uses a frame length of 20ms. Frame-based processing of speech signals is equivalent to adding a rectangular window in the time domain, an operation that can lead to spectral leakage.

[0079] Step 21: Apply a first-order FIR high-pass filter to the speech signal for pre-emphasis. The process is as follows:

[0080] H(z) = 1 - az -1

[0081] Where a is the pre-emphasis coefficient, which ranges from 0.9 to 1; here we let a = 0.98; z represents the Z-transform; H(z) is the transfer function;

[0082] Step 22: Using a Hamming window to perform frame segmentation on the pre-emphasized speech signal can effectively reduce spectral leakage. Adding overlap between frames can make the characteristics of the entire speech change more smoothly. The overlapping part accounts for 50% of the total frame length, making the frame length of the segmented signal less than 50ms.

[0083] The Hamming window function is as follows:

[0084] w(n′)=0.54-0.46cos[2πn′ / (N′-1)]

[0085] Where 0≤n′≤N′-1, n′ represents the n′-th point in the window, N′ represents the window length, and w(n′) represents the Hamming window function.

[0086] Other steps and parameters are the same as in specific implementation method one or two.

[0087] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that, in step three, eCQSCC features are extracted from the training and validation sets of the preprocessed speech signals; the specific process is as follows:

[0088] This invention proposes an extended constant Q-symmetric subband cepstral coefficient extraction method based on the phase sign amplitude-phase spectrum, and the extraction block diagram is shown below:

[0089] Figure 1 The left-hand portion of the spectrum is not uniformly resampled and is an octave spectrum. The right-hand portion is uniformly resampled and is a linear spectrum. Fusing the linear spectrum and the octave spectrum can improve the characteristic performance.

[0090] Step 3: Perform Constant-Q Transform (CQT) processing on the preprocessed speech signal to obtain the CQT-transformed speech signal; the specific process is as follows:

[0091] In music, all musical scales are composed of twelve equal temperaments spanning several octaves. These twelve equal temperaments correspond to the twelve semitones within an octave on a piano. At the same pitch, the higher octave has twice the frequency of the lower octave. In other words, in music, sound is distributed exponentially, while the spectrum obtained from a traditional Fourier transform is linearly distributed. This means that the frequency points of the two cannot have a one-to-one correspondence, leading to errors in pitch frequency estimation. CQT (Cross-Q Transform) refers to a bandwidth where the center frequency follows an exponential distribution. Unlike the Fourier transform, the bandwidth of each filter is different, determined by the center frequency of that segment. The ratio of the center frequency to the bandwidth of each segment is always a constant Q. This constant Q transform, unlike the traditional Fourier transform, results in frequencies with the same distribution as the pitch scale, making it crucial in speech signal analysis.

[0092] In traditional Fourier transforms, the center frequency and bandwidth of each filter are independent, evenly spaced along the frequency axis, and the bandwidth is also the same. CQT filters, however, exhibit an octave distribution, which is quite useful for music analysis. Therefore, CQT has a significant advantage in this regard, offering higher time resolution at high frequencies and higher frequency resolution at low frequencies. Its difference from Fourier transforms is as follows: Figure 2a , 2b As shown.

[0093] The preprocessed speech signal is a discrete signal. For a discrete signal, the center frequency f of the k-th frequency band is...k It can be represented as

[0094] f k =2 (k-1) / b f1

[0095] Where f1 is the center frequency of the lowest frequency band; b is the number of spectral lines contained in an octave. In this paper, b = 96, which means that there are 96 spectral lines in each octave and 8 frequency components in each semitone.

[0096] Based on the center frequency f of the kth frequency band k Obtain a constant Q; the constant Q transform (CQT) has a center frequency to bandwidth ratio of constant Q, expressed as:

[0097]

[0098] in, Center frequency f k Bandwidth at the location;

[0099] The preprocessed speech signal is subjected to a constant Q-transform, which is represented as follows:

[0100]

[0101] Where x(n) is the preprocessed speech signal (time-domain signal); N k Let N be the window length that varies with frequency, satisfying N k =Q·f s / f k , k = 1, ..., K; K is the total number of frequency bands of the speech signal after Q-transformation; f s X is the sampling frequency of the preprocessed speech signal. cqt (k) represents the speech signal after Q-transformation with a constant constant; j represents the narration unit, j 2 =-1; For window functions;

[0102] in Using Hamming windows

[0103]

[0104] Step 32: Based on the constant-constant Q-transformed speech signal obtained in Step 31, calculate the amplitude-phase spectrum of the speech signal; the specific process is as follows:

[0105] X cqt (k) is the complex spectrum, as shown below:

[0106]

[0107] Among them, |X cqt (k)|and Let x(n) represent the amplitude spectrum and phase spectrum, respectively.

[0108] The By calculating X cqt The arctangent of the ratio of the imaginary part to the real part of (k) is obtained;

[0109] The value of is wrapped between -π and π, therefore It can be viewed as a wrapped phase;

[0110] For complex spectrum X cqt (k) Taking the logarithm, we obtain the complex spectrum at the base e logarithmic scale, as shown in the following expression:

[0111]

[0112] ln(X cqt The modulus of (k) is shown below:

[0113]

[0114] The amplitude-phase spectrum (MPS) of the speech signal can then be written as:

[0115]

[0116] Step 33: Divide the amplitude and phase spectrum of the speech signal obtained in Step 32 into symmetrical sub-bands (to ensure each sub-band has a different length, divide it into M sub-bands), and perform a discrete cosine transform on each sub-band to obtain the characteristics of each sub-band; the specific process is as follows:

[0117] Directly performing DCT on the entire spectrum ignores the deceptive information in the subbands, dividing the entire spectrum into individual subbands. The width of each subband is not the same, but changes according to the number of subbands, and they are symmetrically distributed.

[0118] Let the number of subbands be even, and the length of each subband be calculated using the following formula:

[0119]

[0120] in, Indicates the first The length of each sub-belt M represents the number of sub-bands; here, we take M = 16. min Indicates the length of the smallest subband;

[0121] Where L min The method for obtaining the result is as follows:

[0122]

[0123] Where K is the total number of frequency bands of the speech signal after Q transformation, and sum() represents summation;

[0124] The first MPS spectrum of the speech signal Each subband can be represented as

[0125]

[0126] in, Indicates intermediate variables. In the amplitude-phase spectrum, the first One point, In the amplitude-phase spectrum, the first One point, The first MPS spectrum representing the amplitude-phase spectrum of a speech signal Individual belt;

[0127] and Each satisfies

[0128]

[0129]

[0130] MPS spectrum The l-th frequency point of each sub-band can be represented as in

[0131] Perform Discrete Cosine Transform (DCT) on each sub-band separately;

[0132]

[0133]

[0134] Where p represents the p-th frequency band of the sub-band, p = 1, 2, ..., P-1; Represents the coefficients of the Discrete Cosine Transform (DCT);

[0135] At this time, the Features of sub-bands Can represent

[0136]

[0137] The MPS of the preprocessed speech signal (time domain signal) x(n) is then divided into symmetrical subbands, and the logarithm is taken to perform DCT transform on the feature.

[0138] Then the characteristic representation of all subbands

[0139]

[0140] Steps 3 and 4: Based on the constant-constant Q-transformed speech signal obtained in Step 31, perform linear spectrum PMPS processing on the constant-constant Q-transformed speech signal to obtain the linear spectrum PMPS-processed speech signal; the specific process is as follows:

[0141] While the MPS extraction process utilizes the magnitudes of amplitude and phase, it does not address the sign issue. Phase refers to the angle between the vector and the real axis, thus it can be positive or negative. Therefore, we consider adding phase sign information to the MPS to enhance deception detection capabilities.

