CN116778910A - A voice detection method - Google Patents

Abstract

This application provides a speech detection method, which includes: acquiring a target speech, preprocessing the target speech, the preprocessing including pre-emphasis, framing and windowing; determining the first value of the preprocessed target speech. vocal tract features, first sound source wave features and a variety of first correlation features; determining the first principal component feature based on the first vocal tract features, first sound source wave features and a variety of first correlation features; The first principal component feature is input into a trained classifier and a classification result is output, and the classification result is a fake speech or a natural speech. This application utilizes the trace information left by forged speech at the fundamental frequency, and utilizes the differences in sound source and vocal tract characteristics between forged speech and natural speech to achieve forged speech detection. Use the principal component analysis method to screen the sound source and vocal tract features respectively, select principal components with higher correlation as features, reduce feature dimensions and redundant features, and improve the generalization ability and efficiency of the model.

Description

A voice detection method

Technical field

The present application relates to the field of speech detection, and in particular to a speech detection method.

Background technique

With the continuous advancement of technology, speech technology has been widely used, such as speech recognition, speech synthesis, etc. With the booming development of deep learning, artificial intelligence technology has been introduced to improve performance in many tasks in the speech field. However, the development process of voice technology has also introduced some challenges. In order to deal with the major threat of voice spoofing attacks, the development of forged voice detection systems for voice spoofing attacks has attracted much attention in recent years. Although many fake speech detection methods have been proposed, only a few have been implemented. The existing forged speech detection system is not designed for the characteristics of forged speech, resulting in low model generalization ability and difficulty in gaining people's trust.

The current counterfeiting voice detection system relies too much on traditional artificially designed features and the classification performance of deep neural networks. It is not designed for the characteristics of forged voice, resulting in a model that is closely related to the data set and has poor generalization in actual application scenarios. Difficult to deploy practically.

Contents of the invention

In the first aspect, embodiments of the present application provide a speech detection method. The method includes: acquiring the target speech, preprocessing the target speech, and the preprocessing includes pre-emphasis, framing and windowing; determining the preprocessed target speech. Multiple first vocal channel features; determine the first sound source wave feature of the preprocessed target speech; the first sound source wave feature is extracted through an inverse filter; based on multiple first vocal channel features and the first sound source wave The feature determines the first principal component feature; the first principal component feature is input into the trained classifier and the classification result is output. The classification result is fake speech or natural speech.

Therefore, the speech detection method proposed in the embodiment of the present application uses the trace information left by the forged speech at the fundamental frequency, and uses the differences in the sound source and vocal tract characteristics caused by the different generation processes between the fake speech and the natural speech to Implement fake speech detection. At the same time, the principal component analysis method is used to screen the sound source and vocal channel features respectively, and the principal components with higher correlation are selected as features to reduce feature dimensions and redundant features, and improve the generalization ability and efficiency of the model.

In some implementable implementations, the plurality of first vocal channel features include the amplitude-frequency characteristics of the vocal channel filter and the amplitude-frequency characteristics of the inverse filter, and the plurality of first vocal channel features of the preprocessed target speech are determined, It includes: using linear predictive coding to predict the filter parameters of each frame of the preprocessed target speech; calculating and determining the amplitude-frequency characteristics of the vocal channel filter and the amplitude-frequency characteristics of the inverse filter based on the filter parameters of each frame.

Therefore, embodiments of the present application introduce a variety of features related to the vocal tract to enhance the generalization ability of the classifier.

In some implementable implementations, the plurality of first channel features further include fundamental frequency features, fundamental frequency perturbation features, amplitude perturbation features, and Mel cepstrum coefficients.

Therefore, embodiments of the present application introduce a variety of features related to the vocal tract to enhance the generalization ability of the classifier.

In some implementations that can be implemented, determining the first sound source wave characteristics of the preprocessed target speech includes: obtaining the short-time Fourier characteristics of the framed target speech; using an inverse filter to perform the short-time Fourier The features are inversely filtered to obtain the first sound source wave characteristics of the preprocessed target speech.

Therefore, the embodiment of the present application uses the inverse filtering method to separate the sound source features and vocal tract features, so that the classifier can make judgments using specific features introduced due to the different generation mechanisms of fake speech and natural speech, making full use of information. At the same time, too many data preprocessing steps are not introduced.

