Voice recognition wake-up method, device and system for anti-recording detection

Abstract

The present application discloses a voice recognition wake-up method, device and system for anti-recording detection, including: collecting current audio, extracting multi-dimensional features in the current audio; based on the extracted multi-dimensional features, setting the time dimension, aligning the current audio with the historical real voice on the time axis; constructing a multi-dimensional feature scatter plot based on the aligned current audio and historical real voice; comparing the generated multi-dimensional scatter plot of the aligned audio with the multi-dimensional scatter plot of the historical real voice, and calculating the comprehensive similarity; setting a similarity threshold and judging whether to record, when the comprehensive similarity value is higher than the set threshold, the current wake-up audio is the real user voice, otherwise it is a recording. By constructing a historical scatter ring group feature combination and a scatter ring group recognition model, it is possible to more quickly identify whether the current wake-up audio is real, and has a stronger recognition ability for the splicing attack of multiple recordings.

Description

Technical Field

[0001] The present invention relates to the field of speech processing technology, and in particular to a speech recognition wake-up method, device and system for anti-recording detection. Background Art

[0002] In the current era of digitalization and intelligence, speech recognition technology, as a key link in human-computer interaction, is increasingly widely used in many fields, including smart phones, smart homes, smart security, etc. However, with the continuous advancement of technology, speech recognition systems are also facing increasingly severe security challenges, especially the problem of recording attacks. Recording attacks refer to recording the user's real voice, and then editing, splicing or imitating it to bypass the system's authentication mechanism, thereby achieving the purpose of illegally controlling or obtaining sensitive information. In order to effectively resist such attacks, it is particularly important to develop efficient and accurate anti-recording detection technology.

[0003] For example, Chinese patent publication number CN111524528A discloses a voice wake-up method for anti-recording detection, including: when the user's initial voice data is detected, start the voice wake-up module to obtain the initial signal; determine the anti-recording detection strategy corresponding to the initial signal; perform anti-recording detection on the voice data according to the anti-recording detection strategy through the voice wake-up module and the algorithm module; when receiving the target instruction of passing the detection, turn off the microphone and start the preset application. The embodiment of the present application can control the voice wake-up process according to the anti-recording detection strategy, eliminate the use of recording playback to wake up the mobile phone, improve the security of the mobile phone, as well as the accuracy and intelligence of voice wake-up, thereby greatly improving the user experience.

[0004] In the existing patented technologies, traditional speech recognition systems mainly rely on the waveform characteristics and spectrum characteristics of audio signals for recognition. Although these characteristics can reflect the uniqueness of speech to a certain extent, their recognition effect is often unsatisfactory in the face of complex recording attacks. Recording attackers can change the recording environment, adjust the parameters of the recording equipment, use high-quality recording equipment, etc., so that the recorded speech is highly similar to the real speech in waveform and spectrum, thereby deceiving the traditional speech recognition system. Summary of the invention

[0005] The present application provides a speech recognition wake-up method, device and system for anti-recording detection. By extracting multi-dimensional features, different sounds can be described and distinguished more comprehensively, thereby improving recognition accuracy. Time dimension alignment eliminates the impact of time differences on recognition results, thereby further improving recognition accuracy.

[0006] This application provides a voice recognition wake-up method for anti-recording detection, including:

[0007] S101, collecting current audio and extracting multi-dimensional features in the current audio;

[0008] S102, based on the multi-dimensional features extracted in step S101, setting the time dimension, and aligning the current audio with the historical real speech on the time axis;

[0009] S103, constructing a multi-dimensional feature scatter plot based on the aligned current audio and historical real speech, constructing a scatter ring based on the constructed multi-dimensional feature scatter plot, constructing a scatter ring group based on the constructed scatter ring, collecting historical scatter ring group features to construct a historical scatter ring group feature combination, training the historical scatter ring group feature combination, based on the trained historical scatter ring group feature combination recognition model, and constructing a shearing and splicing recognition model based on the recognition model;

[0010] S104, calculating the comprehensive similarity based on comparing the multi-dimensional scatter plot of the aligned audio generated in step S103 with the multi-dimensional scatter plot of the historical real speech;

[0011] S105, according to the calculated comprehensive similarity, a similarity threshold is set and a determination is made as to whether to record. When the comprehensive similarity value is higher than the set threshold, the current wake-up audio is the real user voice, otherwise, it is a recording.

[0012] Preferably, the step of aligning the current audio with the historical real speech on the time axis is as follows: creating an empty data structure, extracting characteristic sampling points such as timbre, pitch, speaking rate and resonance peak of the aligned audio signal; storing the characteristic value and the corresponding time point of each characteristic sampling point in the data structure, and using a visualization tool to draw the data in the data structure, with the X-axis representing the time dimension and the Y-axis representing the value of the characteristic dimension. For each characteristic dimension, a corresponding scatter plot is constructed, and the generated multiple scatter plots are superimposed and displayed, and the multiple scatter plots are overlapped in the same coordinate system.