[0142] The octave spectrum is shown below:

[0143]

[0144] Here, sign(·) represents taking the sign of it;

[0145] The CQT transform yields an octave spectrum. To further extract information from the linear spectrum, X... PMPS (k) Perform uniform resampling to convert the octave spectrum into linear spectrum information (MATLAB). The result is expressed using Y. PMPS (l′) represents; the specific process is as follows:

[0146] The octave scale (frequency domain range) is decomposed into d equal parts using the linear resampling period T′, where d is taken as 16; the linear frequency sampling rate F′ is then calculated.

[0147]

[0148] Using a multiphase anti-aliasing filter and spline interpolation method, the signal X is sampled at a uniform sampling rate F′. PMPS (k) Reconstruction, results are expressed using Y PMPS (l′) represents;

[0149] For Y PMPS (l′) Perform full-band DCT (same as all content under "Perform Discrete Cosine Transform (DCT) for each sub-band" in step 3.3); the result is used... express;

[0150] Where p′ represents the p′-th frequency point of the entire frequency band; l′ represents the l′-th frequency point of the entire frequency band;

[0151] Step 35: Based on Steps 33 and 34, perform dynamic information extraction to obtain the eCQSCC feature; the specific process is as follows:

[0152] δ-δ represents the acceleration coefficient; δ represents the velocity coefficient. The calculation is performed with N′ points as a group, and the upper and lower N′ points are connected. The acceleration coefficient is calculated based on δ with N′ points, so that the characteristics can be dynamically represented. Here, N′ is 3.

[0153] The formula for calculating δ is as follows:

[0154]

[0155] Among them, c t "" represents the signal characteristics of the t-th frame; N′ represents the currently selected frame. or The number of frames;

[0156] The formula for calculating δ-δ is expressed as follows:

[0157]

[0158] calculate Given the dynamic information of δ and δ-δ, the eCQSCC feature of frame t can be expressed as:

[0159]

[0160] The final features generated in this way include both sub-band features and overall linear spectrum features, while also adding phase sign information and dynamic information, theoretically enabling the acquisition of more deception information.

[0161] The Gaussian Mixture Model (GMM) was used to detect spoofed speech based on this feature, and the detection results are as follows: Figure 3a , 3b As shown in 3c and 3d:

[0162] As can be seen, in the PA scenario, the eCQSCC feature can achieve an EER of 7.57% and a t-DCF of around 0.155; in the LA scenario, it can achieve an EER of 6.29% and a t-DCF of around 0.17, demonstrating excellent performance.

[0163] Step 4: Extract fundamental frequency variation (FFV) features from the training and validation sets of the preprocessed speech signals; the specific process is as follows:

[0164] When a person speaks, the airflow causes the glottis to vibrate, creating a voiced sound. When the glottis does not vibrate, it is called a voiceless sound. The fundamental period refers to the duration of each opening and closing of the glottis when a voiced sound is produced. This vibration period is called the fundamental period, and its reciprocal is called the fundamental frequency. When a person produces a voiced sound, the airflow originates from the lungs and impacts the glottis, causing it to open and close, forming periodic pulses. These pulses, through the resonance of the vocal tract and the radiation from the lips, ultimately generate the speech information we hear. The fundamental period reflects this periodicity of the glottis.

[0165] The fundamental period and fundamental frequency are crucial parameters in speech signal analysis because they reflect the fundamental characteristics of the speaker's speech stimulus. They have wide applications in various speaker-related fields.

[0166] Because the prosody (referring to the fundamental frequency) of deceptive speech, especially synthesized speech, differs somewhat from that of natural speech, prosodic features of the speech signal can be used for deceptive speech detection. When synthesizing speech, it is assumed that the fundamental frequency is constant, and the target pitch is predicted using the average frequency of the source loudspeaker or the given text input. However, pitch variations may exist in reality, and the pitch variations of synthesized speech are not expected to resemble those of natural speech. Therefore, pitch variations in the speech signal are also an important clue for speech synthesis detection. Thus, to obtain pitch-related deceptive speech artifacts, fundamental frequency variation features are added.

[0167] Since the fundamental tone represents the first harmonic frequency in a speech signal, it can be considered a case of feature compression. To take advantage of the fact that all harmonics in two adjacent speech frames are equally spaced, and using each spectral element, Laskowski et al. introduced a novel method for estimating pitch variation. This makes the frame-level fundamental frequency variation (FFV) feature inherently multidimensional, which can be modeled using a Geometric Matrix Model (GMM).

[0168] The overall flowchart of the fundamental frequency variation FFV feature extraction process is as follows: Figure 4 As shown;

[0169] The steps for progressively extracting the fundamental frequency variation (FFV) features are as follows:

[0170] The training and validation sets of the preprocessed speech signals are input into the fundamental frequency variation (FFV) feature extraction model.

[0171] (a) Use two Hanning windows;

[0172] The two windows correspond to the left and right halves, respectively, which is F in the image above. L and FR Then calculate the 512-point Fast Fourier Transform and observe it in the frequency domain. These two window functions are as follows: Figure 5 As shown, corresponding to Figure 5 The "Add Window" section in the text.

[0173] (b) By calculating the same-size spectrum |F L |and|F R The vanishing dot product between | yields the FFV spectrum;

[0174] By normalizing the dot product to ∑|F L | 2 ×∑|F R | 2 Taking the square root of the equation, we obtain the final equation representing the energy independence of the FFV spectrum, as shown below.

[0175]

[0176] -N / 2+1≤n≤N / 2, for a fixed α, and The 512 point values ​​were determined using linear interpolation:

[0177]

[0178]

[0179] in

[0180]

[0181]

[0182] The limited 512 sampling points are arranged at equal intervals in the following locations:

[0183]

[0184] Where r∈{-N / 2,-N / 2+1,…,-1,0,+1,…,N / 2-2,N / 2-1}, The original value of the peak separation of the two window functions is FFV characteristics such as Figure 6 As shown:

[0185] (c) From Figure 6 As can be seen, the dimensionality of the FFV spectrum is very high. Here, passing it through a filter bank consisting of seven filters can reduce the dimensionality of the features;

[0186] Of the seven filters, one is for constant pitch, one is for pitches that rise slowly and quickly, one is for pitches that fall slowly and quickly, and two additional filters are for pitches that are indeterminate.

[0187] FFV filter bank such as Figure 7 As shown, the trapezoidal center filter and two rectangular additional filters are designed to capture meaningful rhythmic variations.

[0188] Because the silent frame FFV spectrum has a flat tail, rectangular additional filters are included in the filter bank structure. The region below each filter is considered uniform. This filter bank reduces the feature dimension of each speech frame from 512 to 7.

[0189] (d) Finally, the compressed FFV spectrum is decorrelated using Discrete Cosine Transform (DCT).

[0190] The final FFV coefficients can be modeled using GMM-based techniques in a frame-synchronized manner.

[0191] Its deception detection performance is as follows Figure 8a , 8b As shown in 8c and 8d:

[0192] Compared to eCQSCC, FFV's performance in spoofing detection is not particularly outstanding. However, there are two main reasons for introducing FFV: first, FFV can provide prosodic information that differs from amplitude and phase, serving as a supplement to the amplitude and phase-based features of eCQSCC; second, FFV has a small feature dimension, which can increase the amount of information in the speech signal with only a slight increase in overall dimensionality, thereby improving the system's spoofing detection capability.

[0193] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0194] Specific Implementation Method 5: This implementation method differs from one of the specific implementation methods 1 to 4 in that, in step 9, the BosarisToolkit tool is used to fuse the scores output in step 7 and step 8, the fusion process is trained to obtain a pre-trained BosarisToolkit tool, and the fusion result is obtained.

