An intelligent voiceprint anti-counterfeiting method, device and equipment based on a hop signal

Abstract

This invention discloses an intelligent voiceprint anti-counterfeiting method, device, and equipment based on hops signals. It collects the voiceprint signals generated by hops hitting the bottle wall, then denoises the signals and extracts PLP features. These feature signals are then input into a CNN-GMM-HMM voiceprint model for training to construct an original standard database. After collecting sample voiceprint signals, they are preprocessed and then matched and identified with the voiceprint information in the standard database to achieve authenticity verification. Compared to traditional identification methods, this method can identify genuine and counterfeit liquor faster and more accurately, greatly improving work efficiency while helping to ensure the quality and safety of liquor.

Description

Technical Field

[0001] This invention belongs to the field of anti-counterfeiting technology, and relates to digital information anti-counterfeiting technology, specifically to an intelligent voiceprint anti-counterfeiting method, device and equipment based on hop signal. Background Technology

[0002] As living standards continue to improve, the liquor market is growing, but so is the prevalence of counterfeit and substandard liquor. To better distinguish genuine liquor from counterfeit, anti-counterfeiting technologies have been widely adopted. Currently, the mainstream anti-counterfeiting methods for liquor are mostly physical anti-counterfeiting measures based on packaging, such as labels, bottle caps, and identification codes. However, these methods are easily counterfeited and difficult to identify, leading to the circulation of counterfeit and substandard products in the market.

[0003] Information technology-based anti-counterfeiting, as a method utilizing computer technology and information security technology, has advantages such as strong concealment, scalability, and visualization. Currently, digital information technologies such as digital watermarking, encryption, and QR codes are used to create anti-counterfeiting labels and patterns to improve the anti-counterfeiting performance of liquor packaging. However, these digital information technology anti-counterfeiting methods, whether digital watermarking, encryption, or QR codes, generally suffer from vulnerabilities that allow them to be easily attacked, cracked, or tampered with. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent voiceprint anti-counterfeiting method, device and equipment based on the bubbling signal of liquor. It can identify the authenticity of liquor based on the unique voiceprint signal of liquor bubbling, and has the advantages of being safer, more reliable and more efficient.

[0005] To achieve the above objectives, the present invention adopts the following technical solution.

[0006] A smart voiceprint anti-counterfeiting method based on hop signals includes the following steps:

[0007] Step 1: Collect several raw voiceprint signals;

[0008] Step 2: The acquired voiceprint signal is denoised and pre-characterized using an autoencoder;

[0009] The autoencoder is a stacked sparse autoencoder. The original voiceprint signal is input into a deep neural network model SAE composed of multiple sparse autoencoders for training. Then the decoding layer is removed. The signal that has undergone the first training feature processing is used as the input for the second SAE training to generate the second feature. Then the decoding output is performed to obtain the pre-featured data.

[0010] Step 3: Extract PLP features from the pre-featured data in Step 2 using a speech feature extraction tool, and then input it into the CNN-GMM-HMM voiceprint model for learning and training to obtain the standard database of the CNN-GMM-HMM voiceprint model.

[0011] The CNN-GMM-HMM voiceprint model includes: a CNN module, used to extract and classify PLP features through convolutional layers, pooling layers, and fully connected layers; a GMM module, used to calculate the mean and variance of the PLP feature vector after convolution by the CNN module; and an HMM module, used to train the initial probability, state transition probability A, and final output probability B of the voiceprint signal samples.

[0012] Step 4: Repeatedly collect data on the liquor to be tested using an acoustic vibration sensor to obtain the acoustic signature signal of the sample liquor.

[0013] Step 5: The collected sample hops voiceprint signals are processed using the same denoising and pre-featureization process as in Step 2 using the stacked sparse autoencoder.

[0014] Step 6: Input the preprocessed voiceprint signal from Step 5 into the standard database from Step 3 for identification and comparison;

[0015] Step 7: Display and output the identification and verification results.

