Voice deception attack detection system and method based on microphone array

Abstract

The present invention belongs to the technical field of voice command liveness detection, and discloses a method for performing passive voice command liveness detection based on a circular microphone array of a smart speaker in a smart home environment to resist the threat of voice replay attacks from the device. In the process of liveness detection, through fine-grained analysis of the audio frequency domains of different channels and extraction of multi-channel features, the present invention can efficiently determine whether the voice command is generated by a real user or forged by an electronic device. The present invention can achieve fast and flexible defense against voice deception attacks, highly guarantee the security of voice interfaces in smart homes, and meet the needs of the industry. Relying only on the voice signal collected by the microphone of the smart speaker, it is possible to identify whether the voice command is generated by a real user or by a deceptive device.

Description

Technical Field

[0001] This invention belongs to the technical field of liveness detection for voice interfaces, specifically a microphone array-based voice spoofing attack detection system and method. By analyzing multi-channel voice signals received by smart speakers, it can determine whether the signals are generated by a real user or an electronic device, thus defending against spoofing attacks and protecting the security of voice interfaces. Background Art

[0002] Smart speakers equipped with intelligent voice interfaces (AIs) serve as the hubs of mainstream smart home platforms (such as Amazon Alexa and Google Home), making life more convenient for users. For example, users can remotely control home appliances or query information as long as they are within the sound reception range of the smart speaker. However, due to the broadcast nature of voice signals, attackers can perform a variety of voice spoofing attacks, such as replay attacks, adversarial sample attacks, and "dolphin sound attacks."

[0003] Since real voice commands are generated by the user's mouth movements, while voice spoofing commands are played by electronic speaker devices, the different characteristics between humans and electronic devices can be used to perform voice liveness detection. Voice liveness detection is divided into two types: liveness detection based on two-factor authentication and passive liveness detection that relies only on voice signals. Currently, research on liveness detection for two-factor authentication has been carried out. However, in order to capture the liveness characteristics of real users, many two-factor authentication systems require users to carry special sensors (such as accelerometers and magnetometers). Although two-factor authentication systems based on wireless sensing do not require users to carry devices, they require the deployment of IoT devices in smart home environments to collect channel state information of wireless signals, which brings additional workload to users.

[0004] The passive detection method studied in this paper only considers audio data collected by the voice interface. Its main principle is that there are significant differences between the pronunciation of real people and electronic speakers. Real people's pronunciation involves vocal cord vibration and mouth movement, while electronic speakers rely solely on the vibration of the diaphragm. This difference in pronunciation leads to subtle but significant differences in the collected audio spectrum. Currently, passive detection solutions based on mono and dual-channel audio have been proposed, but they still have numerous defects and shortcomings, making it difficult to meet the needs of liveness detection in smart home scenarios. Some passive verification solutions use the sub-bass low-frequency region of the mono audio spectrum or the human voice region as liveness detection factors. However, the characteristics of these regions are susceptible to changes in the sound propagation channel. Some liveness detection solutions use a unique feature in dual-channel audio—"fieldprint"—as a liveness detection feature. However, this solution requires the user to maintain a fixed posture to ensure the robustness of the fieldprint. Therefore, this solution is difficult to deploy in many real-world scenarios (for example, when the user walks or gestures change). A new passive liveness detection solution with the following advantages is needed:

[0005] (1) No need to rely on devices: Passive detection is performed only based on the audio collected by the smart speaker.

[0006] (2) Resilient to environmental changes: Robust to changes in sound propagation paths and user movements.

[0007] (3) High precision: Compared with existing solutions, the precision is higher.

[0008] Therefore, designing a passive liveness detection method based on microphone array has great potential.

[0009] Patent publication number CN111639519A discloses a liveness detection method and device that uses a microphone array to identify and detect liveness within a test box. This method is based on biometrics and pattern matching. This invention focuses on the source of audio, distinguishing between machine-generated and real-person audio. This method fundamentally differs from the patent in both its purpose and detection method. Simply detecting liveness within the test box cannot protect against voice spoofing attacks such as replay attacks, making it unsafe.