[0013] Preferably, the comprehensive similarity = , where Σ represents the summation operation, dimension similarity represents the similarity value calculated on the current dimension, and weight represents the contribution of the dimension in the comprehensive similarity calculation.

[0014] Preferably, the specific steps of forming a scatter ring based on a multi-dimensional feature scatter plot are:

[0015] S201, constructing a scatter ring based on the scatter plot of the multi-dimensional features constructed in step S103;

[0016] S202, constructing a scattered point ring group according to the scattered point ring obtained in step S201;

[0017] S203, calculating the morphological features of the scattered point ring group and the relationship features of adjacent scattered point rings through the constructed scattered point ring group, and calculating the similarity of the scattered point ring group according to the calculated morphological features of the scattered point ring group and the relationship features of adjacent scattered point rings;

[0018] S204, based on the calculated scatter point ring group similarity, set a scatter point ring group similarity threshold to judge the audio.

[0019] Preferably, the scatter ring is a ring structure that represents multiple feature dimensions of the audio in the form of scattered points at a specific time point and is formed by connecting lines. The scatter ring formation method is to use a drawing library to draw and set the coordinate axis: for the X-axis, several scales are set in the circumferential direction, each scale represents a feature dimension, and the circumferential direction indicates that each feature dimension is arranged in a ring manner around a point, and the X-axis forms a circle; for the Y-axis, the Y axes of all dimensions are overlapped, indicating that data points of different feature dimensions are represented in the same vertical space, and the value of the Y-axis represents the size or intensity of each feature dimension; for each time node, the feature dimension is obtained and the feature value is extracted. For each feature dimension, the position of the feature dimension on the X-axis is determined by calculating the angle of each feature dimension, and the scatter points at the same position are connected in sequence using connecting lines to form a scatter ring.

[0020] Preferably, a scatter ring group is constructed by determining the number of scatter rings, the number of scatter rings = total audio duration ÷ time node interval, wherein the total audio duration is the total duration of the current audio collected, and the time node interval is the time interval between sampling points.

[0021] Preferably, the step of forming a new model by a scattered ring group is:

[0022] S301, collecting historical scattered point ring group features to construct historical scattered point ring group feature combinations;

[0023] S302, training the historical scattered point ring group feature combination, based on the trained historical scattered point ring group feature combination recognition model, building a cutting and splicing recognition model based on the recognition model;

[0024] S303, setting up a model performance monitoring mechanism, and optimizing and updating the model according to the monitoring results.

[0025] Preferably, the specific steps of constructing the cutting and splicing recognition model are:

[0026] S401, collecting real cut and spliced ​​recordings and un-spliced ​​real speech, and using the collected real cut and spliced ​​recordings and un-spliced ​​real speech as training data;

[0027] S402, selecting and training a cutting and splicing recognition model based on the training data collected in step S402;

[0028] S403: Recognize the audio based on the trained cutting and splicing recognition model.

[0029] The present application also provides a voice recognition wake-up device that is anti-recording detection, including an audio acquisition module, a feature extraction and processing unit, a scatter point construction and visualization module, a similarity calculation and judgment module, a model recognition and optimization module, and an output and feedback module, wherein the audio acquisition module uses a high-sensitivity microphone array to capture sound signals in the surrounding environment; the feature extraction and processing unit includes an integrated deep neural network processor and a time dimension alignment module, and the time dimension alignment module is used to align the current audio with the historical real voice on the time axis; the scatter point construction and visualization module includes a data structure storage unit; the output and feedback module includes a recognition result output unit and an abnormal event recording and feedback unit.

[0030] The present application also provides a voice recognition wake-up system that is anti-recording detection, including an audio acquisition layer, a data processing layer, a model recognition layer, a decision and output layer, and an optimization and update layer, wherein the audio acquisition layer is deployed on a microphone array on various smart devices; the data processing layer includes a feature extraction module, a time alignment module, and a scatter plot construction module; the model recognition layer includes a comprehensive similarity calculation module, a scatter ring group recognition module, and a cutting and splicing recognition module; the optimization and update layer includes a model performance monitoring module and a model optimization and update module, and the model performance monitoring module monitors the recognition performance of the model in real time.