In some implementable implementations, determining the first principal component feature based on a plurality of first vocal channel features and a first sound source wave feature includes: splicing a plurality of first vocal channel features and the first sound source wave feature to obtain a first One splicing feature; the number of features of the first splicing feature is n; decentralize each feature in the first splicing feature; calculate the covariance matrix of the decentralized first splicing feature; perform eigenvalue decomposition on the covariance matrix , obtain multiple eigenvalues and corresponding eigenvectors; according to the order of multiple eigenvalues from large to small, select the first k eigenvectors to form a transformation matrix, where k<n; multiply the first splicing feature by the transformation matrix to obtain the One principal component feature.

Therefore, the embodiment of the present application introduces principal component analysis in the feature splicing process to achieve dimensionality reduction and select highly relevant features that contribute more to the task. Reduce redundancy to improve efficiency.

In some implementations that can be implemented, the method also includes the step of training a classifier: obtaining labeled training speech in the training set, preprocessing the training speech, and the preprocessing includes pre-emphasis, framing, and windowing; determining that the preprocessed Multiple second vocal channel features of the training speech; determine the second sound source wave characteristics of the preprocessed training speech; the second sound source wave feature is extracted through the vocal tract inverse filter; based on multiple second vocal channels The second principal component feature and the second sound source wave feature determine the second principal component feature; input the second principal component feature into the classifier, perform iterative training, and obtain the trained classifier when the loss function converges.

Therefore, the classifier trained in the embodiment of the present application has a simple structure and good portability. Other beneficial effects are as mentioned above and will not be repeated here.

In some implementations that can be implemented, the plurality of second channel features include the amplitude-frequency characteristics of the channel filter and the amplitude-frequency characteristics of the inverse filter, and the plurality of second channel features of the preprocessed training speech are determined, Including: using linear predictive coding to predict the filter parameters of each frame of the preprocessed training speech; calculating and determining the amplitude-frequency characteristics of the vocal channel filter and the amplitude-frequency of the inverse filter based on the filter parameters of each frame of the training speech. characteristic.

In some implementable implementations, the plurality of second channel features further include fundamental frequency features, fundamental frequency perturbation features, amplitude perturbation features, and Mel cepstrum coefficients.

In some implementations that can be implemented, determining the second sound source wave characteristics of the preprocessed training speech includes: obtaining the short-time Fourier characteristics of the framed training speech; using an inverse filter to perform the short-time Fourier The features are inversely filtered to obtain the second sound source wave features of the preprocessed training speech.

In some implementable implementations, determining the second principal component feature based on a plurality of second vocal channel features and a second sound source wave feature includes: splicing a plurality of second vocal channel features and the second sound source wave feature to obtain a third Two spliced features; the number of features of the second spliced feature is n; decentralize each feature in the second spliced feature; calculate the covariance matrix of the decentralized second spliced feature; perform eigenvalue decomposition of the covariance matrix , obtain n eigenvalues and corresponding eigenvectors; according to the order of multiple eigenvalues from large to small, select the first k eigenvectors to form a transformation matrix, where k<n; multiply the second splicing feature by the transformation matrix to obtain the Two principal component features.

In a second aspect, embodiments of the present application provide an electronic device, including: at least one memory for storing a program;

At least one processor is configured to execute the program stored in the memory. When the program stored in the memory is executed, the processor is configured to execute the method described in any one of the first aspects.

In a third aspect, embodiments of the present application provide a computer storage medium. Instructions are stored in the computer storage medium. When the instructions are run on a computer, they cause the computer to execute the method provided in the first aspect.

Description of drawings

In order to more clearly illustrate the technical solutions of the multiple embodiments disclosed in this specification, the drawings needed to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only those disclosed in this specification. For multiple embodiments, those of ordinary skill in the art can also obtain other drawings based on these drawings without exerting creative efforts.

The drawings needed to be used in the description of the embodiments or the prior art are briefly introduced below.

Figure 1 is a system architecture diagram of a speech detection method provided by an embodiment of the present application;

Figure 2 is a flow chart of a speech detection method provided by an embodiment of the present application;

Figure 3 is a training flow chart of a speech detection method provided by an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings.

In the description of the embodiments of this application, words such as "exemplary", "for example" or "for example" are used to represent examples, illustrations or explanations. Any embodiment or design described as "exemplary," "such as," or "for example" in the embodiments of the present application is not to be construed as preferred or advantageous over other embodiments or designs. Rather, use of the words "exemplary," "such as," or "for example" is intended to present the concepts in a concrete manner.