[0195] The specific process is as follows:

[0196] The eCQSCC feature + GMM model and the FFV feature + GMM model are both trained to convergence on their own. Each model will produce a score: positive values ​​represent real speech, negative values ​​represent fake speech, and larger absolute values ​​indicate higher confidence.

[0197] The BosarisToolkit tool takes as input the scores and labels of real and fake speech in a trained eCQSCC feature + Gaussian mixture model and a trained FFV feature + Gaussian mixture model.

[0198] The Bosaris Toolkit was used to perform score-level fusion. The fusion process was trained and the fusion result was also a score, which was positive or negative. Positive numbers represented real speech and negative numbers represented fake speech.

[0199] Obtain the pre-trained BosarisToolkit tool.

[0200] By fusing extracted features at different score levels, the overall performance can be further improved. Such a system possesses deception artifacts from two or more systems. Performing deception detection on the fused system can achieve better deception detection performance. Its block diagram is as follows: Figure 9 As shown;

[0201] After the input speech is trained by Model A and Model B, the scores of Model A and Model B are fused together, and the fused score is used as the overall score of the deception detection system to detect deceptive speech.

[0202] Both Model A and Model B use GMM, but they use different features.

[0203] After the input speech is trained on multiple models, the scores of each model are fused, and the fused score is used as the overall score of the deception detection system for deception speech detection. Many approaches fuse multiple models together; this paper only fuses the scores of two models, aiming to achieve better performance with fewer features. Models A and B both use GMM, but different features. Information is supplemented based on the proposed eCQSCC; the supplementary feature chosen here is FFV. The resulting fused information performs as follows in the PA and LA scenarios of the evaluation set: Figure 10a , 10b As shown in 10c and 10d;

[0204] From the perspective of the fusion-based deception detection performance, the fusion system shows a certain improvement in deception detection performance. The improvement is smaller in playback speech deception scenarios but more significant in synthesized speech deception scenarios. This is related to the features proposed in this paper, indicating that eCQSCC is more sensitive to synthesized speech but relatively weaker in detecting playback speech. This is reflected after system fusion.

[0205] This section summarizes the performance of some classic features for deception detection using GMM:

[0206] Table 2. Spoofing detection results based on GMM

[0207]

[0208] This shows that among the multiple features based on GMM, the fusion of eCQSCC and FFV achieved better results in the evaluation set.

[0209] In addition, the performance of features for deception detection based on the ASVspoof2019 evaluation set database and using a Gaussian mixture model was compared with those of features known in the past two or three years.

[0210] Table 3 shows the performance characteristics of each feature based on the ASVspoof2019 evaluation set.

[0211]

[0212] Several methods for feature extraction using the ASVspoof2019 database and deception detection using GMM modeling were compared. Some methods only performed feature deception detection performance in a single scene (PA or LA), and those not performed are indicated by "-" (the same applies below). Systems that fuse multiple models were also included.

[0213] After comparison, it can be found that the feature fusion system proposed in this paper has good deception detection performance in traditional machine learning such as GMM. It has the highest performance in the LA and PA scenarios among the compared features. The reason is that the features in this paper combine information with both linear spectrum and octave spectrum, and also adopt PMPS, which has better detection performance, especially in the LA scenario.

[0214] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0215] The evaluation index for verifying the results of this invention is:

[0216] Calculate the probability of equal errors; the specific process is as follows:

[0217] The commonly used detection metric in voice spoofing attack detection is the equal error rate (EER). Generally speaking, the higher the false rejection rate (FRR), the more stringent the system is, but it may also cause problems for normal users to be unable to recognize the voice. Conversely, the higher the false acceptance rate (FAR), the easier it is for users to pass the recognition, but the opportunity for spoofing attacks will also increase.

[0218] The false rejection rate is the percentage of users who falsely rejected a rating out of the total number of users who were judged to have the same rating.

[0219]

[0220] Where FRR(θ) represents the false rejection rate, N 相同用户但判定得分≤θ N represents the number of users who are the same user but whose judgment score is less than θ, i.e., incorrect judgments. θ represents the judgment threshold. 判定为相同用户 This indicates the number of users identified as the same.

[0221] Correspondingly, if the two voice messages are not actually from the same user, but are classified as belonging to the same user during the scoring process, this situation is called a false acceptance; the false acceptance rate can be expressed as:

[0222]

[0223] Where, N 不同用户但判定得分>θ N represents the number of different users whose judgment score is greater than θ, i.e., the number of incorrect acceptances. 判定为不同用户 This represents the number of users identified as different, and FAR(θ) represents the false acceptance rate.

[0224] According to the definition of equal error probability (EER), the formula for equal error probability (EER) is as follows:

[0225] EER=FRR(θ EER )=FAR(θ EER )

[0226] Where, θ EER FRR(θ) represents the decision threshold under equal error probabilities. EER FAR(θ) represents the false rejection rate under equal error probabilities. EER () represents the error acceptance rate under equal error probabilities;

[0227] The cascaded detection cost function; the specific process is as follows:

[0228] In practical applications, spoofing speech detection systems are cascaded with automatic speaker authentication systems. When a user authenticates their identity, the spoofing speech detection system checks whether the speech is spoofed (either through playback or synthesis). Only after passing this check can identity verification proceed. Alternatively, identity verification can be performed first, followed by spoofing detection. The final outputs of these two systems are logically ANDed; if one is negative, the entire output is negative. The tandem detection cost function (t-DCT) more accurately describes the system performance.

[0229] The Detection Cost Function (DCF) is defined as follows:

[0230]

[0231] Among them, C miss It is the cost of wrong rejection, C fa It is the price of wrong acceptance. and These represent the false rejection rate and false acceptance rate of the Automatic Speaker Authentication System (ASV), respectively; the smaller the DCF value, the better the performance of the ASV system; π tar Represents the prior probability of the target;

[0232] The automatic speaker authentication system and the spoofed speech detection system are cascaded, taking into account overall system performance; the cascaded detection cost function is defined as follows:

[0233]

[0234] in, This indicates the cost of an erroneous rejection by the automated speaker authentication system. This indicates the cost of an erroneous acceptance by the automated speaker authentication system. This represents the cost of deceiving the voice detection system into making an incorrect rejection. This represents the cost of deceiving the voice detection system into accepting incorrect information. This indicates the false rejection rate of the automatic speaker authentication system. This indicates the false acceptance rate of the automatic speaker authentication system. This indicates the false rejection rate of the deceptive voice detection system. π represents the false acceptance rate of a deceptive voice detection system. non π represents the prior probability of a non-target. spoof This represents the prior probability of a deception attack;

[0235] The t-DCF parameters used in the ASVspoof2019 challenge are shown in the table below.

[0236] Table 1. Parameter values ​​of the t-DCF cost function

[0237]

[0238] Specific Implementation Method Six: The specific process of a speech feature extraction and detection method in this implementation method is as follows:

[0239] Step 1: Obtain the audio database of the speech signal and divide the audio database of the speech signal into a training set and a validation set;

[0240] Step 2: Preprocess the training set and validation set of the speech signal in the audio database to obtain the preprocessed training set and validation set of the speech signal.

[0241] Step 3: Extract eCQSCC features from the training and validation sets of the preprocessed speech signals;

[0242] Step 4: Input the eCQSCC features of the training set of the preprocessed speech signal into the attention mechanism model for training;

[0243] Step 5: Input the eCQSCC features of the validation set of the preprocessed speech signal into the attention mechanism model for validation. If the result meets the requirements, the trained attention mechanism model is obtained; if the result does not meet the requirements, continue to step 4.

[0244] Step 6: Acquire the speech signal to be tested in the LA scene, preprocess the speech signal to be tested, and obtain the preprocessed speech signal to be tested.

[0245] The LA stands for Logical Access (LA).

[0246] Step 7: Extract eCQSCC features from the preprocessed speech signal to be tested, and obtain the eCQSCC features of the preprocessed speech signal to be tested.