[0031] One or more technical solutions provided in the present application have at least the following technical effects or advantages: by extracting multi-dimensional features, different sounds can be described and distinguished more comprehensively, improving the accuracy of recognition; time dimension alignment eliminates the influence of time differences on recognition results, further improving the accuracy of recognition; the calculation of the comprehensive similarity value considers the contribution of multiple dimensions and weights them according to their importance, making the recognition result more reliable; by setting the similarity threshold, the real user voice and the recording can be effectively distinguished, reducing the misrecognition rate;

[0032] By forming scatter rings and scatter ring groups, different sound features can be described and distinguished more comprehensively, especially for the recognition of audio with a lengthened or shortened time axis. The calculation of ring group morphological features and adjacent ring relationship features further refines the description of sound features and improves the accuracy of recognition. By setting the ring group similarity threshold and combining the comprehensive similarity of the scatter diagram in Example 1 for judgment, the real user voice and recording attack can be more effectively identified, reducing the misrecognition rate.

[0033] By building a historical scattered ring group feature combination and a scattered ring group recognition model, it is possible to more quickly identify whether the current wake-up audio is real, especially for splicing attacks of multiple recordings, which makes the recognition result more objective and accurate, reduces the subjectivity of human judgment, and by setting the threshold, the accuracy and false recognition rate of recognition can be further controlled to meet the needs of practical applications;

[0034] By constructing a cutting and splicing recognition model, this technical solution can more accurately identify forged recordings formed by cutting and splicing, especially for splicing attacks on specific segments such as unlocking voice. The quantitative relationship of the model recognition process makes the recognition results more objective and accurate, reducing the false recognition rate. By setting the threshold and parameter range, the accuracy and stability of the recognition can be further controlled to meet the needs of practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of the voice recognition wake-up method for anti-recording detection of the present invention;

[0036] Figure 2 A schematic diagram of a process of forming a multi-dimensional scatter point ring based on a multi-dimensional feature scatter point graph according to an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of a process of forming a new model by combining features of a historical scattered point ring group according to an embodiment of the present invention;

[0038] Figure 4 A schematic diagram of a process for constructing a cutting and splicing recognition model according to an embodiment of the present invention;

[0039] Figure 5 It is a structural block diagram of a voice recognition wake-up device for preventing recording detection according to an embodiment of the present invention;

[0040] Figure 6 This is a structural block diagram of a voice recognition wake-up system for preventing recording detection according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] To facilitate the understanding of the present invention, the present application will be described more comprehensively below with reference to the relevant drawings; the drawings show preferred embodiments of the present invention, but the present invention can be implemented in many different forms and is not limited to the embodiments described herein; on the contrary, the purpose of providing these embodiments is to enable a more thorough and comprehensive understanding of the disclosed content of the present invention.

[0042] It should be noted that the terms “vertical”, “horizontal”, “up”, “down”, “left”, “right” and similar expressions used in this document are only for illustrative purposes and do not represent the only implementation method.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which the present invention belongs; the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention; the term "and / or" used herein includes any and all combinations of one or more related listed items.

[0044] Embodiment 1: Figure 1 1 is a flow chart of a voice recognition wake-up method for anti-recording detection according to an embodiment of the present invention, comprising the following steps:

[0045] S101, collecting current audio and extracting multi-dimensional features in the current audio;

[0046] Specifically, the collected audio includes current audio and historical real speech. The multi-dimensional features include timbre, pitch, speaking rate and resonance peaks. Timbre reflects the relative strength and distribution of different frequency components in the sound; pitch is the high or low pitch of the sound, that is, the change in frequency. In speech, the change in pitch is used to distinguish word meanings and express emotions; speaking rate refers to the speed of speaking, which reflects the rate at which the speaker pronounces; resonance peaks are the peaks of the spectrum produced by vocal tract resonance, which reflect the influence of vocal tract shape and size on the sound.

[0047] The original audio signal is preprocessed, and the background noise in the audio is removed using a noise reduction algorithm (bandpass filtering). The audio signal is then divided into multiple segments, and the audio signal is converted from the time domain to the frequency domain using Fourier transform. A deep neural network (DNN) is constructed and trained, and the preprocessed audio signal is used as input. The deep neural network calculates through its own algorithm and outputs feature vectors such as timbre, pitch, speaking rate, and resonance peaks.

[0048] S102, based on the multi-dimensional features extracted in step S101, setting the time dimension, and aligning the current audio with the historical real speech on the time axis;

[0049] Furthermore, the time dimension is set as the X-axis, and the time interval between sampling points is set according to the characteristics of the audio signal, and the audio signal is evenly sampled according to the set time interval; according to the collected historical real speech, the current audio is aligned with the historical real speech on the time axis, and dynamic time warping DTW is selected as the time alignment algorithm. The minimum distance between the current audio sequence and the historical real speech sequence is calculated to find the alignment path.