In the description of the embodiments of this application, the term "and/or" is only an association relationship describing associated objects, indicating that there can be three relationships. For example, A and/or B can mean: A alone exists, and A alone exists. There is B, and there are three situations A and B at the same time. In addition, unless otherwise stated, the term "plurality" means two or more. For example, multiple systems refer to two or more systems, and multiple terminals refer to two or more terminals.

In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Therefore, features defined as "first" and "second" may explicitly or implicitly include one or more of these features. The terms “including,” “includes,” “having,” and variations thereof all mean “including but not limited to,” unless otherwise specifically emphasized.

In the description of the embodiments of this application, reference is made to "some embodiments", which describes a subset of all possible embodiments, but it can be understood that "some embodiments" may be the same subset or different subsets of all possible embodiments. sets and can be combined with each other without conflict.

In the description of the embodiments of this application, the terms "first\second\third, etc." or module A, module B, module C, etc. are only used to distinguish similar objects and do not represent a specific ordering of the objects. , it is understood that the specific order or sequence may be interchanged where permitted, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein.

In the description of the embodiments of this application, the labels indicating steps involved, such as S110, S120..., etc., do not necessarily mean that this step will be executed. If allowed, the order of the previous and subsequent steps can be interchanged, or they can be performed simultaneously. implement.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

Figure 1 is a system architecture diagram of a speech detection method provided by an embodiment of the present application. As shown in Figure 1, the speech preprocessing module 11 first preprocesses the target speech, which can be the acquired real or fake speech; the preprocessing includes pre-emphasis, framing and windowing. The vocal tract feature extraction module 121 uses linear prediction coding (LPC) to extract the coefficients of the vocal tract filter of each frame of the speech signal in the target speech, and estimates the amplitude-frequency characteristics of the vocal channel filter of each frame of the speech signal based on the coefficients, and its Inverse the amplitude-frequency characteristics of the filter to obtain the vocal channel characteristics. The sound source wave feature extraction module 122 calculates the short-time Fourier transform (STFT) spectrum for each frame of the preprocessed speech signal, and uses the vocal tract inverse filter to inversely filter the short-time Fourier transform (STFT) spectrum. , obtain the sound source wave characteristics. The sound source channel related feature extraction module 123 obtains related features such as fundamental frequency features, fundamental frequency perturbation features, Mel Cepstral Coefficients (MFCC), etc. for each frame of preprocessed speech signal. The principal component determination module 13 splices related features such as vocal tract features, sound source wave features, fundamental frequency features, fundamental frequency perturbation features, Mel Cepstral Coefficients (MFCC), etc., and uses principal component analysis (PCA) to remove redundancy Features to obtain principal component features with high correlation. The classification module 14 inputs the principal component features into the classifier for two classifications, and outputs the result of natural speech or fake speech. Embodiments of the present application provide a speech detection method that makes full use of the differences in sound sources and vocal channels to achieve generalized forged speech recognition in actual usage scenarios for speech whose authenticity is difficult to distinguish; using the main The feature selection ability of component analysis reduces redundancy and improves efficiency.

Figure 2 is a schematic diagram of a speech detection method provided by an embodiment of the present application. As shown in Figure 2, the speech detection method includes: S11, obtain the target speech, and preprocess the target speech. The preprocessing includes pre-emphasis, framing and windowing; S12, determine the first channel of the preprocessed target speech. Features, first sound source wave features and multiple first correlation features; S14, determine the first principal component features based on the first vocal channel features, first sound source wave features and multiple first correlation features; S15, determine the first principal component features The principal component features are input to the trained classifier, and the classification result is output. The classification result is fake speech or natural speech.

Each step of the voice detection method provided by this application will be introduced in detail below with reference to the embodiments.

S11, obtain the target speech and preprocess the target speech. The preprocessing includes pre-emphasis, framing and windowing.

Since high-frequency parts of speech usually have smaller amplitudes compared to lower frequency parts, pre-emphasis balances the spectrum, avoids numerical problems during Fourier transform operations, improves signal-to-noise ratio (SNR), and eliminates the voicing process. The effect of the vocal cords and lips compensates for the high-frequency part of the speech signal that is suppressed by the articulatory system, highlighting the high-frequency formants.