[0247] Step 8: Input the preprocessed eCQSCC features of the speech signal to be tested into the trained attention mechanism model for detection, and obtain the detection result of the speech signal to be tested.

[0248] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Method Six in that, in step one, the audio database of the speech signal is obtained, and the audio database of the speech signal is divided into a training set and a validation set; the specific process is as follows:

[0249] Synthetic speech data from the ASVspoof 2019 competition database was selected as the synthetic speech dataset in the audio database;

[0250] The synthesized speech dataset consists of two audio sets: a training set and a validation set.

[0251] The audio in the ASVspoof 2019 event database includes both Physical Access (PA) spoofing detection and Logical Access (LA) spoofing detection.

[0252] Physical access is mainly for various playback voice spoofing scenarios, while logical access is for synthetic voice spoofing scenarios.

[0253] The other steps and parameters are the same as in Specific Implementation Method Seven.

[0254] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods Six or Seven in that, in step two, the training set and validation set of the audio database of the speech signal are preprocessed to obtain the preprocessed training set and validation set of the speech signal; the specific process is as follows:

[0255] The transmission loss of high-frequency components of speech signals in the air is higher than that of low-frequency components. In order to compensate for the loss of high-frequency components and protect the information of the vocal tract, it is necessary to pre-emphasize the speech signal to achieve compensation for the high-frequency components.

[0256] A first-order FIR high-pass filter is used for speech pre-emphasis, with the transfer function H(z) = 1 - az. -1 'a' is the pre-emphasis coefficient, which ranges from 0.9 to 1. Here, we let a = 0.98.

[0257] Speech signals are non-stationary, and their characteristic parameters change over time. Macroscopically, speech signals are generated by the continuous movement of the oral cavity, leading to changes in the vocal tract. Changes in the vocal tract result in corresponding changes in the emitted speech signal. However, within a short timeframe, the changes in the vocal tract are very slow relative to the frequency of the speech signal; that is, the speech signal is short-term stationary. Therefore, signal analysis is required after the speech signal is framed. Generally, a frame length of less than 50ms can be considered as a stationary speech signal within the frame; this paper uses a frame length of 20ms. Frame-based processing of speech signals is equivalent to adding a rectangular window in the time domain, an operation that can lead to spectral leakage.

[0258] Step 21: Apply a first-order FIR high-pass filter to the speech signal for pre-emphasis. The process is as follows:

[0259] H(z) = 1 - az -1

[0260] Where a is the pre-emphasis coefficient, which ranges from 0.9 to 1; here we let a = 0.98; z represents the Z-transform; H(z) is the transfer function;

[0261] Step 22: Using a Hamming window to perform frame segmentation on the pre-emphasized speech signal can effectively reduce spectral leakage. Adding overlap between frames can make the characteristics of the entire speech change more smoothly. The overlapping part accounts for 50% of the total frame length, making the frame length of the segmented signal less than 50ms.

[0262] The Hamming window function is as follows:

[0263] w(n′)=0.54-0.46cos[2πn′ / (N′-1)]

[0264] Where 0≤n′≤N′-1, n′ represents the n′-th point in the window, N′ represents the window length, and w(n′) represents the Hamming window function.

[0265] The other steps and parameters are the same as in specific implementation methods six or seven.

[0266] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods Six to Eight in that, in step three, eCQSCC features are extracted from the training and validation sets of the preprocessed speech signals; the specific process is as follows:

[0267] This invention proposes an extended constant Q-symmetric subband cepstral coefficient extraction method based on the phase sign amplitude-phase spectrum, and the extraction block diagram is shown below:

[0268] Figure 1 The left-hand portion of the spectrum is not uniformly resampled and is an octave spectrum. The right-hand portion is uniformly resampled and is a linear spectrum. Fusing the linear spectrum and the octave spectrum can improve the characteristic performance.

[0269] Step 3: Perform Constant-Q Transform (CQT) processing on the preprocessed speech signal to obtain the CQT-transformed speech signal; the specific process is as follows:

[0270] In music, all musical scales are composed of twelve equal temperaments spanning several octaves. These twelve equal temperaments correspond to the twelve semitones within an octave on a piano. At the same pitch, the higher octave has twice the frequency of the lower octave. In other words, in music, sound is distributed exponentially, while the spectrum obtained from a traditional Fourier transform is linearly distributed. This means that the frequency points of the two cannot have a one-to-one correspondence, leading to errors in pitch frequency estimation. CQT (Cross-Q Transform) refers to a bandwidth where the center frequency follows an exponential distribution. Unlike the Fourier transform, the bandwidth of each filter is different, determined by the center frequency of that segment. The ratio of the center frequency to the bandwidth of each segment is always a constant Q. This constant Q transform, unlike the traditional Fourier transform, results in frequencies with the same distribution as the pitch scale, making it crucial in speech signal analysis.

[0271] In traditional Fourier transforms, the center frequency and bandwidth of each filter are independent, evenly spaced along the frequency axis, and the bandwidth is also the same. CQT filters, however, exhibit an octave distribution, which is quite useful for music analysis. Therefore, CQT has a significant advantage in this regard, offering higher time resolution at high frequencies and higher frequency resolution at low frequencies. Its difference from Fourier transforms is as follows: Figure 2a , 2b As shown.

[0272] The preprocessed speech signal is a discrete signal. For a discrete signal, the center frequency f of the k-th frequency band is...k It can be represented as

[0273] f k =2 (k-1) / b f1

[0274] Where f1 is the center frequency of the lowest frequency band; b is the number of spectral lines contained in an octave. In this paper, b = 96, which means that there are 96 spectral lines in each octave and 8 frequency components in each semitone.

[0275] Based on the center frequency f of the kth frequency band k Obtain a constant Q; the constant Q transform (CQT) has a center frequency to bandwidth ratio of constant Q, expressed as:

[0276]

[0277] Where, δ fk Center frequency f k Bandwidth at the location;

[0278] The preprocessed speech signal is subjected to a constant Q-transform, which is represented as follows:

[0279]

[0280] Where x(n) is the preprocessed speech signal (time-domain signal); N k Let N be the window length that varies with frequency, satisfying N k =Q·f s / f k , k = 1, ..., K; K is the total number of frequency bands of the speech signal after Q-transformation; f s X is the sampling frequency of the preprocessed speech signal. cqt (k) represents the speech signal after Q-transformation with a constant constant; j represents the narration unit, j 2 =-1; For window functions; where Using Hamming windows

[0281]

[0282] Step 32: Based on the constant-constant Q-transformed speech signal obtained in Step 31, calculate the amplitude-phase spectrum of the speech signal; the specific process is as follows:

[0283] X cqt (k) is the complex spectrum, as shown below:

[0284]

[0285] Among them, |X cqt (k)|and Let x(n) represent the amplitude spectrum and phase spectrum, respectively.

[0286] The By calculating X cqt The arctangent of the ratio of the imaginary part to the real part of (k) is obtained;

[0287] The value of is wrapped between -π and π, therefore It can be viewed as a wrapped phase;

[0288] For complex spectrum X cqt (k) Taking the logarithm, we obtain the complex spectrum at the base e logarithmic scale, as shown in the following expression:

[0289]

[0290] ln(X cqt The modulus of (k) is shown below:

[0291]

[0292] The amplitude-phase spectrum (MPS) of the speech signal can then be written as:

[0293]

[0294] Step 33: Divide the amplitude and phase spectrum of the speech signal obtained in Step 32 into symmetrical sub-bands (to ensure each sub-band has a different length, divide it into M sub-bands), and perform a discrete cosine transform on each sub-band to obtain the characteristics of each sub-band; the specific process is as follows:

[0295] Directly performing DCT on the entire spectrum ignores the deceptive information in the subbands, dividing the entire spectrum into individual subbands. The width of each subband is not the same, but changes according to the number of subbands, and they are symmetrically distributed.