[0010] Patent publication number CN112581978A discloses a mechanism for detecting and locating sound events, focusing on accurately locating the direction and distance of a sound source. Through frequency domain processing, the precise location of the sound source is achieved. This invention does not focus on positioning and direction; its mission is to distinguish the source of the sound and determine whether the voice command received by the smart speaker comes from a real user or is forged by a machine. This serves as a defense against voice replay attacks. Summary of the Invention

[0011] To achieve robust, device-free passive liveness detection, this paper proposes a liveness detection system based on a microphone array. This system leverages the microphone arrays widely adopted by mainstream smart speakers (for example, Amazon Echo 3rd Gen and Google Home Max both have six microphones) to effectively defend against voice spoofing attacks generated by electronic devices. Because different microphones in the microphone array are positioned differently, the diversity of the collected audio is significantly enhanced. Leveraging audio diversity, this paper can extract more liveness information relevant to the target user, which is expected to significantly improve the robustness and accuracy of liveness detection.

[0012] The technical solutions of the present invention are as follows:

[0013] A voice deception attack detection system based on a microphone array is characterized by including:

[0014] The data acquisition module uses a microphone array to collect voice commands and save them as multi-channel voice;

[0015] A pre-processing module is used to perform time domain and frequency domain analysis on the speech signal of each channel, as well as direction detection of the speech signal;

[0016] The feature extraction module is used to calculate the standard deviation between the audio spectra of multiple microphones and convert it into a spectrum-based array fingerprint F SAP , according to the differences in the macroscopic features of the audio spectra from different microphones, the fingerprint F based on the spectrum distribution of different channels is extracted SDP , combined with the linear prediction cepstral coefficient LPCC, fingerprint F based on multi-channel linear prediction cepstral coefficient LPC , construct array fingerprint X = [F SAP ,F SDP ,F LPC ];

[0017] The attack detection module uses a feedforward back-propagation neural network to perform binary classification based on the array fingerprint constructed by S3, effectively distinguishing between the user's real voice and the machine-generated playback speech to complete the detection.

[0018] A method for detecting voice spoofing attacks based on a microphone array is characterized by:

[0019] S1. Sampling: N microphones are sampled at a sampling rate of F s , sampling time T, sampling, and the collected voice sample is recorded as V, the audio data dimension is M×N, where M=F s *T, M is the number of samples obtained within the sampling time T;

[0020] S2. Preprocessing: Perform multi-channel signal frequency analysis and voice signal direction detection on the voice sample V;

[0021] S3. Feature extraction: Calculate the standard deviation between the audio spectra of multiple microphones and convert it into a spectrum-based array fingerprint F SAP , according to the differences in the macroscopic features of the audio spectra from different microphones, the fingerprint F based on the spectrum distribution of different channels is extracted SDP , combined with the linear prediction cepstral coefficient LPCC, fingerprint F based on multi-channel linear prediction cepstral coefficient LPC , construct array fingerprint X = [F SAP ,F SDP ,F LPC ];

[0022] S4. Attack Detection: For the array fingerprint constructed in S3, a feedforward back-propagation neural network is selected to perform a binary classification task. That is, the existing feature data is labeled and divided into a training set and a test set. The array fingerprint is input into the neural network classifier for classification. The prediction results of 0 and 1 are obtained on the test set. 0 corresponds to a real person's voice, and 1 corresponds to a machine-generated playback sound. This effectively distinguishes the user's real voice from the machine-generated playback speech, completing the detection.

[0023] This paper defines a new feature called "array fingerprinting" for audio captured from multiple channels. Specifically, array fingerprinting leverages the correlation between data from different channels to eliminate distortion caused by factors such as air flow and user position variations. This paper proposes using a circular microphone array in a smart speaker to perform passive liveness detection in a smart home environment.

[0024] In order to achieve clear and high-fidelity multi-channel data collection, the data acquisition module needs to rely on an open development board with a voice interface and set a suitable sampling rate for audio collection.

[0025] To pre-process the collected data and ensure its high availability, it is necessary to perform multi-channel signal frequency analysis and sound source direction detection, find key features that can be used for identification, and determine the approximate location of the sound source.

[0026] The present invention mainly selects three representative features for feature extraction: spectrum-based array fingerprint (SpectrogramArray Fingerprint, F SAP Spectrogram Distribution Fingerprint (F SDP ) and features based on multi-channel LPCC coefficients (Channel LPCC Feature, FLPC By observing the collected audio spectrum itself and the differences between the spectra of different microphones (i.e., array fingerprints), the present invention extracts and captures features. By combining the three features, a final feature vector is formed for subsequent classification and judgment. At the same time, during feature extraction, the present invention takes into account factors such as time overhead, ensuring low-latency and high-accuracy smart home liveness detection throughout the entire process.

