A method for detecting a replay voice attack for a voiceprint security authentication system

Abstract

The application provides a replay voice attack detection method for a voiceprint security authentication system. First, the voice signal is preprocessed, then a plurality of sub-band signals are obtained through linear equal-width Gabor filters, each sub-band signal is processed through FDEO to obtain an instantaneous amplitude and an instantaneous frequency, then the instantaneous amplitude and the instantaneous frequency are respectively taken as inputs of an SENet to obtain enhanced IACC and IFCC features, and the features are respectively processed through windowing and averaging and discrete cosine transformation to obtain respective low-dimensional feature vectors. Then, the extracted IACC and IFCC feature vectors are respectively used to train respective Gaussian mixture model classifiers to obtain respective classifier model parameters. In detection, the IACC and IFCC feature vectors of the voice to be detected are respectively input into the respective GMM classifiers and are scored for credibility, and finally score level fusion is performed to realize discrimination of true and false voices.

Description

Technical Field

[0001] This invention belongs to the field of voiceprint recognition technology, specifically relating to a method for detecting replay speech attacks in a voiceprint security authentication system. Background Technology

[0002] Voiceprint recognition technology has been widely applied in many aspects of our daily lives. As one of the most popular biometric recognition technologies, it has unique advantages in areas such as remote authentication and has attracted increasing attention. Voiceprint authentication refers to extracting the required features from a speaker's spoken words and then using these features to decide whether to accept or reject the user. Currently, voiceprint authentication systems are vulnerable to attacks involving spoofed speech (including synthesized speech, converted speech, replay speech, and imitation speech). Replay speech, on the other hand, involves recording speech segments using recording equipment and playing them back using different speaker devices, which does not require extensive audio domain knowledge like synthesized or converted speech. Therefore, voiceprint authentication systems are easily vulnerable to replay speech attacks. Commonly used anti-replay attack methods employ feature vectors such as Mel-frequency cepstral coefficients, constant Q-cepstral coefficients, linear prediction cepstral coefficients, corrected group delay, and instantaneous frequency cepstral coefficients. However, the accuracy of these feature vectors in attack detection needs improvement and cannot adequately meet the security requirements of replay speech detection systems. Summary of the Invention

[0003] This invention proposes a method that combines the Frequency-Domain Energy Operator (FDEO) with Squeeze-and-Excitation Networks (SENet) to extract instantaneous amplitude cosine coefficients (IACC) and instantaneous frequency cosine coefficients (IFCC) with an attention mechanism. Then, using these two feature parameters as feature vectors and a Gaussian mixture model as a back-end classifier, a replay speech detection system is constructed to distinguish between genuine and fake speech.

[0004] First, the speech signal is preprocessed, then passed through a linear, uniformly wide Gabor filter to obtain several sub-band signals. The instantaneous amplitude and frequency of each sub-band signal are obtained using FDEO, and then used as inputs to SENet to obtain enhanced IACC and IFCC features. These are then processed by windowing averaging and discrete cosine transform to obtain their respective low-dimensional feature vectors. The extracted IACC and IFCC feature vectors are then used to train their respective Gaussian Mixture Model (GMM) classifiers, obtaining their respective classifier model parameters. During detection, the IACC and IFCC feature vectors of the test speech are input into their respective GMM classifiers and their confidence scores are assigned. Finally, the scores are fused to achieve the distinction between genuine and fake speech.

[0005] A method for detecting replay speech attacks in a voiceprint security authentication system includes the following steps:

[0006] Step 1: Preprocess the training speech;

[0007] Step 2: Pass the preprocessed speech signal through a Gabor filter bank to obtain N sub-band signals;

[0008] Step 3: Perform AM-FM analysis on the subband signal using the frequency domain energy operator FDEO to obtain the energy estimate of the speech signal, and obtain the instantaneous amplitude IACC and instantaneous frequency IFCC;

[0009] Step 4: The instantaneous amplitude and instantaneous frequency of the speech signal are used to obtain the feature vector after feature enhancement through the squeeze excitation network SENet;

[0010] Step 5: Obtain a low-dimensional feature vector representation from the feature vectors in Step 4 by averaging and performing DCT.