[0016] Furthermore, in step 3, the pre-featured data from step 2 is processed using the speech feature extraction tool OpenSMILE to extract PLP features. First, the voiceprint data is processed by framing and windowing to extract the collected voiceprint signal segments. Then, a fast Fourier transform is performed to obtain the spectrum. Next, the amplitude is squared for amplification. Then, the Bark filter bank is used for processing, followed by equal loudness pre-emphasis and intensity-loudness conversion. Finally, an inverse Fourier transform is performed for linear prediction to obtain PLP features.

[0017] Furthermore, in step 6, the PLP features in the voiceprint model are first processed by using a separate convolutional layer with amplitude spectrum as input, operating independently on each channel, and performing max pooling across channels. The channel with the largest response in each node is selected to extract and classify the PLP features. Then, the extracted PLP features are used to enter the GMM model for cluster training. After that, the phonemes are represented as hidden variables and imported into the HMM model for training.

[0018] Training of the GMM-HMM begins by initializing the GMM model using an average distribution approach. Then, the maximum likelihood estimate, i.e., the sample mean, is calculated using the existing estimates of the hidden variables. and variance 2 :

[0019]

[0020] 2

[0021] Then, the state transition probability is obtained statistically; next, realignment is performed, and the voiceprint signal is aligned according to the transition probability, mean, and variance obtained above. After setting the basic parameters, the training is repeated a certain number of times until convergence is achieved, thus obtaining the standard database of CNN-GMM-HMM voiceprint models.

[0022] Furthermore, the identification and comparison in step 6 specifically involves:

[0023] First, calculate the false acceptance rate and false rejection rate for each threshold point of the voiceprint signal after preprocessing in step 5.

[0024] Then, ROC curves are plotted based on the calculated precision and recall for each threshold point;

[0025] Finally, find the x-coordinate X value corresponding to the intersection of the ROC curve and the diagonal in the ROC space. The smaller the value, the higher the authenticity of the tested liquor sample.

[0026] A smart voiceprint anti-counterfeiting device based on hop signals, comprising:

[0027] The hop soundprint signal excitation unit generates a cyclic alternating magnetic field through electromagnetic induction, which uses the magnetic force of attraction / repulsion to make the test sample shake repeatedly, thereby generating a hop soundprint signal.

[0028] The hops acoustic signature signal acquisition unit is an acoustic vibration sensor made of piezoelectric ceramics and damping materials. It is used to detect the vibration generated locally between the hops in the bottle and the internal wall of the bottle, and obtain the sample hops acoustic signature signal.

[0029] The hop voiceprint signal recognition unit is configured to execute the voiceprint anti-counterfeiting method to identify the voiceprint features collected by the hop voiceprint signal acquisition unit and a standard database to obtain the verification result of the hop voiceprint signal.

[0030] Furthermore, the hop voiceprint signal excitation unit includes a first braking frame and a second braking frame.

[0031] The first braking frame, as an integral support, is composed of a horizontal moving base and a vertical drive frame. The vertical drive frame is mounted on the horizontal moving base, and the second braking frame is suspended on the vertical drive frame. The second braking frame for placing the liquor sample consists of a liquor bottle mouth slot, an AC-current-conducting coil, and a liquor bottle support. The AC-current-conducting coil is located at one end of the liquor bottle support, the liquor bottle mouth slot is located at the other end of the liquor bottle support, and a magnet is located on the vertical drive frame near the AC-current-conducting coil. The acoustic vibration sensor of the liquor foam acoustic signal acquisition unit is placed at the center of the liquor bottle mouth slot.

[0032] A smart voiceprint anti-counterfeiting device based on hop signals includes: a cloud server ECS and a file storage NAS;

[0033] The cloud server ECS is used to execute the program of the intelligent voiceprint anti-counterfeiting method based on hop signal;

[0034] The file storage NAS is used to store computer programs and a standard database of constructed CNN-GMM-HMM voiceprint models. When the computer programs are executed by a processor, they implement the method described above.

[0035] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects:

[0036] This invention discloses an intelligent voiceprint anti-counterfeiting device and equipment based on hops signals. It collects the voiceprint signals generated by hops hitting the bottle wall, then denoises the signals and extracts PLP features. These features are then input into a CNN-GMM-HMM voiceprint model for training to construct an original standard database. After collecting sample voiceprint signals, they are preprocessed and then matched and identified with the voiceprint information in the standard database to achieve authenticity verification. Compared to traditional identification methods, this method can identify genuine and counterfeit liquor faster and more accurately, greatly improving work efficiency while helping to ensure the quality and safety of liquor.