In this embodiment of the present application, a high-pass filter can be used to pre-emphasize the target speech. The pre-emphasized speech signal y(n) is:

y(n)＝x(n)-0.79·x(n-1) (1)

Where x(n) represents the n-th frame speech signal in the target speech, and x(n-1) represents the n-1th frame speech signal in the target speech.

In most cases, the speech signal is non-stationary and it does not make sense to Fourier transform the entire signal. Therefore, Fourier transform can be performed on short-time frames, and a good approximation of the time-frequency transformation of the signal can be obtained by connecting adjacent frames.

In this embodiment of the present application, after pre-emphasis, the target speech is divided into multiple short-term frames, and each frame is a short-term smooth speech signal.

For example, the frame length can be set to 25ms, the frame movement can be set to 10ms, and the original speech is divided into multiple speech frames to calculate short-term time-frequency characteristics.

After segmenting the target speech into multiple short-term frames, each frame is multiplied by a window function to increase the continuity between the left and right ends of the frame and reduce spectrum leakage. The window function can be a Hamming window (Hamming), and the window function formula w(n) is:

Among them, N represents the number of sampling points in the window.

S12. Determine the first vocal channel characteristics, first sound source wave characteristics and various first related characteristics of the preprocessed target speech.

The following introduces the first vocal channel characteristics, first sound source wave characteristics and various first related characteristics of the target speech respectively.

S121. Determine the first vocal channel characteristics of the preprocessed target speech.

The first vocal channel features include the amplitude-frequency characteristics of the vocal channel filter and the amplitude-frequency characteristics of the inverse filter. Linear predictive coding (LPC) can be used to predict the vocal channel filter parameters of each frame of the preprocessed target speech; according to the required The amplitude-frequency characteristics of the vocal channel filter of each frame of the target speech and the amplitude-frequency characteristics of the inverse filter can be obtained by calculating the filter parameters of each frame.

In the embodiment of the present application, the zero points and poles of the vocal channel filter under z-transformation of each frame of speech can be calculated based on linear predictive coding (LPC), and the zero points and poles of the vocal channel filter are determined according to the zero points and poles of the vocal channel filter. Amplitude-frequency characteristics. Further, for each frame of the target speech, the zero points and poles of the vocal tract filter are exchanged to obtain the inverse vocal tract filter. The amplitude-frequency characteristics of the inverse filter calculated from the swapped zeros and poles.

S122, determine the first sound source wave characteristic of the preprocessed target speech; the first sound source wave characteristic is the amplitude-frequency characteristic of the inverse filter multiplied by the original speech in the frequency domain, and then the vocal tract characteristics are filtered out to obtain the sound source wave characteristic. .

In the embodiment of the present application, the short-time Fourier features of each frame of the target speech are obtained; the short-time Fourier features are inversely filtered through an inverse filter to obtain the first sound source wave features.

S123. Determine multiple first correlation features of the preprocessed target speech, including Mel cepstrum coefficients, fundamental frequency features, fundamental frequency perturbation features and amplitude perturbation features.

Among them, the Mel cepstrum coefficient uses a Mel filter bank to filter the short-time Fourier spectrum of each frame of the target speech through the Mel filter to obtain the Mel spectrum feature (fbank).

Specifically, each frame of the target speech is windowed through short-time Fourier transform and time-frequency analysis is performed, and then the logarithmic amplitude spectrum is taken to obtain the spectrogram. The short-time Fourier transform is defined as:

Among them, x(τ) is a single frame speech signal, h(τ-t) is the analysis window function, and τ is the offset.

The spectrogram is filtered through an 80-dimensional Mel filter bank to obtain the Mel spectrum features.

The Mel filter bank imitates the human auditory perception system, and its sensitivity to signals of different frequencies is different. The Mel filter bank is not uniformly distributed on the frequency axis. In the low-frequency region, the Mel filters are densely distributed, but in the high-frequency region, the Mel filters are sparsely distributed. Therefore, the Mel filter bank can simulate the human ear's nonlinear perception of sound, which is more discriminative at lower frequencies and less resolution at higher frequencies. The conversion formula between Mel frequency f _mel and linear frequency f is:

According to this formula, the Mel filter bank can be calculated. Each filter in the filter bank is multiplied and accumulated with the short-time Fourier spectrum (STFT) in turn to obtain the Mel filter bank characteristics. Perform cepstrum analysis on the calculated Mel filter bank features to obtain the final Mel Cepstrum (MFCC) features.