[0011] Step 6: Perform first-order and second-order difference operations and concatenation features to obtain IACC+△+△△ and IFCC+△+△△.

[0012] eigenvectors;

[0013] Step 7: Construct Gaussian mixture model parameters using the feature vectors of the training set, and input the speech to be tested into the classifier for discrimination.

[0014] Preferably, step 2 includes:

[0015] Let the sampling frequency be F. s Then the highest frequency of the sampled signal is F. s / 2, divide it into 200Hz equal-width filters, the number of filters is

[0016] The sign indicating rounding up indicates the center frequency is F. i =200i-100(Hz), i∈[1,N], therefore a filter bank consisting of N Gabor filters is needed, and the unit impulse response of the i-th filter is expressed as:

[0017] g i (n)=exp(-b 2 n 2 cos(Ω) i n) (1)

[0018] Ω i =2πF i T (2)

[0019] Among them, Ω i The center frequency is represented by , T is the sampling period, and b is used to control the bandwidth; after the i-th Gabor sub-band filter...

[0020] The speech signal expression after the filter can be represented as:

[0021] γ i (n)=s(n)*g i (n) (3)

[0022] Where the symbol * represents the convolution operation, s(n) is the speech signal, and g i (n) represents the impulse response of the i-th Gabor filter.

[0023] Preferably, step 3 includes:

[0024] γ i (n) x is obtained after N:1 downsampling. i (n), then x i (n) Obtain the energy estimate of the speech signal through FDEO; for discrete signal x i (n), using a three-point symmetric difference, is defined as:

[0025]

[0026] Then its frequency domain weighted energy operator discrete expression is:

[0027]

[0028] Where h i (n) is the signal x i (n) Hilbert transform, defined as h i (n)=H[x i (n)];

[0029] After estimating the energy of the speech signal, the instantaneous amplitude and instantaneous frequency of each sub-band signal are obtained using an energy separation algorithm.

[0030] The rate, expressed as:

[0031]

[0032]

[0033] Among them, a i [n] and ω i [n] represents the instantaneous amplitude and instantaneous frequency of the i-th sub-band signal, respectively.

[0034] Preferably, the SENet architecture in step 4 mainly consists of a squeezing layer and an excitation layer;

[0035] By extracting features from each frame of speech, the instantaneous amplitude feature vector a[n] = {a1[n], a2[n], ..., a...} is obtained. N [n]} and instantaneous frequency eigenvector ω[n]={ω1[n],ω2[n],……,ω N [n]}, thus for a speech signal with frame number τ, two [n] can be obtained.

[0036] Let A and Ω be two feature vector sequences.

[0037] A and Ω are used as inputs to the extrusion layer, collectively referred to as U = {u1, u2, ..., u}. c ,…,u N} T T represents the matrix transpose operator;

[0038] When the input is A, U = A; when the input is Ω, U = Ω; and c is the frequency exponent and c ∈ [1, N], u c ={u c (1),u c (2),……,u c (τ)}, and then the squeezed descriptor z = {z1, z2, ..., z} is obtained by global average pooling. c ,…,z N}, where the squeeze descriptor z c The expression is:

[0039]

[0040] After obtaining the squeeze descriptor, the weights of each feature dimension are obtained through matrix operations in the activation layer. The weight parameters are represented as s = {s1, s2, ..., s}. c ,…,s N}, its expression is:

[0041] s=σ(W2*δ(W1*z)) (9)

[0042] Where δ(·) represents the Rectified Linear Unit (ReLU) function, W1 and W2 represent the dimension reduction matrix of dimension a×N and the dimension increase matrix of dimension N×a, respectively, which are mainly used to reduce parameter calculations, and σ(·) represents the Sigmoid function.