[0037] The formation of bubbles in baijiu (Chinese white liquor) is the result of the combined effects of multiple factors, including alcohol concentration, temperature, and time. Therefore, these bubbles, as a unique characteristic of baijiu, are difficult to replicate or counterfeit. Based on this theory, this invention uses the unique sound signature signal of baijiu bubbles to distinguish genuine baijiu from counterfeit products. This is a novel and effective anti-counterfeiting technology for baijiu, providing baijiu brands with a safer, more reliable, and efficient method.

[0038] This invention provides an intelligent voiceprint anti-counterfeiting method, device, and equipment based on hop signals, which can identify the authenticity of products without damaging the bottle / box packaging. At the same time, it uses unique voiceprint features for anti-counterfeiting verification, which can not only effectively prevent the emergence of counterfeit and shoddy products, but also bring more market promotion opportunities to product brands. Attached Figure Description

[0039] Figure 1 This is a flowchart of an intelligent voiceprint anti-counterfeiting method based on hop signals in one embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram illustrating the working principle of a stacked sparse autoencoder in one embodiment of the present invention;

[0041] Figure 3 This is a flowchart illustrating the principle of PLP feature extraction in one embodiment of the present invention;

[0042] Figure 4 This is a flowchart illustrating the training of voiceprint signals in a voiceprint model according to one embodiment of the present invention;

[0043] Figure 5 This is a flowchart illustrating sample voiceprint signal recognition in one embodiment of the present invention;

[0044] Figure 6a This is a schematic diagram of the composition of an intelligent voiceprint anti-counterfeiting device based on hop signals in one embodiment of the present invention;

[0045] Figure 6b This is a schematic diagram of the specific structure of the hop voiceprint signal excitation unit in one embodiment of the present invention;

[0046] Figure 7 This is a schematic diagram of an intelligent voiceprint anti-counterfeiting device based on hop signals according to an embodiment of the present invention. Detailed Implementation

[0047] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0048] Figure 1 This is a flowchart of an intelligent voiceprint anti-counterfeiting method based on hop signals according to an embodiment of the present invention. The intelligent voiceprint anti-counterfeiting method based on hop signals is used in an electronic device and is executed by the electronic device, including the following steps:

[0049] In step S11, acoustic vibration sensors are used to collect acoustic signature data of the bubbles hitting the bottle wall and save it in WAV format. The acoustic signature data consists of acoustic signatures of ten different types of liquor. The acoustic signatures of each type of liquor are collected 100 times in a silent environment, resulting in 1,000 voice data points with noise.

[0050] In step S12, all collected voiceprint signals are subjected to denoising and pre-featureization processing using an autoencoder, wherein the autoencoder is designed as a stacked sparse autoencoder. Figure 2This is a schematic diagram illustrating the working principle of Stacked Sparse Autoencoders (SAEs). The original acquired speaker signal is used as input for SAE training; then the decoding layer is removed; the signal processed by the first training feature is used as input for the second SAE training to generate the second set of features, which are then decoded and output.

[0051] In step S13, the pre-featured data from step 2 is saved to a folder as input. Using the cmd command-line interface, switch to the directory where the pre-featured data is saved, and use the speech feature extraction tool OpenSMILE to execute the feature extraction command SMILExtract_Release -C configuration file path -I "audio path to be processed" -O "path and filename to save feature vectors" to extract PLP features. Specifically, first, frame-segmentation and windowing processing are performed to extract segments of the acquired voiceprint signal. The frame shift size is set to 5-10ms, and the window length is 20-30ms. Then, a fast Fourier transform is performed, followed by amplitude squaring for amplification. Next, the Bark filter bank is used for processing, followed by equal loudness pre-emphasis and intensity-loudness conversion. Finally, an inverse Fourier transform is performed for linear prediction to obtain PLP features. The principle and process are as follows: Figure 3 As shown.