[0027] The classification model generates feature vectors from multi-channel speech signals collected by a microphone array and then selects a feedforward back-propagation neural network to perform liveness detection. A lightweight neural network is used because it enables fast decision-making, which is crucial for devices in smart home environments.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] (1) A novel passive detection system. The array liveness detection system of the present invention can prevent voice spoofing attacks. The present invention uses only audio collected from smart speakers, without requiring users to carry any equipment or perform other operations.

[0030] (2) Highly robust detection feature. This paper theoretically analyzes the principles behind passive detection and proposes a highly robust liveness detection feature: array fingerprint. This novel feature not only improves efficiency but also broadens the application scenarios of passive liveness detection.

[0031] (3) In specific applications, the present invention can still ensure the effectiveness and usability of liveness detection tasks when facing multiple factors (e.g., distance, direction, deceptive devices, background noise). Specifically, the array-based characteristics and effective feature selection ensure the superiority of this solution compared to existing methods.

[0032] (4) Invention patent CN111639519A is limited to detecting the presence of living organisms or other active objects inside a box and is unable to identify and defend against voice spoofing commands. This invention proposes a defense against voice replay attacks based on a microphone array, which can specifically determine the nature of the sound source and effectively distinguish whether it is a machine-generated or human voice.

[0033] (5) Invention patent CN112581978A can only use a microphone array to identify the location and behavior of the sound source. It cannot identify whether the sound source comes from a real user or an electronic device, and cannot defend against voice deception commands. The present invention proposes a defense method against voice replay attacks based on a microphone array. It has the characteristics of being lightweight and easy to deploy. It does not require precise positioning and position matching of sound events. Instead, it analyzes the type of sound source from the spectral characteristics, whether it comes from a machine forgery or a human voice. The present invention has the advantages of stable features and simple detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Flow chart of the present invention

[0035] Figure 2 Schematic diagram of a circular microphone array DETAILED DESCRIPTION

[0036] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. It should be noted that those skilled in the art may make various modifications and improvements without departing from the scope of the present invention, and these modifications and improvements fall within the scope of protection of the present invention.

[0037] The present invention provides a passive microphone array-based liveness detection solution that can be deployed in smart homes, including the following steps:

[0038] Step 1: Data collection. Use the built-in microphone arrays in popular smart speakers like Amazon Echo and Google Home to collect voice audio. More generally, you can use a development board with a voice interface to collect audio data.

[0039] Step 2: Data preprocessing. First, use signal processing methods such as short-time Fourier transform, fast Fourier transform, and signal segmentation to perform frequency domain analysis of multi-channel signals. Preprocessing methods can improve the usability of collected data.

[0040] Step 3: After data preprocessing, the present invention performs further feature extraction and obtains three representative features: spectrum-based array fingerprint, spectrum distribution-based fingerprint, and multi-channel LPCC coefficient-based features, preparing for final classification and detection.

[0041] In step 4, after generating feature vectors from the multi-channel speech signals collected by the microphone array, the present invention deploys a lightweight feedforward back-propagation neural network to perform liveness detection, which can achieve rapid response to decision-making.

[0042] The technical solutions provided by the present invention are described in further detail below with reference to the accompanying drawings.

[0043] like Figure 1 As shown, the research method and analysis steps of the present invention include a data acquisition module, a preprocessing module, a feature extraction module, and an attack detection module. The detailed steps are as follows:

[0044] The present invention first uses a development board with a voice interface to collect data, and uses its equipment similar to that of a commercial smart speaker to achieve effective data collection. Specifically, for a smart speaker with N microphones, s The sampling rate is set to 100, and the data collection time is defined as T. After obtaining the data, the collected voice sample is recorded as V M×N , where M = F s *T, the dimension of the audio data is M × N. The collected V is transmitted to the data preprocessing module.

[0045] During the preprocessing process, the identity of the audio (i.e., the real user or the voice deception device) is hidden in the spectrum of the collected multi-channel audio. Therefore, before feature extraction, it is necessary to perform frequency analysis on the signal of each channel. In addition, in order to better utilize the characteristics of multi-channel, the present invention detects the direction of the audio. First, multi-channel signal frequency analysis must be performed, because the audio spectrum in the time-frequency domain contains key features for further liveness detection. The present invention first performs a short-time Fourier transform to obtain a two-dimensional spectrum of the audio signal of each channel. For the audio V of the i-th channel containing M samples, i , the Hanning window is applied to divide the signal into several blocks. The sliding window method is adopted when dividing the signal. The length and activity size of each window should be set to an appropriate length. Finally, a 4096-point fast Fourier transform is performed on each block, and the spectrum graph S is obtained. i Another task in the preprocessing process is to perform sound source direction detection. M×N , it is necessary to calculate the direction of the sound source in order to better utilize the multi-channel characteristics. First, a high-pass filter with a cutoff frequency of 100Hz is used to process V M×N The audio signal V′ on each channel i . Then, for the i-th microphone M i The filtered signal is used to calculate the error E i =mean((V′(:,i-1)-V′(:,i)) 2 ). Finally, from the calculated results, find the microphone with the smallest error and regard it as the microphone closest to the sound source.