[0043] The obtained weight parameter vector is multiplied by the original input vector sequence channels to obtain the weighted output x. c The expression is:

[0044] x c =F scale (u c ,s c (10)

[0045] Among them, F scale (·,·) denotes channel multiplication, then the hidden layer representation of the speech is X={x1,x2,…,x c ,…,x N} T ;

[0046] The hidden representation X of the speech is fed into a linear layer as a classifier for training. This linear layer outputs the probability of true or false speech, which is used to calculate the cross-entropy loss and update the parameters within the module. During testing, this linear layer is removed, and X is directly divided into feature vectors for each frame along the time dimension. The average of each frame is then calculated to obtain the vector-level representation of the speech, namely the enhanced instantaneous amplitude feature vector and instantaneous frequency feature vector, which are represented as follows:

[0047] a′={a1′,a2′,…,a i ′,…,a′ N} and ω′={ω1′,ω2′,…,ω i ′,…,ω′ N}

[0048] Preferably, in step 5:

[0049] The DCT expression is:

[0050]

[0051]

[0052] Where f(i) is the original signal, which is respectively the a mentioned above. i ′ and ω iF(u) is the coefficient after DCT transformation; where F(u) takes the first 13 coefficients, i.e., F(u), u=1,2,…13; N is the dimension of the eigenvector; ρ(u) is a compensation coefficient that makes the DCT transformation matrix an orthogonal matrix.

[0053] Preferably, in step 7:

[0054] For the eigenvector x∈R D×1 The likelihood function expression for its mixing density is:

[0055]

[0056] Where, ω i Let be the weight of the i-th Gaussian component, and have Gaussian probability density distribution function p i (x) is represented by the D×1 dimensional mean vector μ i The variance matrix Σ of D×D i Parameterization, the expression is:

[0057]

[0058] The model parameters are estimated using an iterative mean-maximization algorithm, the expression of which is as follows:

[0059] θ (i+1) =arg max θ E[logp(x,γ|θ)|x,θ (i) (15)

[0060] Where, θ (i) This represents the value of θ obtained in the i-th iteration, where θ represents the parameter set. γ represents the response of different Gaussian functions to the input sample x.

[0061] The essential features of this invention are as follows: Based on the differences in spectral envelope and harmonic structure between genuine and pseudo-speech, FDEO is used to extract the IACC and IFCC feature vectors of the speech signal, solving the problem of complex phase unfolding in traditional AM-FM speech signal analysis. The frequency domain energy operator used employs Hilbert transform, which, compared to the Teager energy operator, avoids negative energy, making it more practical and offering stronger noise resistance. SENet is used to enhance the extracted features, making them more directional and increasing the difference between genuine and pseudo-speech in the feature vectors. Fractional fusion takes into account both amplitude and frequency information of the speech signal, making the entire system more robust to replayed speech. Attached Figure Description

[0062] Figure 1Flowchart of replay speech detection using FDEO;

[0063] Figure 2 SENet architecture diagram. Detailed Implementation

[0064] The technical solution of the present invention will be further described in detail below through specific embodiments and in conjunction with the accompanying drawings.

[0065] Example 1

[0066] Figure 1 The following describes the system structure for replaying spoofed speech detection according to the present invention, and the working principle of each module in the system will be introduced next.

[0067] The proposed replay speech detection system consists of three parts: feature extraction, feature enhancement, and back-end classification. The system utilizes FDEO to perform AM-FM analysis on the detected speech, obtaining the instantaneous frequency and amplitude of the speech signal. Since the harmonic structure and spectral envelope of the replay speech differ from those of the real speech, the system extracts IACC and IFCC feature vectors to distinguish these differences. This feature vector is then input into a Gaussian mixture model for training to obtain a true / false classification model. Finally, the feature vector of the detected speech is input into the classification model to identify true / false speech. This entire process can distinguish between real and replay speech, thus achieving replay speech detection.