[0296] Let the number of subbands be even, and the length of each subband be calculated using the following formula:

[0297]

[0298] in, Indicates the first The length of each sub-band M represents the number of sub-bands; here, we take M = 16. min Indicates the length of the smallest subband;

[0299] Where L min The method for obtaining the result is as follows:

[0300]

[0301] Where K is the total number of frequency bands of the speech signal after Q transformation, and sum() represents summation;

[0302] The first MPS spectrum of the speech signal Each subband can be represented as

[0303]

[0304] in, Indicates intermediate variables. In the amplitude-phase spectrum, the first One point, In the amplitude-phase spectrum, the first One point, The first MPS spectrum representing the amplitude-phase spectrum of a speech signal Individual belt;

[0305] and Each satisfies

[0306]

[0307]

[0308] MPS spectrum The l-th frequency point of each sub-band can be represented as in

[0309] Perform Discrete Cosine Transform (DCT) on each sub-band separately;

[0310]

[0311]

[0312] Where p represents the p-th frequency band of the sub-band, p = 1, 2, ..., P-1; Represents the coefficients of the Discrete Cosine Transform (DCT);

[0313] At this time, the Features of sub-bands Can represent

[0314]

[0315] The MPS of the preprocessed speech signal (time domain signal) x(n) is then divided into symmetrical subbands, and the logarithm is taken to perform DCT transform on the feature.

[0316] Then the characteristic representation of all subbands

[0317]

[0318] Steps 3 and 4: Based on the constant-constant Q-transformed speech signal obtained in Step 31, perform linear spectrum PMPS processing on the constant-constant Q-transformed speech signal to obtain the linear spectrum PMPS-processed speech signal; the specific process is as follows:

[0319] While the MPS extraction process utilizes the magnitudes of amplitude and phase, it does not address the sign issue. Phase refers to the angle between the vector and the real axis, thus it can be positive or negative. Therefore, we consider adding phase sign information to the MPS to enhance deception detection capabilities.

[0320] The octave spectrum is shown below:

[0321]

[0322] Here, sign(·) represents taking the sign of it;

[0323] The CQT transform yields an octave spectrum. To further extract information from the linear spectrum, X... PMPS (k) Perform uniform resampling to convert the octave spectrum into linear spectrum information (MATLAB). The result is expressed using Y. PMPS (l′) represents; the specific process is as follows:

[0324] The octave scale (frequency domain range) is decomposed into d equal parts using the linear resampling period T′, where d is taken as 16; the linear frequency sampling rate F′ is then calculated.

[0325]

[0326] Using a multiphase anti-aliasing filter and spline interpolation method, the signal X is sampled at a uniform sampling rate F′. PMPS (k) Reconstruction, results are expressed using Y PMPS (l′) represents;

[0327] For Y PMPS (l′) Perform full-band DCT (same as all content under "Perform Discrete Cosine Transform (DCT) for each sub-band" in step 3.3); the result is used... express;

[0328] Where p′ represents the p′-th frequency point of the entire frequency band; l′ represents the l′-th frequency point of the entire frequency band;

[0329] Step 35: Based on Steps 33 and 34, perform dynamic information extraction to obtain the eCQSCC feature; the specific process is as follows:

[0330] δ-δ represents the acceleration coefficient; δ represents the velocity coefficient. The calculation is performed with N′ points as a group, and the upper and lower N′ points are connected. The acceleration coefficient is calculated based on δ with N′ points, so that the characteristics can be dynamically represented. Here, N′ is 3.

[0331] The formula for calculating δ is as follows:

[0332]

[0333] Where, c″ t N represents the signal characteristics of the t-th frame; N′ represents the currently selected frame. or The number of frames;

[0334] The formula for calculating δ-δ is expressed as follows:

[0335]

[0336] calculate Given the dynamic information of δ and δ-δ, the eCQSCC feature of frame t can be expressed as:

[0337]

[0338] The final features generated in this way include both sub-band features and overall linear spectrum features, while also adding phase sign information and dynamic information, theoretically enabling the acquisition of more deception information.

[0339] The Gaussian Mixture Model (GMM) was used to detect spoofed speech based on this feature, and the detection results are as follows: Figure 3a , 3b As shown in 3c and 3d:

[0340] As can be seen, in the LA scenario, an EER of 6.29% and a t-DCF of around 0.17 can be obtained, demonstrating excellent performance.

[0341] The other steps and parameters are the same as those in specific implementation methods six to eight.

[0342] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods Six to Nine in that the attention mechanism model is specifically as follows:

[0343] The network structure of the attention mechanism model is as follows: input layer, first convolutional unit, first BN layer, first ReLU layer, first convolutional attention module CBAM, first residual unit, second residual unit, third residual unit, fourth residual unit, second convolutional attention module CBAM, average pooling layer, FC layer, softmax classification layer;

[0344] The first convolutional unit is a first two-dimensional convolutional layer;

[0345] The first residual unit includes, in sequence: a first residual block, a second residual block, and a third residual block;

[0346] The second residual unit includes, in sequence: a fourth residual block, a fifth residual block, a sixth residual block, and a seventh residual block;

[0347] The third residual unit sequentially includes: the eighth residual block, the ninth residual block, the tenth residual block, the eleventh residual block, the twelfth residual block, the thirteenth residual block, the fourteenth residual block, the fifteenth residual block, the sixteenth residual block, the seventeenth residual block, the eighteenth residual block, the nineteenth residual block, the twentieth residual block, the twenty-first residual block, the twenty-second residual block, the twenty-third residual block, the twenty-fourth residual block, the twenty-fifth residual block, the twenty-sixth residual block, the twenty-seventh residual block, the twenty-eighth residual block, the twenty-ninth residual block, and the thirtieth residual block;

[0348] The fourth residual unit includes, in sequence, the thirty-first residual block, the thirty-second residual block, and the thirty-third residual block;

[0349] The structure of each residual block from the first residual block to the thirty-third residual block is as follows:

[0350] Each residual block consists of, in sequence: an input layer, a second two-dimensional convolutional layer, a second batch normalization (BN) layer, a second ReLU layer, a third two-dimensional convolutional layer, a third batch normalization (BN) layer, a third ReLU layer, and an output layer;

[0351] The connection relationships of each residual block are as follows:

[0352] The feature map is input into the input layer, and then passes through the second 2D convolutional layer, the second BN layer, the second ReLU layer, the third 2D convolutional layer, and the third BN layer in sequence. The feature map output from the third BN layer and the feature map input from the input layer are input into the third ReLU layer. The feature map output from the third ReLU layer is then output through the output layer.

[0353] Attention-based deceptive speech detection;

[0354] Attention Mechanism: In each convolutional layer of a convolutional neural network, a set of filters fuses spatial and channel information within the local receptive field. By using non-linear activation functions and downsampling operations, CNNs can capture and obtain the global theoretical receptive field. A crucial and unavoidable issue in network research is how to make the network focus on the parts that are most needed or contribute the most to the final result, emphasizing the capture of only the most distinctive attributes among the features to further improve the network's recognition ability.

[0355] Attention mechanisms are commonly used modules in neural network training, and with improvements, they have evolved into various models. However, the core of each is similar: to allow the network to focus on areas that require more attention. For example, in face recognition, it's desirable for the network to focus on facial information rather than blank areas. Similarly, in speech training, it's desirable for the network to pay more attention to the unique aspects of each individual, such as their stimulus sources and vocal tract information. Attention mechanisms are one way to achieve adaptive attention in the network. Attention mechanisms can be categorized into several types, including channel attention, spatial attention, a combination of channel and spatial attention, and self-attention, among others.