The function of cepstral analysis is to extract the envelope of the frequency spectrum, which in the time-frequency characteristics of speech is to extract the formant characteristics of speech, that is, the vocal tract characteristics. The discrete cosine transform (DCT) can be used to perform cepstral analysis on the fbank feature to obtain the Mel cepstral coefficient MFCC _i :

where S _j is the output value of the jth filter in fbank, MFCC _i is the Mel cepstrum coefficient of the i th Mel filter, and N is the number of Mel filter banks.

The fundamental frequency feature F ₀ is a feature obtained by calculating the pitch period of each frame signal.

The fundamental frequency perturbation feature is a characteristic that represents the non-periodic changes of the fundamental frequency perturbation. The fundamental frequency perturbation is a measure of the change in the fundamental frequency between adjacent cycles. The fundamental frequency perturbation is expressed in percentage (%). The formula of the fundamental frequency perturbation Jitter is:

Among them, F ₀ is the fundamental frequency of each frame of speech, and J is the number of gene cycles of each frame of speech.

The amplitude perturbation feature is a feature that represents the non-periodic changes of the amplitude perturbation. Amplitude perturbation is a measure of the change in amplitude between adjacent periods of the speech signal, and the amplitude perturbation is measured in decibels (dB).

The formula of amplitude perturbation Shimmer is:

Where A is the peak-to-peak amplitude of each pitch period, and L is the number of amplitudes.

The above-mentioned first vocal channel features, first sound source features and various first related features all represent the sound source and vocal tract information of the speech, and these feature information indicate the difference in the generation process of the fake speech and the real speech.

In the embodiment of the present application, a variety of relevant features are obtained during the speech feature extraction process. However, too many features contain redundant information, which has an impact on the generalization and calculation speed of the model. Therefore, features need to be screened to obtain features with higher relevance to the speech forgery detection task.

S13: Determine the first principal component feature based on the first vocal channel feature, the first sound source wave feature and a variety of first correlation features.

The embodiment of this application uses principal component analysis (PCA) to screen multiple first vocal channel features and first sound source wave features, including the following steps:

S131. The first splicing feature is obtained by splicing the first vocal channel feature, the first sound source wave feature and a variety of first related features; the number of features of the splicing feature is n.

In the embodiment of this application, the amplitude-frequency characteristics of the vocal tract filter, the amplitude-frequency characteristics of the inverse filter, the fundamental frequency characteristics, the fundamental frequency perturbation characteristics, the amplitude perturbation characteristics, and the Mel cepstral spectrum of each frame of the target speech are The number is spliced with n features such as the first sound source wave feature to obtain the first splicing feature.

S132, decentralize each feature in the spliced features.

In the embodiment of the present application, decentralization includes subtracting the mean value of the dimension from each feature in the first concatenated feature.

S133. Calculate the covariance matrix based on the decentralized first splicing feature.

S134. Perform eigenvalue decomposition on the covariance matrix to obtain multiple first eigenvalues and corresponding first eigenvectors.

S135, select the first k eigenvectors to form the first transformation matrix according to the order of multiple eigenvalues from large to small, where k is the number of features after dimensionality reduction, k<n.

S136: Multiply the first splicing feature by the first transformation matrix to obtain the first principal component feature.

Among them, the first principal component feature is a high-correlation feature after dimensionality reduction.

In the embodiment of the present application, the process of principal component feature analysis can project high-dimensional features onto low-dimensional features and retain features with high relevance to the task to the greatest extent possible.

S14. Input the first principal component feature into the trained classifier and output the classification result. The classification result is fake speech or natural speech.

The speech detection method provided by the embodiment of the present application also includes the step of training a classifier.

Figure 3 is a flow chart of training a classifier in the speech detection method provided by the embodiment of the present application. As shown in Figure 3, the training process of the classifier includes: S31, obtain the labeled training speech of the training set, and preprocess the training speech. The preprocessing includes pre-emphasis, framing and windowing. S32. Determine the second vocal channel feature, the second sound source wave feature and a plurality of second related features of the preprocessed training speech. S34: Determine the second principal component feature based on the second vocal channel feature, the second sound source wave feature and the multiple second related features of the training speech. S35, input the second principal component feature into the classifier, perform iterative training, and obtain the trained classifier when the loss function converges.

The various steps of training the classifier are introduced in detail below.