[0068] The detailed algorithm process is as follows:

[0069] (1) The preprocessed speech signal is passed through a Gabor filter bank to obtain N sub-band signals.

[0070] (2) FDEO is used to perform AM-FM analysis on the subband signal to obtain the energy estimate of the speech signal and the instantaneous amplitude and instantaneous frequency.

[0071] (3) The instantaneous amplitude and instantaneous frequency of the speech signal are used to obtain the feature vector after feature enhancement through SENet.

[0072] (4) Low-dimensional feature vector representation is obtained by averaging and DCT.

[0073] (5) Perform first-order and second-order difference operations and concatenation features to obtain the eigenvectors IACC+△+△△ and IFCC+△+△△.

[0074] (6) Construct Gaussian mixture model parameters using the feature vectors of the training set, and input the speech to be tested into the classifier for discrimination.

[0075] The feature extraction module utilizes FDEO to extract IACC and IFCC. Compared to real speech, the spectral energy of replayed speech is distorted due to playback equipment and channel characteristics, resulting in significant harmonic distortion in the spectral envelope. Therefore, the authenticity of the speech can be determined by extracting instantaneous frequency and amplitude features. The frequency domain energy operator is used to estimate the energy of the real signal, and an energy separation algorithm preserves amplitude and frequency information, which not only reduces computational complexity but also improves time and frequency resolution.

[0076] Extracting IACC and IFCC requires preprocessing the speech signal under test, such as sampling and pre-emphasis, and then obtaining the sub-band signal through a bandpass filter. Here, a linear Gabor filter with equal width is used. If the sampling frequency is F... s Then the highest frequency of the sampled signal is F. s / 2, divide it into 200Hz equal-width filters, the number of filters is The sign indicating rounding up indicates the center frequency is F. i =200i-100(Hz), i∈[1,N], therefore a filter bank consisting of N Gabor filters is needed, and the unit impulse response of the i-th filter is expressed as:

[0077] g i (n)=exp(-b 2 n 2 cos(Ω) i n) (1)

[0078] Ω i =2πF i T (2)

[0079] Among them, Ω i Let be the center frequency, T be the sampling period, and b be used to control the bandwidth. The speech signal after passing through the i-th Gabor subband filter can be expressed as:

[0080] γ i (n)=s(n)*g i (n) (3)

[0081] Where the symbol * represents the convolution operation, s(n) is the speech signal, and g i (n) represents the impulse response of the i-th Gabor filter.

[0082] γ i (n) x is obtained after N:1 downsampling. i (n), then x i (n) Energy estimation of the speech signal is obtained through FDEO. For discrete signal x i(n), using a three-point symmetric difference, is defined as:

[0083]

[0084] Then its frequency domain weighted energy operator discrete expression is:

[0085]

[0086] Where h i (n) is the signal x i (n) Hilbert transform, defined as h i (n)=H[x i (n)].

[0087] After estimating the energy of the speech signal, the instantaneous amplitude and instantaneous frequency of each sub-band signal are obtained using an energy separation algorithm, expressed as follows:

[0088]

[0089]

[0090] Among them, a i [n] and ω i [n] represents the instantaneous amplitude and instantaneous frequency of the i-th sub-band signal, respectively.

[0091] Feature enhancement is achieved through SENet. The principle is to calculate weighting factors derived from the input feature vectors and then control the scale of each filter channel, i.e., the output weight of each dimension. This enhances important features and weakens less important ones, thus making the features more targeted. The SENet architecture mainly consists of squeezing layers and activation layers, as shown in the diagram below. Figure 2 As shown.