[0052] Then, it is fed into the CNN-GMM-HMM voiceprint model for learning. The CNN-GMM-HMM voiceprint model includes: a CNN module, which extracts and classifies PLP features through convolutional layers, pooling layers, and fully connected layers; a GMM module, which calculates the mean and variance of the PLP feature vector after convolution by the CNN module; and an HMM module, which trains the initial probability, state transition probability A, and final output probability B of the voiceprint signal samples.

[0053] In the voiceprint model, a separate convolutional layer is first used with the amplitude spectrum as input, operating independently on each channel and performing max pooling across channels. The channel with the largest response in each node is selected to extract and classify PLP features. Then, the extracted PLP features are used to enter the GMM model for cluster training. This is mainly because each PLP feature is determined by a phoneme, and the phoneme is then represented as a hidden variable and imported into the HMM model. Therefore, each feature is determined by several states.

[0054] Training of GMM-HMM begins by first initializing the GMM model using an average allocation method. Then, the maximum likelihood estimate, i.e., the sample mean, is calculated using the existing estimates of the hidden variables. ) and variance ( 2 ):

[0055]

[0056] 2

[0057] Then, the state transition probability is obtained statistically; next, realignment is performed, and the voiceprint signal is aligned according to the parameters obtained above (transition probability, mean, variance). After setting the basic parameters, the number of repetitions for training is set:

[0058] def EM_step(data, pi, mean, var, update_times):

[0059] for i in range(update_times):

[0060] W, pi = E_step(data, pi, mean, var)

[0061] mean, var = M_step(data, W)

[0062] return pi, mean var

[0063] Convergence yields the standard database for CNN-GMM-HMM voiceprint models. The entire database training and construction process is as follows: Figure 4 As shown.

[0064] In step S14, the acoustic signature signal of the liquor to be tested is collected. This collection can be performed repeatedly by an acoustic vibration sensor.

[0065] In step S15, the collected voiceprint signal is input into the stacked sparse autoencoder for the first encoding. The resulting feature vector is used as the hidden layer to enter the next SAE training layer to generate the second feature vector before decoding and output.

[0066] Preferably, for the same hop, multiple sample voiceprint signals are acquired; these multiple sample voiceprint signals are input into a stacked sparse autoencoder, which has encoding and decoding capabilities. After encoding, a feature vector is obtained as a hidden layer, and then the feature vector of the hidden layer is decoded to form a second feature vector. The entire process can maintain the previous voiceprint signal features fixed and ignore the interaction with subsequent features, thereby reducing the search of the parameter space. The formula is expressed as follows:

[0067] F L =

[0068] Among them, F L This represents the representation learned from the top-level L.

[0069] In step S16, the voiceprint signal feature vector after the autoencoder encoding-decoding process is input into the original standard database for matching and recognition. Specifically, firstly, the false acceptance rate and false rejection rate corresponding to each threshold point of the voiceprint signal after pre-featured processing in step 5 are calculated; then, an ROC curve (Receiver Operating Characteristic) is plotted based on the calculated precision and recall for each threshold point, with the vertical axis representing the true positive rate (TPR) and the horizontal axis representing the false positive rate (FPR). Next, the x-coordinate value corresponding to the intersection of the ROC curve and the diagonal in the ROC space is found. The smaller the value, the higher the authenticity of the detected liquor sample. It is considered that an X-value between 0 and 0.2 indicates a probability of over 98% for the identification result being genuine liquor. The voiceprint signal recognition process for the detected sample is as follows. Figure 5 As shown.

[0070] In step S15, the analysis and settlement results will be used to output the identification and detection results.

[0071] Corresponding to the method embodiments, this application also provides an intelligent voiceprint anti-counterfeiting device based on hop signal.

[0072] like Figure 6a The diagram shown is a structural schematic of an intelligent voiceprint anti-counterfeiting device based on hop signals provided in an embodiment of this application, including:

[0073] Hop voiceprint signal excitation unit;

[0074] Hop voiceprint signal acquisition unit;

[0075] Hops voiceprint signal recognition unit.

[0076] The specific manner in which each component of the apparatus in the above embodiments performs its operation is described in detail below.