S31, obtain the labeled training speech of the training set, and preprocess the training speech. The preprocessing includes pre-emphasis, framing and windowing.

For the implementation of pre-emphasis, framing and windowing, reference can be made to the implementation of step S11 and will not be described again here.

S32. Determine the second vocal channel feature, the second sound source wave feature and multiple second related features of the preprocessed training speech.

The following introduces the second vocal channel features, second sound source wave features and multiple second related features of the training speech respectively.

S321. Determine the second channel characteristics of the preprocessed training speech.

The second channel features include the amplitude-frequency characteristics of the vocal channel filter of the training speech and the amplitude-frequency characteristics of the inverse filter. The linear predictive coding (LPC) method can be used to predict the filter parameters of each frame of the preprocessed training speech. ; Calculate the amplitude-frequency characteristics of the vocal tract filter for each frame of the training speech based on the filter parameters of each frame of the training speech, and the amplitude-frequency characteristics of the inverse filter.

The relationship and calculation method between the amplitude-frequency characteristics of the inverse filter and the amplitude-frequency characteristics of the filter for each frame of the training speech can be referred to step S121, which will not be described again here.

S322, determine the second sound source wave feature of the preprocessed training speech; the second sound source wave feature is the amplitude-frequency characteristic of the inverse filter of the training speech and the original speech multiplied in the frequency domain and filtered out the vocal tract features to obtain the sound Source wave characteristics.

S323. Determine multiple second correlation features of the preprocessed training speech, including Mel cepstrum coefficients, fundamental frequency features, fundamental frequency perturbation features and amplitude perturbation features.

For various related features and calculation methods of each frame of the training speech, please refer to step S123, which will not be described again here.

The above-mentioned second vocal channel features, second sound source features and various related features of the training voice all represent the sound source and vocal tract information of the training voice. These features indicate the difference in the generation process of the fake voice and the real voice.

Similarly, in the embodiments of this application, many types of sound source and vocal tract related features were obtained during the feature extraction process of training speech. Too many features contain redundant information, which has an impact on the generalization of the model and the calculation speed. It is necessary to These features are screened to obtain features with higher relevance to the speech forgery detection task.

S33: Determine the second principal component feature based on the second vocal channel feature, the second sound source feature and multiple related features.

The embodiment of this application uses principal component analysis (PCA) to screen the second channel features, second sound source features and multiple related features, including the following steps:

S331, splicing the second channel feature, the second sound source feature and multiple related features to obtain the second splicing feature; the number of features of the second splicing feature is n.

In the embodiment of this application, the amplitude-frequency characteristics of the filter of each frame of the training speech, the amplitude-frequency characteristics of the inverse filter, the fundamental frequency characteristics, the fundamental frequency perturbation characteristics, the amplitude perturbation characteristics, the Mel cepstrum coefficients and The second splicing feature is obtained by splicing n features such as the second sound source feature.

S332, decentralize each feature in the second spliced feature.

In the embodiment of the present application, decentralization includes subtracting the mean value of the dimension from each feature in the second concatenated feature.

S333. Calculate the covariance matrix based on the decentralized second splicing feature.

S334: Perform eigenvalue decomposition on the covariance matrix to obtain multiple second eigenvalues and corresponding second eigenvectors.

S335: Select the first k second eigenvectors to form a second transformation matrix in descending order of multiple second eigenvalues, where k is the number of dimensions after dimensionality reduction, k<n.

S336: Multiply the second splicing feature by the second transformation matrix to obtain the second principal component feature. Among them, the second principal component feature is a high-correlation feature after dimensionality reduction.

S34, input the second principal component feature into the classifier, perform iterative training, and obtain the trained classifier when the loss function converges.

In the embodiment of this application, the classifier used is a hidden layer feature extractor, and the hidden layer feature extractor has a residual network (ResNet) with a squeeze excitation module (SE-block).

The residual network is a variant of the convolutional neural network (CNN). It solves the degradation problem caused by network deepening by adding residual connections between convolutional layers. Each residual module usually contains a multi-layer structure, and these layers can be the constituent layers of any neural network, which makes the residual network more scalable. After the input x of the residual module undergoes forward calculation F(x), the original input without forward calculation will be added to the output, forming a short-circuit connection. This calculation method makes even if the forward calculation is 0, the residual module is only equivalent to doing identity mapping, ensuring that the network performance will not be degraded. The residual module allows the convolutional neural network to avoid degradation problems and deepen the network layers, making the deep neural network have better performance.