[0092] The compression layer is designed to address channel dependency issues by compressing global spatial information into a single channel descriptor. This is achieved by extracting features from each frame of speech, resulting in the instantaneous amplitude feature vector a[n] = {a1[n], a2[n], ..., a...}. N [n]} and instantaneous frequency eigenvector ω[n]={ω1[n],ω2[n],……,ω N [n]}, thus for a speech signal with frame number τ, two feature vector sequences can be obtained, denoted as A and Ω. A and Ω are used as the inputs to the squeezing layer, collectively referred to as U={u1,u2,…,u c ,…,u N} T T represents the matrix transpose operator. That is, when the input is A, U = A, and when the input is Ω, U = Ω; while c is the frequency exponent and c∈[1,N], u c={u c (1),u c (2),……,u c (τ)}, and then the squeezed descriptor z = {z1, z2, ..., z} is obtained by global average pooling. c ,…,z N}, where the squeeze descriptor z c The expression is:

[0093]

[0094] After obtaining the squeeze descriptor, the weights of each feature dimension are obtained through matrix operations in the activation layer. The weight parameters are represented as s = {s1, s2, ..., s}. c ,…,s N}, its expression is:

[0095] s=σ(W2*δ(W1*z)) (9)

[0096] Where δ(·) represents the Rectified Linear Unit (ReLU) function, W1 and W2 represent the dimension-reduced matrix of dimension a×N and the dimension-increasing matrix of dimension N×a, respectively, which are mainly used to reduce parameter calculations, and σ(·) represents the Sigmoid function.

[0097] Finally, the obtained weight parameter vector is multiplied by the original input vector sequence channels to obtain the weighted output x. c The expression is:

[0098] x c =F scale (u c ,s c (10)

[0099] Among them, F scale (·,·) denotes channel multiplication, then the hidden layer representation of the speech is X={x1,x2,…,x c ,…,x N} T .

[0100] During the training of the feature enhancement module, the hidden representation X of the speech is fed into a linear layer as a training classifier. The output of this linear layer is the probability of real versus fake speech, which is used to calculate the cross-entropy loss and update the parameters within the module. During testing, this linear layer is removed, and X is directly divided into feature vectors for each frame along the time dimension. The average of each frame is then calculated to obtain the vector-level representation of the speech, namely the enhanced instantaneous amplitude feature vector and instantaneous frequency feature vector, which are represented as a′={a′1,a′2,…,a′ i ,…,a′ N} and ω′={ω′1,ω′2,…,ω′ i ,…,ω′ N}

[0101] The low-dimensional representations of IFCC and IACC are obtained using the Discrete Cosine Transform (DCT). The DCT expression is as follows:

[0102]

[0103]

[0104] Where f(i) is the original signal, i.e., a′ as described above. i and ω′ i F(u) are the coefficients after the DCT transformation. This invention uses the first 13 coefficients of F(u), i.e., F(u), u = 1, 2, ..., 13. N is the dimension of the eigenvectors. ρ(u) is a compensation coefficient that makes the DCT transformation matrix orthogonal.

[0105] The obtained 13-dimensional IACC and IFCC are subjected to first-order and second-order difference operations respectively, and then concatenated to obtain two 39-dimensional IACC+△+△△ and IFCC+△+△△ vectors.

[0106] Gaussian Mixture Model Classifier: After extracting the IACC+△+△△ and IFCC+△+△△ feature vectors from the speech signals in the training set, the parameters of their respective classifier models are trained using these feature vectors. One class represents real speech, and the other represents replayed speech. The Gaussian Mixture Model is a widely used generative model in speech processing, which represents the feature parameter space of each speech class as a weighted sum of M multivariate Gaussian functions.

[0107] For the eigenvector x∈R D×1 The likelihood function expression for its mixing density is:

[0108]

[0109] Where, ω i Let be the weight of the i-th Gaussian component, and have Gaussian probability density distribution function p i (x) is represented by the D×1 dimensional mean vector μ i The variance matrix Σ of D×D i Parameterization, the expression is:

[0110]

[0111] The model parameters are estimated using an iterative mean-maximization algorithm, the expression of which is as follows:

[0112] θ (i+1) =arg max θ E[logp(x,γ|θ)|x,θ (i) (15)

[0113] Where, θ (i) This represents the value of θ obtained in the i-th iteration, where θ represents the parameter set. γ represents the response of different Gaussian functions to the input sample x.