[0046] In the feature extraction process, based on the audio spectrum differences between normal speech and deceptive forged speech, that is, the collected audio spectrum itself, and the differences between the spectra of different microphones (that is, array fingerprints), the present invention proposes three representative feature selections: spectrum-based array fingerprints, spectrum distribution-based fingerprints, and multi-channel LPCC coefficient-based features.

[0047] Starting with the sound propagation model of speakers, spectrum-based array fingerprinting primarily addresses the following three questions: How can we model sound propagation in smart speaker scenarios and answer why existing features perform poorly in such scenarios? How can we extract liveness detection features (i.e., array fingerprints) from multi-channel speech samples that are robust to user location and microphone array layout? Can array fingerprints effectively address environmental factors and robustly perform liveness detection?

[0048] like Figure 2 As shown in Figure 2, based on the characteristics of the circular layout of the array microphone, the array fingerprint is defined as follows:

[0049] A F =std(log[y1,y2,...,y N ])

[0050] =std(C-α(s(f,t))·lg(e)·[d1,d2,...,d N ])

[0051] =-α(s(f,t))·lg(e)·std([d1,d2,...,d k ])

[0052] =A F (s(f, t),|σ d )

[0053] where y k Represents the audio V collected by the kth microphone k The location of the sound source is (L, 0). The distance from the sound source to the kth microphone is d k s(f, t) is the audio source, which contains the identity information of whether the audio is from a real user or a spoofing device. C is the attenuation of the transmission channel. F The formula shows that the feature is composed of the source audio s(f, t) and the standard deviation σ of the distance d =std([d1, d2, ..., d N ]) is determined by Figure 2 Medium k The definition of σ can be deduced from the above. dBasically a constant value. Based on the defined array fingerprint, the spectrum-based array fingerprint is further improved. In order to reduce the computational overhead, for the spectrum S of size Ms×Ns k , only keep the frequencies less than the cutoff frequency f sap In this study, we empirically sap The resized spectrum is expressed as Spec = [Spec1, Spec2, ..., Spec k ], where Spec k =S k (:M spec ,:),M spec By f sap It is determined together with Ms and is set to 426 in the present invention. Then the obtained Spec is converted into M G ×N G The grid matrix G k Then, according to the defined array fingerprint formula, the initial array fingerprint feature F based on the spectrum is obtained. G (i,j)=std([G1(i,j),G2(i,j),...,G N (i, j)]), after five-point moving average and normalization to remove excess noise, the final spectrum-based array fingerprint F SAP .

[0054] The second representative feature proposed by this invention is the fingerprint based on the spectrum distribution of different channels. The distribution characteristics of the spectrum itself also provide useful information related to the identity behind the sound source. Therefore, the fingerprint F based on the spectrum distribution is extracted. SDP Used for liveness detection. The spectrum from real people and spoofing devices is very different. Therefore, the frequency amplitude intensity of the audio from different channels is used to detect the presence of liveness. k , to average and resample its length to N Ch , as F SDP At the same time, for the same audio, the amplitude and distribution of Ch in different channels are slightly different. The present invention slices the spectrum and calculates the cumulative distribution function to obtain the corresponding index matrix μ. Finally, its average value and standard deviation are obtained as part of the spectrum graph feature, which are the latter two components. Therefore, the fingerprint based on spectrum distribution can be recorded as

[0055] The third representative feature proposed by the present invention is a feature based on multi-channel linear prediction cepstral coefficients (LPCC) coefficients. Therefore, retaining the LPCC that characterizes a given audio signal can further improve the detection performance. k The collected audio signal yk (t), in order to calculate the LPCC of order p, the present invention first calculates its linear predictive coding (LPC). After obtaining the LPC parameters, the feature LPCC coefficients of the last dimension can be obtained by calculating according to the iterative formula. For LPCCs from different channels, the present invention calculates their average value as the last set of features F LPC Finally, the final feature vector X=[F SAP , F SDP , F LPC ].