[0356] Obtaining the Convolutional Block Attention Module: In 2018, Woo et al. proposed the Convolutional Block Attention Module (CBAM), which combines channel attention and spatial attention mechanisms. When the network learns features, only task-relevant regions need attention. The essence of spatial attention is to locate the main target, perform relevant transformations, and obtain weights. Google DeepMind's Spatial Transformer Network (STN) is a typical example. It achieves better adaptability in preprocessing through learning and transforming the input. CBAM processes the input feature layers using both channel attention and spatial attention mechanisms, such as... Figure 12 .

[0357] Spoofing speech detection based on attention residual networks: The structure of the residual blocks used in this paper is as follows... Figure 13 As shown, each residual block first passes through a Conv2D layer (16 filters, 3×3 kernel size, stride 1, padding 1), then a batch normalization layer and a traditional ReLU activation function. After that, it passes through another Conv2D layer (16 filters, 3×3 kernel size, stride 1, padding 1) and another batch normalization layer. Skip connections are established by directly adding the input to the output. The result is then passed through another ReLU activation function as the output of a residual block. Dropout layers are not used here; instead, they are added to the overall network as a regularizer to reduce overfitting in the entire model.

[0358] To avoid disrupting the connections between residual blocks, attention modules are placed before and after multiple residual blocks. In other words, attention modules are not placed inside the residual blocks; instead, one attention mechanism module is placed before and after each of the four residual blocks. The network structure is as follows: Figure 14 As shown;

[0359] The attention mechanism modules are all CBAM, and max pooling has been removed during implementation. Inputting eCQSCC features, after 100 rounds of training, the deception detection performance is as follows: Figure 15a , 15b As shown in 15c and 15d;

[0360] The table below compares the performance of network models in the field of deception detection in recent years.

[0361] Table 4 Comparison of deception detection performance based on neural networks

[0362]

[0363]

[0364] As can be seen, the neural network with the proposed features and attention mechanism achieves excellent performance in the LA scenario, with EER and t-DCF as low as 0.04% and 0.001, respectively, far exceeding the performance of other methods in recent years. Furthermore, the first-place method in the ASVspoof2019 challenge for the LA scenario has an EER and t-DCF of 0.22% and 0.0069, respectively, which is weaker than the method presented in this paper.

[0365] The other steps and parameters are the same as those in one of the specific implementation methods six to nine.

[0366] The evaluation index for verifying the results of this invention is:

[0367] Calculate the probability of equal errors; the specific process is as follows:

[0368] The commonly used detection metric in voice spoofing attack detection is the equal error rate (EER). Generally speaking, the higher the false rejection rate (FRR), the more stringent the system is, but it may also cause problems for normal users to be unable to recognize the voice. Conversely, the higher the false acceptance rate (FAR), the easier it is for users to pass the recognition, but the opportunity for spoofing attacks will also increase.

[0369] The false rejection rate is the percentage of users who falsely rejected a rating out of the total number of users who were judged to have the same rating.

[0370]

[0371] Where FRR(θ) represents the false rejection rate, N 相同用户但判定得分≤θ N represents the number of users who are the same user but whose judgment score is less than θ, i.e., incorrect judgments. θ represents the judgment threshold. 判定为相同用户This indicates the number of users identified as the same.

[0372] Correspondingly, if the two voice messages are not actually from the same user, but are classified as belonging to the same user during the scoring process, this situation is called a false acceptance; the false acceptance rate can be expressed as:

[0373]

[0374] Where, N 不同用户但判定得分>θ N represents the number of different users whose judgment score is greater than θ, i.e., the number of incorrect acceptances. 判定为不同用户 This represents the number of users identified as different, and FAR(θ) represents the false acceptance rate.

[0375] According to the definition of equal error probability (EER), the formula for equal error probability (EER) is as follows:

[0376] EER=FRR(θ EER )=FAR(θ EER )

[0377] Where, θ EER FRR(θ) represents the decision threshold under equal error probabilities. EER FAR(θ) represents the false rejection rate under equal error probabilities. EER () represents the error acceptance rate under equal error probabilities;

[0378] The cascaded detection cost function; the specific process is as follows:

[0379] In practical applications, spoofing speech detection systems are cascaded with automatic speaker authentication systems. When a user authenticates their identity, the spoofing speech detection system checks whether the speech is spoofed (either through playback or synthesis). Only after passing this check can identity verification proceed. Alternatively, identity verification can be performed first, followed by spoofing detection. The final outputs of these two systems are logically ANDed; if one is negative, the entire output is negative. The tandem detection cost function (t-DCT) more accurately describes the system performance.

[0380] The Detection Cost Function (DCF) is defined as follows:

[0381]

[0382] Among them, C miss It is the cost of wrong rejection, C fa It is the price of wrong acceptance. and These represent the false rejection rate and false acceptance rate of the Automatic Speaker Authentication System (ASV), respectively; the smaller the DCF value, the better the performance of the ASV system; π tar Represents the prior probability of the target;

[0383] The automatic speaker authentication system and the spoofed speech detection system are cascaded, taking into account overall system performance; the cascaded detection cost function is defined as follows:

[0384]

[0385] in, This indicates the cost of an erroneous rejection by the automated speaker authentication system. This indicates the cost of an erroneous acceptance by the automated speaker authentication system. This represents the cost of deceiving the voice detection system into making an incorrect rejection. This represents the cost of deceiving the voice detection system into accepting incorrect information. This indicates the false rejection rate of the automatic speaker authentication system. This indicates the false acceptance rate of the automatic speaker authentication system. This indicates the false rejection rate of the deceptive voice detection system. π represents the false acceptance rate of a deceptive voice detection system. non π represents the prior probability of a non-target. spoof This represents the prior probability of a deception attack;

[0386] The t-DCF parameters used in the ASVspoof2019 challenge are shown in the table below.

[0387] Table 1. Parameter values ​​of the t-DCF cost function

[0388]