[0114] The iterative process involves first assigning initial values ​​to the Gaussian distribution parameters θ, which include the weights, mean, and variance of the Gaussian distribution function. Then, the mean of the log-likelihood function is calculated at the 0th iteration, and the value of θ with the maximum mean is obtained. This updated value is then used as the value for the first iteration and fed into the next iteration. This process is repeated until convergence.

[0115] After obtaining the Gaussian mixture models for the real speech class and the replayed speech class, the IACC and IFCC feature vectors of the detected speech are fed into their respective classifiers. The log-likelihood ratio score of each frame is calculated and averaged over time. The final score is obtained by subtracting the score of the replayed speech class from the score of the real speech class. The two scores obtained from IACC and IFCC are then fused with equal weights and compared with a set score threshold. If the score exceeds the threshold, it is judged as real speech; otherwise, it is considered as replayed speech.

Claims

1. A method for detecting a replayed voice attack for a voiceprint security authentication system, characterized in that, It comprises the following steps: Step 1, pre-processing the training voice; Step 2, passing the pre-processed voice signal through a Gabor filter bank to obtain N sub-band signals; Step 3, using a frequency domain energy operator FDEO to perform AM-FM analysis on the sub-band signals to obtain energy estimates of the voice signal, and to obtain instantaneous amplitude IACC and instantaneous frequency IFCC; Step 4, passing the instantaneous amplitude and instantaneous frequency of the voice signal through a squeeze-and-excitation network SE-Net to obtain a feature vector after feature enhancement; Step 5, passing the feature vector of step 4 through averaging and DCT to obtain low-dimensional feature vectors IACC" and IFCC"; Step 6, performing first-order difference operation and second-order difference operation on the low-dimensional feature vector IACC" of step 5, and concatenating the low-dimensional feature vector IACC", the first-order difference operation result and the second-order difference operation result to obtain 39-dimensional IACC"+△IACC"+△△IACC"; performing first-order difference operation and second-order difference operation on the low-dimensional feature vector IFCC" of step 5, and concatenating the low-dimensional feature vector IFCC", the first-order difference operation result and the second-order difference operation result to obtain 39-dimensional IFCC"+△IFCC"+△△IFCC"; Step 7, using the feature vectors of the training set to construct Gaussian mixture model parameters, and inputting the voice to be tested into a classifier for discrimination.

2. The method for detecting a replay voice attack for a voiceprint security authentication system according to claim 1, wherein The step 2 comprises: Let the sampling frequency be F s Then the highest frequency of the sampling signal is F s / 2, divide it into 200Hz equal width, the number of filters is The symbol represents the ceiling, then the center frequency is F i = 200i - 100(Hz), i ∈ [1, N], so it needs a filter bank composed of N Gabor filters, the unit impulse response of the i-th filter is represented as: g i (n) = exp(-b 2 n 2 ) cos(Ω i n) ⑴ Ω i = 2πF i T ⑵ where Ω i is the center frequency, T is the sampling period, and b is used to control the bandwidth; the speech signal after the i-th Gabor sub-band filter can be expressed as gamma i (n) = s(n) * g i (n) ⑶ where the symbol * denotes convolution operation, s(n) is the speech signal, g i (n) is the impulse response of the i-th Gabor filter.

3. The method of claim 2, wherein the method further comprises: The step 3 comprises: The γ i (n) is obtained by N:1 down-sampling x i (n), and then x i (n) is obtained by FDEO to obtain the energy estimation of the speech signal. For a discrete signal x i (n), a three-point symmetric difference is used, which is defined as: Then the discrete expression of the frequency domain weighted energy operator is: where h i (n) is the signal x i (n) is the Hilbert transform, defined as h i (n) = H[x i (n)]; After completing the energy estimation of the voice signal, the instantaneous amplitude and the instantaneous frequency of each sub-band signal are obtained by using an energy separation algorithm, and the expression is: where a i [n] and ω i [n] are the instantaneous amplitude and instantaneous frequency of the i-th subband signal, respectively.