[0056] After feature extraction, classification is required to achieve the task purpose of liveness detection of the present invention. The present invention proposes to use a neural network to perform the classification task. Specifically, a feedforward back-propagation neural network is selected to perform liveness detection. The neural network is a lightweight network that only contains three hidden layers with linear rectification activation functions (the three layer sizes are 64, 32, and 16 respectively). The dropout after the 64 and 32 node layers is set to 20%, and the output layer is an activation node of 1 sigmoid function. The reason for using a lightweight neural network is that it can achieve a fast response to decision-making, which is crucial for devices in a smart home environment.

[0057] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various variations or modifications within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A voice deception attack detection system based on microphone array, characterized in that: include: The data acquisition module uses a microphone array to collect voice commands and save them as multi-channel voice; A pre-processing module is used to perform time domain and frequency domain analysis on the speech signal of each channel, as well as direction detection of the speech signal; The feature extraction module is used to calculate the standard deviation between the audio spectra of multiple microphones and convert it into a spectrum-based array fingerprint F SAP , according to the differences in the macroscopic features of the audio spectra from different microphones, the fingerprint F based on the spectrum distribution of different channels is extracted SDP , combined with the linear prediction cepstral coefficient LPCC, fingerprint F based on multi-channel linear prediction cepstral coefficient LPC , construct array fingerprint X = [F SAP ,F SDP ,F LPC ]; The attack detection module uses a feedforward back-propagation neural network to perform binary classification based on the array fingerprint constructed by S3 feature extraction. This effectively distinguishes between the user's real voice and the machine-generated playback speech, completing the detection. The spectrum-based array fingerprint F in the S3 feature extraction SAP , the specific steps are as follows: S3.1 Based on the characteristics of the circular layout of the array microphone, the array fingerprint is defined as follows: A F =std(log[y1,y2,…y k …,y N ]) =std(C-α(s(f,t))·lg(e)·[d1,d2,…d k …,d N ]) =-α(s(f,t))·lg(e)·std([d1,d2,…d k …,d N ]) =A F (s(f,t),∣σ d ) Among them, y k represents the audio collected by the kth microphone, d k is the distance from the sound source to the kth microphone, s(f,t) is the audio at the sound source, which contains the identity information of whether the audio is from a real user or a spoofing device; C is the attenuation of the transmission channel; S3.2 adjusts the spectrum Spec, the formula is as follows: Spec=[Spec1,Spec2,…,Spec k ], Where Spec k =S k (∶M spec ,∶), M spec By the sampling cutoff frequency f sap , the sampling rate Fs, and the number of FFT points are jointly determined; S3.3 Convert the obtained Spec into M G ×N G The grid matrix G k , the formula is as follows: G k (i,j)=sum(Spec k (1+(i-1)·S M :i·S M 1+(j-1)·S N :j·S N )) In the formula and are the width and length of each block; S3.4 Spectral array fingerprint initial feature F G (i,j)=std([G1(i,j),G2(i,j),…,G N (i, j)]), after five-point moving average and normalization to remove excess noise, the final spectrum-based array fingerprint F is obtained SAP .

2. A method for detecting voice spoofing attacks based on a microphone array, characterized in that: include: S1. Sampling: N microphones are sampled at a sampling rate of F s , sampling time T, sampling, and the collected voice sample is recorded as V, the audio data dimension is M×N, where M=F s *T, M is the number of samples obtained within the sampling time T; S2. Preprocessing: Perform multi-channel signal frequency analysis and voice signal direction detection on the voice sample V; S3. Feature extraction: Calculate the standard deviation between the audio spectra of multiple microphones and convert it into a spectrum-based array fingerprint F SAP , according to the differences in the macroscopic features of the audio spectra from different microphones, the fingerprint F based on the spectrum distribution of different channels is extracted SDP , combined with the linear prediction cepstral coefficient LPCC, fingerprint F based on multi-channel linear prediction cepstral coefficient LPC , construct array fingerprint X = [F SAP ,F SDP ,F LPC ]; S4. Attack Detection: A feedforward back-propagation neural network is used to perform a binary classification task based on the array fingerprint constructed in S3. This involves labeling the existing feature data and dividing it into a training set and a test set. The array fingerprint is then input into the neural network classifier for classification. The prediction results on the test set are 0 and 1, with 0 corresponding to a real person's voice and 1 corresponding to a machine-generated playback sound. This effectively distinguishes the user's real voice from the machine-generated playback sound, completing the detection. The spectrum-based array fingerprint F in the S3 feature extraction SAP , the specific steps are as follows: S3.1 Based on the characteristics of the circular layout of the array microphone, the array fingerprint is defined as follows: A F =std(log[y1,y2,…y k …,y N ]) =std(C-α(s(f,t))·lg(e)·[d1,d2,…d k …,d N ]) =-α(s(f,t))·lg(e)·std([d1,d2,…d k …,d N ]) =A F (s(f,t),∣σ d ) Among them, y k represents the audio collected by the kth microphone, d k is the distance from the sound source to the kth microphone, s(f,t) is the audio at the sound source, which contains the identity information of whether the audio is from a real user or a spoofing device; C is the attenuation of the transmission channel; S3.2 adjusts the spectrum Spec, the formula is as follows: Spec=[Spec1,Spec2,…,Spec k ], Where Spec k =S k (∶M spec ,∶), M spec By the sampling cutoff frequency f sap , the sampling rate Fs, and the number of FFT points are jointly determined; S3.3 Convert the obtained Spec into M G ×N G The grid matrix G k , the formula is as follows: G k (i,j)=sum(Spec k (1+(i-1)·S M :i·S M 1+(j-1)·S N :j·S N )) In the formula and are the width and length of each block; S3.4 Spectral array fingerprint initial feature F G (i,j)=std([G1(i,j),G2(i,j),…,G N (i, j)]), after five-point moving average and normalization to remove excess noise, the final spectrum-based array fingerprint F is obtained SAP .