The squeeze excitation module is an extension module of the convolutional neural network, which can explicitly model the interdependence between feature channels and improve system performance from the channel perspective. The squeeze excitation module first squeezes the input feature map, squeezes the spatial dimension through the global average pooling layer, and turns each two-dimensional feature channel into a real number with a global receptive field. The output one-dimensional channel features are then excited, and weights are generated for each feature channel through parameters to explicitly model the correlation between feature channels. Finally, the calculated weights are reweighted into the initial two-bit channel features to complete the recalibration of the original features in the channel dimension.

The hidden layer features extracted by the hidden layer feature extractor are then classified into two categories through the linear layer to obtain the final forged speech detection result.

In the embodiment of this application, the Adam optimizer and the additive angularmargin softmax loss (AAM-Softmax) are used to train the classifier. The calculation formula of the AAM-Softmax loss function is:

In the formula, m is the boundary value, and m=0.2, s=40 is the scale parameter. f _i is the embedded feature of the input, is the weight matrix, _yi represents the label information carried by the speech, c is the number of categories, here is 2, and N is the number of speech items in the training speech. The optimizer parameters are: β ₁ =0.9, β ₂ =0.999, ∈ =10 ^-8 , and the weight attenuation is 10 ^-4 .

The speech detection method provided by the embodiment of this application is applied to electronic devices that deploy and configure the PyTorch running environment.

The speech detection method provided by the embodiment of the present application uses the inverse filtering method to separate the sound source characteristics and vocal tract characteristics, and utilizes the different generation mechanisms of fake speech and natural speech, so that the classifier can introduce information of specific features and make full use of the specific features. analysis without introducing too many data preprocessing steps, consuming less resources and saving training time.

The speech detection method provided by the embodiment of the present application introduces a variety of sound source and vocal channel-related features to enhance the generalization ability of the model.

The speech detection method provided by the embodiment of the present application introduces principal component analysis in the feature splicing process to achieve dimensionality reduction and select high-correlation features that contribute more to the task. Reduce redundancy to improve efficiency.

The speech detection method provided by the embodiments of the present application has a simple structure and good portability, and the classifier used can be other two-class convolutional neural networks.

An embodiment of the present application provides an electronic device, including: at least one memory for storing a program; at least one processor for executing the program stored in the memory. When the program stored in the memory is executed, the The processor is used to execute the voice detection method.

Embodiments of the present application provide a computer storage medium. Instructions are stored in the computer storage medium. When the instructions are run on a computer, they cause the computer to execute the speech detection method.

It can be understood that the processor in the embodiments of the present application may be a central processing unit (CPU), or other general-purpose processor, digital signal processor (DSP), application specific integrated circuit (Application Specific Integrated Circuit). specific integrated circuit (ASIC), field programmable gate array (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

The method steps in the embodiments of the present application can be implemented by hardware or by a processor executing software instructions. Software instructions can be composed of corresponding software modules. The software modules can be stored in random access memory (random access memory, RAM), flash memory, read-only memory (read-only memory, ROM), programmable read-only memory (programmable rom). , PROM), erasable programmable read-only memory (erasable PROM, EPROM), electrically erasable programmable read-only memory (electrically EPROM, EEPROM), register, hard disk, mobile hard disk, CD-ROM or other well-known in the art any other form of storage media. An exemplary storage medium is coupled to the processor such that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and storage media may be located in an ASIC.

In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable device. The computer instructions may be stored in or transmitted over a computer-readable storage medium. The computer instructions may be transmitted from one website, computer, server or data center to another website through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) means. , computer, server or data center for transmission. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that contains one or more available media integrated. The available media may be magnetic media (eg, floppy disk, hard disk, magnetic tape), optical media (eg, DVD), or semiconductor media (eg, solid state disk (SSD)), etc.

It can be understood that the various numerical numbers involved in the embodiments of the present application are only for convenience of description and are not used to limit the scope of the embodiments of the present application.

Claims (10)

1. A method of voice detection, the method comprising:

obtaining target voice, and preprocessing the target voice, wherein the preprocessing comprises pre-emphasis, framing and windowing;

determining a first channel characteristic, a first sound source wave characteristic and a plurality of first related characteristics of the preprocessed target voice;

determining the first principal component feature based on the first channel feature, a first source wave feature, and a plurality of first correlation features;

inputting the first principal component characteristics into a trained classifier, and outputting a classification result, wherein the classification result is fake voice or natural voice.