4. The method of claim 1, wherein the method further comprises: In the SE-Net architecture of step 4, it mainly consists of a squeeze layer and an excitation layer; By extracting the features of each frame of speech, the above-mentioned instantaneous amplitude feature vector a[n] = {a1[n], a2[n],..., a N [n]} and the instantaneous frequency feature vector ω[n] = {ω1[n], ω2[n],..., ω N [n]} are obtained. Thus, two feature vector sequences, denoted as A and Ω, can be obtained for a speech signal with a frame number of τ. A and Ω as inputs to the squeezing layer, collectively denoted as U = {u1, u2,..., u c ,…,u N} T , T denotes the matrix transpose operator; When the input is A, U = A, when the input is Ω, U = Ω; and c is a frequency index and c ∈ [1, N], u c = {u c (1), u c (2), …, u c (τ)} is obtained, and then global average pooling is performed to obtain the squeeze descriptor z = {z1, z2, …, z c ,…, z N}, wherein the squeeze descriptor z c The expression is: After the extrusion descriptor is obtained, the weight of each feature dimension is obtained through matrix operation at the excitation layer, and the weight parameter is expressed as s={s1, s2, …, s c ,…,s N}, and the expression is as follows: Wherein, δ(·) represents a rectified linear unit (Relu) function, W1 and W2 represent a dimension-reduction matrix with a dimension of a×N and a dimension-increase matrix with a dimension of N×a respectively, which are mainly used to reduce parameter calculation, σ(·) represents a Sigmoid function; The obtained weight parameter vector is multiplied with the original input vector sequence channel to obtain a weighted output x c The expression is: x c = F scale (u c , s c ) ⑽ where F scale (·,·) denotes channel multiplication, then the hidden layer representation of the speech is X = {x1,x2,...,x c ,...,x N} T ; The hidden layer representation X of the voice is connected to a linear layer as a classifier for training, and the output result of the linear layer is the probability of true or false voice, which is used to calculate the cross-entropy loss to update the parameters in the module; when testing, the linear layer is cancelled, and X is directly divided into feature vectors of each frame according to the time dimension, and each frame is averaged to obtain the vector-level representation of the voice, i.e. the enhanced instantaneous amplitude feature vector and the instantaneous frequency feature vector, which are represented as a' = {a1', a'2,..., a i ′,...,a′ N} and ω' = {ω1', ω2',..., ω i ′,...,ω′ N}.

5. The method for detecting a replay voice attack for a voiceprint security authentication system according to claim 4, wherein, In step 5: The DCT expression is: where f(i) is the original signal, i.e. a i and ω i , F(u) is the DCT transformed coefficient; where F(u) takes the first 13 coefficients, i.e. F(u), u = 1, 2,..., 13; N is the dimension of the eigenvector; and p(u) is a compensation coefficient, making the DCT transformation matrix an orthogonal matrix.

6. The method of claim 1, wherein the method further comprises: In step 7: For the feature vector x ∈ R D×1 The likelihood function expression of the mixture density is: where ω i is the weight of the i-th Gaussian component, and has The Gaussian probability density distribution function p i (x) is parameterized by a D x 1 dimensional mean vector μ i and a D x D dimensional covariance matrix Σ i and is expressed as: The estimation of the model parameters uses a mean-maximization algorithm to iteratively solve, and the expression is as follows: The estimation of the model parameters uses a mean-maximization algorithm to iteratively solve, and the expression is as follows: θ (i+1) = argmax θ E[log p(x,γ|θ)|x,θ (i) ] ⒂ where θ (i) denotes the value of θ obtained at the i-th iteration, θ denotes the parameter set γ represents the degree of response of different Gaussian functions to input samples x.

Images