[0389] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for speech feature extraction and detection, characterized in that: The specific process of the method is as follows: Step 1: Obtain the audio database of the speech signal and divide the audio database of the speech signal into a training set and a validation set; Step 2: Preprocess the training set and validation set of the speech signal in the audio database to obtain the preprocessed training set and validation set of the speech signal. Step 3: Extract the extended constant Q-symmetric subband cepstral coefficients eCQSCC features based on the phase symbol amplitude-phase spectrum from the training and validation sets of the preprocessed speech signal; Step 4: Extract fundamental frequency variation (FFV) features from the training and validation sets of the preprocessed speech signals; Step 5: Input the eCQSCC features of the training set of the preprocessed speech signal into the Gaussian mixture model for training until convergence, and obtain the trained eCQSCC features + Gaussian mixture model. Step 6: Input the fundamental frequency variation (FFV) features of the training set of the preprocessed speech signal into the Gaussian mixture model for training until convergence, and obtain the trained FFV features + Gaussian mixture model. Step 7: Input the training set into the trained eCQSCC feature + Gaussian mixture model, and the eCQSCC feature + Gaussian mixture model outputs the score of the training set. Step 8: Input the training set into the trained FFV feature + Gaussian mixture model, and the FFV feature + Gaussian mixture model outputs the score of the training set; Step 9: Use the BosarisToolkit tool to fuse the scores output in Step 7 and Step 8 into a score level, train the fusion process to obtain a pre-trained BosarisToolkit tool, and obtain the fusion result. Step 10: Input the eCQSCC features of the validation set of the preprocessed speech signal into the trained eCQSCC feature + Gaussian mixture model, and the eCQSCC feature + Gaussian mixture model outputs the score of the validation set. The FFV features of the validation set of the preprocessed speech signal are input into the trained FFV feature + Gaussian mixture model, and the FFV feature + Gaussian mixture model outputs the score of the validation set. The pre-trained BosarisToolkit is used to fuse the scores of the validation set output by the eCQSCC feature + Gaussian mixture model and the scores of the validation set output by the FFV feature + Gaussian mixture model to obtain the fusion result. When the result meets the requirements, the trained BosarisToolkit is obtained. When the result does not meet the requirements, continue to step nine. Step 11: Preprocess the speech signal to be tested to obtain the preprocessed speech signal to be tested; eCQSCC features are extracted from the preprocessed speech signal to be tested, and the eCQSCC features of the preprocessed speech signal to be tested are obtained. The fundamental frequency variation (FFV) feature is extracted from the preprocessed speech signal to obtain the fundamental frequency variation (FFV) feature of the preprocessed speech signal to be tested. Step 12: Input the preprocessed eCQSCC features of the speech signal to be tested into the trained eCQSCC feature + Gaussian mixture model for detection. The eCQSCC feature + Gaussian mixture model outputs the score of the speech signal to be tested. Step 13: Input the fundamental frequency variation (FFV) feature of the preprocessed speech signal to be tested into the trained FFV feature + Gaussian mixture model for detection. The FFV feature + Gaussian mixture model outputs the score of the speech signal to be tested. Step 14: Using the trained BosarisToolkit, the scores of the test speech signal output by the eCQSCC feature + Gaussian mixture model and the scores of the test speech signal output by the FFV feature + Gaussian mixture model are fused to obtain the fusion result of the test speech signal. In step three, eCQSCC features are extracted from the training and validation sets of the preprocessed speech signals. The specific process is as follows: Step 3: Perform constant-constant Q-transform on the preprocessed speech signal to obtain the constant-constant Q-transformed speech signal; the specific process is as follows: The preprocessed speech signal is a discrete signal. For discrete signals, the first... The center frequency of each frequency band Represented as in, It is the center frequency of the lowest frequency band; The number of spectral lines contained within an octave; Based on the Center frequency of each frequency band Obtain the constant Q; expressed as in, Center frequency Bandwidth at the location; The preprocessed speech signal is subjected to a constant Q-transform, which is represented as follows: in, This is the preprocessed speech signal; Let be the window length that varies with frequency, satisfying , ; The total number of frequency bands in the speech signal after constant Q-transformation; The sampling frequency of the preprocessed speech signal. The speech signal after constant Q transformation; As a unit of narration, ; For window functions; in Using Hamming windows Step 32: Based on the constant-constant Q-transformed speech signal obtained in Step 31, calculate the amplitude-phase spectrum of the speech signal; the specific process is as follows: The complex spectrum is shown below: in, and They represent The amplitude spectrum and phase spectrum; The Through calculation The arctangent of the ratio of the imaginary part to the real part is obtained; ; For complex spectrum Taking the logarithm, we get The complex spectrum at a base-logarithmic scale is expressed as follows: The modulus is shown below: The amplitude-phase spectrum (MPS) of the speech signal is: Step 33: Divide the amplitude and phase spectrum of the speech signal obtained in Step 32 into symmetrical sub-bands, and perform discrete cosine transform on each sub-band to obtain the characteristics of each sub-band; the specific process is as follows: Let the number of subbands be even, and the length of each subband be calculated using the following formula: in, Indicates the first The length of each sub-belt ; Indicates the number of sub-bands. Indicates the length of the smallest subband; in The method for obtaining the result is as follows: in, The total number of frequency bands in the speech signal after the constant Q-transformation. To express summation; The first MPS spectrum of the speech signal The individual bands are: in, , Indicates intermediate variables. In the amplitude-phase spectrum, the first One point, In the amplitude-phase spectrum, the first One point, The first MPS spectrum representing the amplitude-phase spectrum of a speech signal Individual belt; and Each satisfies MPS spectrum The first of the sub-bands Each frequency point is ,in ; Perform Discrete Cosine Transform (DCT) on each sub-band separately; in, Indicates the first subband Each frequency band ; Represents the coefficients of the Discrete Cosine Transform (DCT); At this time, the Features of sub-bands for: Then the characteristic representation of all subbands Steps 3 and 4: Based on the constant-constant Q-transformed speech signal obtained in Step 31, perform linear spectrum PMPS processing on the constant-constant Q-transformed speech signal to obtain the linear spectrum PMPS-processed speech signal; the specific process is as follows: The octave spectrum is shown below: in, This indicates that a sign is taken from it; right Uniform resampling is performed to convert the octave spectrum into linear spectrum information, and the result is used... The specific process is as follows: Use linear resampling period Decompose the octave into d equal parts; solve for the linear frequency sampling rate. : Using multiphase anti-aliasing filters and spline interpolation methods to achieve a uniform sampling rate For signals Reconstruction, results used express; right Perform full-band DCT, and use the results express; in, Represents the first of the full-band frequency range One frequency point; Represents the first of the full-band frequency range One frequency point; Step 35: Based on Steps 33 and 34, perform dynamic information extraction to obtain the eCQSCC feature; the specific process is as follows: δ-δ represents the acceleration coefficient; δ represents the velocity coefficient; The formula for calculating δ is as follows: in, Indicates the first Signal characteristics of a frame; Indicates the currently retrieved or The number of frames; The formula for calculating δ-δ is expressed as follows: calculate , The dynamic information of δ and δ-δ, then the first The eCQSCC feature of the frame is: 。 2. The speech feature extraction and detection method according to claim 1, characterized in that: In step one, the audio database of the speech signal is obtained, and the audio database of the speech signal is divided into a training set and a validation set; the specific process is as follows: A portion of the replay audio data from the ASVspoof 2019 event database was selected as the replay audio dataset in the audio database; Synthetic speech data from the ASVspoof 2019 competition database was selected as the synthetic speech dataset in the audio database; The playback audio dataset consists of two audio sets: a training set and a validation set. The synthesized speech dataset consists of two audio sets: a training set and a validation set.

3. The speech feature extraction and detection method according to claim 2, characterized in that: In step two, the training set and validation set of the speech signal in the audio database are preprocessed to obtain the preprocessed training set and validation set of the speech signal; the specific process is as follows: Step 21: Apply a first-order FIR high-pass filter to the speech signal for pre-emphasis. The process is as follows: in, This is the pre-emphasis coefficient; Represents the Z-transform; For transfer functions; Step 22: Use a Hamming window to perform frame segmentation on the pre-emphasized speech signal, and add overlap between frames to make the frame length of the segmented signal less than 50ms. The Hamming window function is as follows: in, , Indicates the first in the window One point, Indicates the length of the window. This represents the Hamming window function.

4. The speech feature extraction and detection method according to claim 3, characterized in that: In step nine, the BosarisToolkit is used to fuse the scores output in step seven and step eight. The fusion process is trained to obtain a pre-trained BosarisToolkit, resulting in the fusion result. The specific process is as follows: The BosarisToolkit tool takes as input the scores and labels of real and fake speech in a trained eCQSCC feature + Gaussian mixture model and a trained FFV feature + Gaussian mixture model. The Bosaris Toolkit was used to perform score-level fusion. The fusion process was trained and the fusion result was also a score, which was positive or negative. Positive numbers represented real speech and negative numbers represented fake speech. Obtain the pre-trained BosarisToolkit tool.