[0077] The specific structure of the hop voiceprint signal excitation unit is as follows: Figure 6bAs shown, it mainly consists of two parts: a first braking frame 1 and a second braking frame 2. The first braking frame 1, as an integral support, is composed of a horizontal moving base 11 and a vertical drive frame 12. The vertical drive frame 12 is mounted on the horizontal moving base 11, and the second braking frame 2 is suspended on the vertical drive frame 12. It can move horizontally and vertically through the horizontal moving base 11 and the vertical drive frame 12, respectively. The second braking frame 2 is used to place the liquor sample and consists of three parts: a liquor bottle mouth slot 21, an AC-current-conducting coil 22, and a liquor bottle support 23. The AC-current-conducting coil 22 is located at one end of the liquor bottle support 23, the liquor bottle mouth slot 21 is located at the other end of the liquor bottle support 23, and the magnet 13 is located on the vertical drive frame 12 near the AC-current-conducting coil 22. The liquor foam acoustic signature signal acquisition unit is an acoustic vibration sensor made of piezoelectric ceramic and damping material. The acoustic vibration sensor of the liquor foam acoustic signature signal acquisition unit is placed at the center of the liquor bottle mouth slot 21.

[0078] The entire excitation unit works as follows: By supplying alternating current to the coil 22, an alternating magnetic field is generated due to the principle of electromagnetism. This magnetic field, combined with the magnet 13, creates an alternating attraction-repulsion force on the second braking frame 2, causing it to oscillate cyclically. This causes the foam to collide with the bottle wall, generating a sound signature signal, which is then collected by the acoustic vibration sensor of the foam sound signature signal acquisition unit on the bottle neck slot 21. The sensor then transmits the collected signal to the electronic device that integrates the foam sound signature signal recognition unit.

[0079] like Figure 7 The diagram shown is a schematic of an intelligent voiceprint anti-counterfeiting device based on hop signal provided in this application embodiment, which includes: a cloud server ECS and a file storage NAS;

[0080] The cloud server ECS is used to execute the program of the intelligent voiceprint anti-counterfeiting method based on hop signals, implementing each step of the intelligent voiceprint anti-counterfeiting method based on hop signals as described above. The server has 256 vCPUs, 6TB of memory, a main frequency of 3.8GHz, a maximum performance of 24 million PPS, 80Gbps, and a network latency of 20µs+.

[0081] The file storage NAS is used to store the entire program and the original standard database built on it, enabling data persistence under computing elasticity, while CNFS acceleration capability eliminates access latency caused by storage-compute separation.

Claims

1. A smart voiceprint anti-counterfeiting method based on hop signal, characterized in that... Includes the following steps: Step 1: Collect several raw voiceprint signals; Step 2: The acquired voiceprint signal is denoised and pre-characterized using an autoencoder; The autoencoder is a stacked sparse autoencoder. The original voiceprint signal is input into a deep neural network model SAE composed of multiple sparse autoencoders for training. Then the decoding layer is removed. The signal that has undergone the first training feature processing is used as the input for the second SAE training to generate the second feature. Then the decoding output is performed to obtain the pre-featured data. Step 3: Extract PLP features from the pre-featured data in Step 2 using a speech feature extraction tool, and then input it into the CNN-GMM-HMM voiceprint model for learning and training to obtain the standard database of the CNN-GMM-HMM voiceprint model. The CNN-GMM-HMM voiceprint model includes: a CNN module, used to extract and classify PLP features through convolutional layers, pooling layers, and fully connected layers; a GMM module, used to calculate the mean and variance of the PLP feature vector after convolution by the CNN module; and an HMM module, used to train the initial probability, state transition probability A, and final output probability B of the voiceprint signal samples. Step 4: Repeatedly collect data on the liquor to be tested using an acoustic vibration sensor to obtain the acoustic signature signal of the sample liquor. Step 5: The collected sample hops voiceprint signals are processed using the same denoising and pre-featureization process as in Step 2 using the stacked sparse autoencoder. Step 6: Input the preprocessed voiceprint signal from Step 5 into the standard database from Step 3 for identification and comparison; Step 7: Display and output the identification and verification results.