2. The method of claim 1, wherein the first channel characteristics include an amplitude-frequency characteristic of a channel filter and an amplitude-frequency characteristic of an inverse filter, and wherein determining the first channel characteristics of the preprocessed target speech comprises:

predicting the filter parameters of each frame of the target voice after pretreatment by using linear prediction coding;

and calculating and determining the amplitude-frequency characteristic of the sound channel filter and the amplitude-frequency characteristic of the inverse filter according to the filter parameters of each frame.

3. The method of claim 1, wherein the plurality of first correlation features includes a fundamental frequency feature, a fundamental frequency perturbation feature, an amplitude perturbation feature, and a mel-frequency cepstral coefficient.

4. The method of claim 1, wherein said determining the first channel characteristic, the first source wave characteristic, and the plurality of first correlation characteristics of the pre-processed target speech comprises:

acquiring short-time Fourier features of the target voice after framing;

and carrying out inverse filtering on the short-time Fourier features through an inverse filter to obtain first sound source wave features of the preprocessed target voice.

5. The method of claim 1, wherein the determining the first principal component feature based on the first channel feature, the first source wave feature, and the plurality of first correlation features comprises:

splicing the first sound channel characteristic, the first sound source wave characteristic and a plurality of first related characteristics to obtain a first splicing characteristic; the feature quantity of the first splicing features is n;

decentralizing each of the first stitching features; calculating a covariance matrix of the first decentralised stitching feature;

performing eigenvalue decomposition on the covariance matrix to obtain a plurality of eigenvalues and corresponding eigenvectors;

selecting the first k eigenvectors to form a conversion matrix according to the sequence of the eigenvalues from large to small, wherein k is smaller than n;

multiplying the first splicing characteristic by the conversion matrix to obtain the first principal component characteristic.

6. The method according to any one of claims 1-5, further comprising the step of training a classifier:

acquiring training voice of a training set with a label, and preprocessing the training voice, wherein the preprocessing comprises pre-emphasis, framing and windowing;

determining a second sound characteristic, a second sound source characteristic and a plurality of second related characteristics of the preprocessed training speech;

determining a second principal component feature based on a second sound feature, a second sound source feature, and a plurality of second correlation features of the training speech;

and inputting the second principal component characteristics into a classifier, performing iterative training, and obtaining the trained classifier under the condition of convergence of a loss function.

7. The method of claim 6, wherein the plurality of second channel characteristics includes an amplitude-frequency characteristic of a channel filter and an amplitude-frequency characteristic of an inverse filter, and wherein determining the second channel characteristics of the pre-processed training speech comprises:

predicting filter parameters for each frame of the pre-processed training speech using linear predictive coding;

and calculating and determining the amplitude-frequency characteristic of the second channel filter and the amplitude-frequency characteristic of the inverse filter according to the filter parameters of each frame of the training voice.

8. The method of claim 6, wherein the plurality of second correlation features further comprises a fundamental frequency feature, a fundamental frequency perturbation feature, an amplitude perturbation feature, and a mel-frequency cepstral coefficient for each frame of the training speech.

9. The method of claim 6, wherein the determining the second sound characteristic, the second sound source characteristic, and the plurality of second correlation characteristics of the pre-processed training speech comprises:

acquiring short-time Fourier features of the training voice after framing;

and carrying out inverse filtering on the short-time Fourier features through the inverse filter to obtain second sound source wave features of the preprocessed training voice.

10. The method of claim 6, wherein the determining the second principal component feature based on the second sound feature, the second sound source feature, and the plurality of second correlation features of the training speech comprises:

splicing the second channel characteristics, the second sound source wave characteristics and a plurality of second correlation characteristics of the training voice to obtain second splicing characteristics; the feature quantity of the second splicing features is n;

decentralizing each of the second stitching features; calculating a covariance matrix of the second decentralised stitching feature;

performing eigenvalue decomposition on the covariance matrix to obtain n eigenvalues and corresponding eigenvectors;

selecting the first k eigenvectors to form a conversion matrix according to the sequence of the eigenvalues from large to small, wherein k is smaller than n;

multiplying the second stitching feature by the transformation matrix to obtain the second principal component feature.