5. A method for speech feature extraction and detection, characterized in that: The specific process of the method is as follows: Step 1: Obtain the audio database of the speech signal and divide the audio database of the speech signal into a training set and a validation set; Step 2: Preprocess the training set and validation set of the speech signal in the audio database to obtain the preprocessed training set and validation set of the speech signal. Step 3: Extract the extended constant Q-symmetric subband cepstral coefficients eCQSCC features based on the phase symbol amplitude-phase spectrum from the training and validation sets of the preprocessed speech signal; Step 4: Input the eCQSCC features of the training set of the preprocessed speech signal into the attention mechanism model for training; Step 5: Input the eCQSCC features of the validation set of the preprocessed speech signal into the attention mechanism model for validation. If the result meets the requirements, the trained attention mechanism model is obtained; if the result does not meet the requirements, continue to step 4. Step 6: Acquire the speech signal to be tested in the LA scene, preprocess the speech signal to be tested, and obtain the preprocessed speech signal to be tested. The LA refers to logical access; Step 7: Extract eCQSCC features from the preprocessed speech signal to be tested, and obtain the eCQSCC features of the preprocessed speech signal to be tested. Step 8: Input the preprocessed eCQSCC features of the speech signal to be tested into the trained attention mechanism model for detection, and obtain the detection result of the speech signal to be tested. In step three, eCQSCC features are extracted from the training and validation sets of the preprocessed speech signals. The specific process is as follows: Step 3: Perform constant-constant Q-transform on the preprocessed speech signal to obtain the constant-constant Q-transformed speech signal; the specific process is as follows: The preprocessed speech signal is a discrete signal. For discrete signals, the first... The center frequency of each frequency band Represented as in, It is the center frequency of the lowest frequency band; The number of spectral lines contained within an octave; Based on the Center frequency of each frequency band Obtain the constant Q; expressed as in, Center frequency Bandwidth at the location; The preprocessed speech signal is subjected to a constant Q-transform, which is represented as follows: in, This is the preprocessed speech signal; Let be the window length that varies with frequency, satisfying , ; The total number of frequency bands in the speech signal after constant Q-transformation; The sampling frequency of the preprocessed speech signal. The speech signal after constant Q transformation; As a unit of narration, ; For window functions; in Using Hamming windows Step 32: Based on the constant-constant Q-transformed speech signal obtained in Step 31, calculate the amplitude-phase spectrum of the speech signal; the specific process is as follows: The complex spectrum is shown below: in, and They represent The amplitude spectrum and phase spectrum; The Through calculation The arctangent of the ratio of the imaginary part to the real part is obtained; ; For complex spectrum Taking the logarithm, we get The complex spectrum at a base-logarithmic scale is expressed as follows: The modulus is shown below: The amplitude-phase spectrum (MPS) of the speech signal is: Step 33: Divide the amplitude and phase spectrum of the speech signal obtained in Step 32 into symmetrical sub-bands, and perform discrete cosine transform on each sub-band to obtain the characteristics of each sub-band; the specific process is as follows: Let the number of subbands be even, and the length of each subband be calculated using the following formula: in, Indicates the first The length of each sub-belt ; Indicates the number of sub-bands. Indicates the length of the smallest subband; in The method for obtaining the result is as follows: in, The total number of frequency bands in the speech signal after the constant Q-transformation. To express summation; The first MPS spectrum of the speech signal The individual bands are: in, , Indicates intermediate variables. In the amplitude-phase spectrum, the first One point, In the amplitude-phase spectrum, the first One point, The first MPS spectrum representing the amplitude-phase spectrum of a speech signal Individual belt; and Each satisfies MPS spectrum The first of the sub-bands Each frequency point is ,in ; Perform Discrete Cosine Transform (DCT) on each sub-band separately; in, Indicates the first subband Each frequency band ; Represents the coefficients of the Discrete Cosine Transform (DCT); At this time, the Features of sub-bands for: Then the characteristic representation of all subbands Steps 3 and 4: Based on the constant-constant Q-transformed speech signal obtained in Step 31, perform linear spectrum PMPS processing on the constant-constant Q-transformed speech signal to obtain the linear spectrum PMPS-processed speech signal; the specific process is as follows: The octave spectrum is shown below: in, This indicates that a sign is taken from it; right Uniform resampling is performed to convert the octave spectrum into linear spectrum information, and the result is used... The specific process is as follows: Use linear resampling period Decompose the octave into d equal parts; solve for the linear frequency sampling rate. : Using multiphase anti-aliasing filters and spline interpolation methods to achieve a uniform sampling rate For signals Reconstruction, results used express; right Perform full-band DCT, and use the results express; in, Represents the first of the full-band frequency range One frequency point; Represents the first of the full-band frequency range One frequency point; Step 35: Based on Steps 33 and 34, perform dynamic information extraction to obtain the eCQSCC feature; the specific process is as follows: δ-δ represents the acceleration coefficient; δ represents the velocity coefficient; The formula for calculating δ is as follows: in, Indicates the first Signal characteristics of a frame; Indicates the currently retrieved or The number of frames; The formula for calculating δ-δ is expressed as follows: calculate , The dynamic information of δ and δ-δ, then the first The eCQSCC feature of the frame is: 。 6. The speech feature extraction and detection method according to claim 5, characterized in that: In step one, the audio database of the speech signal is obtained, and the audio database of the speech signal is divided into a training set and a validation set; the specific process is as follows: Synthetic speech data from the ASVspoof 2019 competition database was selected as the synthetic speech dataset in the audio database; The synthesized speech dataset consists of two audio sets: a training set and a validation set.

7. The speech feature extraction and detection method according to claim 6, characterized in that: In step two, the training set and validation set of the speech signal in the audio database are preprocessed to obtain the preprocessed training set and validation set of the speech signal; the specific process is as follows: Step 21: Apply a first-order FIR high-pass filter to the speech signal for pre-emphasis. The process is as follows: in, This is the pre-emphasis coefficient; Represents the Z-transform; For transfer functions; Step 22: Use a Hamming window to perform frame segmentation on the pre-emphasized speech signal, and add overlap between frames to make the frame length of the segmented signal less than 50ms. The Hamming window function is as follows: in, , Indicates the first in the window One point, Indicates the length of the window. This represents the Hamming window function.

8. The speech feature extraction and detection method according to claim 7, characterized in that: The attention mechanism model is specifically as follows: The network structure of the attention mechanism model is as follows: input layer, first convolutional unit, first BN layer, first ReLU layer, first convolutional attention module CBAM, first residual unit, second residual unit, third residual unit, fourth residual unit, second convolutional attention module CBAM, average pooling layer, FC layer, softmax classification layer; The first convolutional unit is a first two-dimensional convolutional layer; The first residual unit includes, in sequence: a first residual block, a second residual block, and a third residual block; The second residual unit includes, in sequence: a fourth residual block, a fifth residual block, a sixth residual block, and a seventh residual block; The third residual unit sequentially includes: the eighth residual block, the ninth residual block, the tenth residual block, the eleventh residual block, the twelfth residual block, the thirteenth residual block, the fourteenth residual block, the fifteenth residual block, the sixteenth residual block, the seventeenth residual block, the eighteenth residual block, the nineteenth residual block, the twentieth residual block, the twenty-first residual block, the twenty-second residual block, the twenty-third residual block, the twenty-fourth residual block, the twenty-fifth residual block, the twenty-sixth residual block, the twenty-seventh residual block, the twenty-eighth residual block, the twenty-ninth residual block, and the thirtieth residual block; The fourth residual unit includes, in sequence, the thirty-first residual block, the thirty-second residual block, and the thirty-third residual block; The structure of each residual block from the first residual block to the thirty-third residual block is as follows: Each residual block consists of, in sequence: an input layer, a second two-dimensional convolutional layer, a second batch normalization (BN) layer, a second ReLU layer, a third two-dimensional convolutional layer, a third batch normalization (BN) layer, a third ReLU layer, and an output layer; The connection relationships of each residual block are as follows: The feature map is input into the input layer, and then passes through the second 2D convolutional layer, the second BN layer, the second ReLU layer, the third 2D convolutional layer, and the third BN layer in sequence. The feature map output from the third BN layer and the feature map input from the input layer are input into the third ReLU layer. The feature map output from the third ReLU layer is then output through the output layer.

Images