2. The intelligent voiceprint anti-counterfeiting method based on hop signal as described in claim 1, characterized in that: In step 3, the pre-featured data from step 2 is processed using the speech feature extraction tool OpenSMILE to extract PLP features. First, the voiceprint data is processed by framing and windowing to extract the collected voiceprint signal segments. Then, a fast Fourier transform is performed to obtain the spectrum. Next, the amplitude is squared for amplification. Then, the Bark filter bank is used for processing, followed by equal loudness pre-emphasis and intensity-loudness conversion. Finally, an inverse Fourier transform is performed for linear prediction to obtain PLP features.

3. The intelligent voiceprint anti-counterfeiting method based on hop signal as described in claim 1, characterized in that: In step 6, the PLP features are first extracted and classified in the voiceprint model using a separate convolutional layer with amplitude spectrum as input, operating independently on each channel and performing max pooling across channels. The channel with the largest response in each node is selected. Then, the extracted PLP features are used to enter the GMM model for cluster training. After that, the phonemes are represented as hidden variables and imported into the HMM model for training. Training of the GMM-HMM begins by initializing the GMM model using an average distribution approach. Then, the maximum likelihood estimate, i.e., the sample mean, is calculated using the existing estimates of the hidden variables. and variance 2 : 2 Then, the state transition probability is obtained statistically; next, realignment is performed, and the voiceprint signal is aligned according to the transition probability, mean, and variance obtained above. After setting the basic parameters, the training is repeated a certain number of times until convergence is achieved, thus obtaining the standard database of CNN-GMM-HMM voiceprint models.

4. The intelligent voiceprint anti-counterfeiting method based on hop signal as described in claim 1, characterized in that: The identification and comparison in step 6 specifically involves: First, calculate the false acceptance rate and false rejection rate for each threshold point of the voiceprint signal after preprocessing in step 5. Then, ROC curves are plotted based on the calculated precision and recall for each threshold point; Finally, find the x-coordinate X value corresponding to the intersection of the ROC curve and the diagonal in the ROC space. The smaller the value, the higher the authenticity of the tested liquor sample.

5. A smart voiceprint anti-counterfeiting device based on hop signal, characterized in that... include: The hop soundprint signal excitation unit generates a cyclic alternating magnetic field through electromagnetic induction, which uses the magnetic force of attraction / repulsion to make the test sample shake repeatedly, thereby generating a hop soundprint signal. The hops acoustic signature signal acquisition unit is an acoustic vibration sensor made of piezoelectric ceramics and damping materials. It is used to detect the vibration generated locally between the hops in the bottle and the internal wall of the bottle, and obtain the sample hops acoustic signature signal. The hops voiceprint signal recognition unit is configured to perform the intelligent voiceprint anti-counterfeiting method based on hops signals as described in any one of claims 1-4, and to identify the voiceprint features collected by the hops voiceprint signal acquisition unit against a standard database to obtain the verification result of the hops voiceprint signal.

6. The intelligent voiceprint anti-counterfeiting device based on hop signal as described in claim 5, characterized in that: The hop soundprint signal excitation unit includes a first braking frame (1) and a second braking frame (2). The first braking frame (1) is composed of a horizontal moving seat (11) and a vertical drive frame (12) as an integral support. The vertical drive frame (12) is installed on the horizontal moving seat (11), and the second braking frame (2) is suspended on the vertical drive frame (12). The second braking frame (2) for placing the liquor sample is composed of a liquor bottle mouth slot (21), a coil (22) that can conduct AC power, and a liquor bottle support (23). The coil (22) that can conduct AC power is set at one end of the liquor bottle support (23), the liquor bottle mouth slot (21) is set at the other end of the liquor bottle support (23), the magnet (13) is set on the vertical drive frame (12) on the side close to the coil (22) that can conduct AC power, and the acoustic vibration sensor of the liquor sound pattern signal acquisition unit is placed at the center of the liquor bottle mouth slot (21).

7. A smart voiceprint anti-counterfeiting device based on hop signal, characterized in that... include: ECS (Elastic Compute Service) cloud servers and NAS (Network Attached Storage); The cloud server ECS is used to execute the program of the intelligent voiceprint anti-counterfeiting method based on hop signal as described in any one of claims 1-4; The file storage NAS is used to store computer programs and a standard database of constructed CNN-GMM-HMM voiceprint models. When the computer programs are executed by a processor, they implement the method as described in any one of claims 1 to 4.

Images