3. The method for detecting voice spoofing attacks based on a microphone array according to claim 2, wherein: include: The S2 pre-processing, multi-channel signal frequency analysis specifically includes: first performing a short-time Fourier transform to obtain a two-dimensional spectrogram of each channel audio signal, where the two dimensions are time and frequency; For the audio V from the i-th channel i , which contains M samples, the Hanning window is applied to divide the signal into several blocks; finally, a fast Fourier transform is performed on each block to obtain a two-dimensional spectrum map S i .

4. The method for detecting voice spoofing attacks based on a microphone array according to claim 2, wherein: include: The S2 pre-processing, the direction detection of the voice signal, specifically includes, first, using a high-pass filter to process V M×N The audio signal V on each channel i '; Then, for the i-th microphone M i The filtered signal is used to calculate the error E i =mean((V′(:,i-1)-V′(:,i)) 2 ); Finally, from the calculated results, find the microphone with the smallest error and regard it as the microphone closest to the sound source.

5. The method for detecting voice spoofing attacks based on a microphone array according to claim 2, wherein: include: The fingerprint of the S3 feature extraction based on the spectrum distribution of different channels The specific steps are as follows: S3.5 Using the frequency amplitude intensity Ch of audio from different channels k , take the average value of the channel frequency strength and resample its length to N Ch , the formula is as follows: As F SDP The first component of S3.6 To characterize the characteristic distribution of each vocal channel Ch, calculate the cumulative distribution function of each channel, slice the spectrum, and obtain the corresponding index matrix μ through the probability accumulation function; S3.7 Calculate the average value D for each column of the obtained index matrix μ mean and standard deviation D std As part of the spectrogram features, namely the last two components.

6. The method for detecting voice spoofing attacks based on a microphone array according to claim 2, wherein: include: The fingerprint F based on multi-channel linear prediction cepstral coefficients in the S3 feature extraction LPC , the specific steps are as follows: S3.8 For the kth microphone M of the smart speaker k The collected audio signal y k (t), and calculate the linear predictive code LPC, record it as a, a=LPC(y k (t),p), where p is the order of LPC, and the collected LPC representation a=[a0,a1,…,a p ]; S3.9 For LPCC coefficients c=[c0,c1,…,c p ], c0=ln(p) Finally, the feature LPCC coefficients of the last dimension are obtained. For LPCCs from different channels, their average value is calculated as the last set of features F LpC .

7. The method for detecting voice spoofing attacks based on a microphone array according to claim 2, wherein: Including the attack detection step S4, the specific steps are as follows: S4.1 selects a feedforward back-propagation neural network consisting of three hidden layers with a linear rectifier activation function; S4.2 Input the calibrated feature vector into the neural network classifier, divide the feature data into 7:3 parts, and use them as training and test sets. Verify the classification results on the test set. Please explain i·S M 1+(j-1)·S N The input of the classifier is the feature vector obtained by combination, and the classification result is the judgment value of 0 or 1, where 0 corresponds to the real voice and 1 corresponds to the replayed sound generated by the machine, thereby effectively distinguishing the user's real voice and the replayed speech generated by the machine and completing the detection.

Images