[0050] S103, constructing a multi-dimensional feature scatter plot based on the aligned current audio and historical real speech, constructing a scatter ring based on the constructed multi-dimensional feature scatter plot, constructing a scatter ring group based on the constructed scatter ring, collecting historical scatter ring group features to construct a historical scatter ring group feature combination, training the historical scatter ring group feature combination, based on the trained historical scatter ring group feature combination recognition model, and constructing a shearing and splicing recognition model based on the recognition model;

[0051] Specifically, an empty data structure is created to store the feature value and corresponding time point of each sampling point. According to the set time interval, the aligned audio signal is traversed in order so that each sampling point is accessed, and the features such as timbre, pitch, speaking rate and resonance peak of the aligned audio signal are extracted. The process of extracting timbre, pitch, speaking rate and resonance peak has been described in detail in step S101 and will not be repeated here. The feature value and corresponding time point of each sampling point are added as data points to the previously initialized data structure to ensure that the order of adding data points is consistent with the time order. The data in the data structure is plotted using the Matplotlib visualization tool, where the X-axis represents the time dimension and the Y-axis represents the value of the feature dimension (timbre, pitch, speaking rate and resonance peak). For each feature dimension, a corresponding scatter plot is constructed, and the generated multiple scatter plots are superimposed and displayed, and the multiple scatter plots are overlapped in the same coordinate system. Points of multiple feature dimensions are displayed in the same space, and the similarities and differences between them can be observed.

[0052] S104, calculating the comprehensive similarity based on comparing the multi-dimensional scatter plot of the aligned audio generated in step S103 with the multi-dimensional scatter plot of the historical real speech;

[0053] Furthermore, by calculating the dot product of the aligned audio (i.e., current speech) and the historical real speech feature vector, and then calculating the modulus (i.e., the length of the vector) of the current wake-up audio and the historical real speech feature vector, the dot product of the two feature vectors is divided by the product of the modulus of the two feature vectors to obtain the cosine value. The closer the calculated cosine value is to 1, the more similar the directions of the two audios in the corresponding dimension are. The linear transformation formula is used to normalize the range of the cosine value similarity from [-1, 1] to [0, 1]. The normalized dimension similarity = (original cosine similarity value + 1) / 2. The conversion of the value range is achieved by shifting the cosine similarity value up by 1 unit and scaling it by half. When the original cosine similarity is -1, the normalized value is 0, and when the original cosine similarity is 1, the normalized value is 1.

[0054] According to expert experience and experimental data, the weights of timbre, pitch, speaking speed and formant are assigned, and the comprehensive similarity is calculated. The comprehensive similarity = , where Σ represents the summation operation, dimension similarity represents the similarity value calculated on the current dimension, and weight represents the contribution of the dimension in the comprehensive similarity calculation.

[0055] In some embodiments, a current wake-up audio (denoted as A) and a historical real speech (denoted as B) are used to calculate the comprehensive similarity between A and B. The comprehensive similarity is based on four dimensions: timbre, pitch, speaking rate and formant. Feature vectors representing timbre, pitch, speaking rate and formant are extracted from audio A and B. Each dimension is represented by a numerical value. For the timbre feature vector, it is represented as A_color = [0.5, 0.8, 0.3], B_color = [0.6, 0.7, 0.4]; for the pitch feature vector, it is represented as A_pitch = [0.2, 0.4, 0.6], B_pitch = [0.3, 0.3, 0.5]; for the speaking rate feature vector, it is represented as A_speed = [0.7, 0.1, 0.9], B_speed = [0.8, 0.2, 0.7]; for the formant feature vector, it is represented as A_formant = [0.1, 0.5, 0.2], B_formant = [0.2, 0.4, 0.3], calculate the cosine similarity based on the feature vector:

[0056] ,in,

[0057] Dot Product =0.5×0.6+0.8×0.7+0.3×0.4=0.3+0.56+0.12=0.98,

[0058] The modulus of the vector A_color ≈0.99,

[0059] The magnitude of the vector B_color ≈1,

[0060] therefore, ≈ ≈0.99. Similarly, the cosine similarities of the other three dimensions are calculated: cos_sim_pitch≈0.95; cos_sim_speed≈0.88; cos_sim_formant≈0.90. The cosine similarities of the four dimensions are normalized to the interval [0,1].

[0061] For the sound: =0.995,

[0062] For the tone: =0.975,

[0063] For speech rate: =0.94,

[0064] For formant: =0.95,

[0065] According to expert experience, weights are assigned to timbre, pitch, speaking speed and formant: the timbre weight is W_color=0.4, W_pitch=0.3, W_speed=0.2, W_formant=0.1. The comprehensive similarity calculated based on the above data is 0.4×0.995+0.3×0.975+0.2×0.94+0.1×0.95=0.9735. The comprehensive similarity value is 0.9735, which is very close to 1, indicating that the comprehensive features of the current wake-up audio A and the historical real voice B in the four dimensions of timbre, pitch, speaking speed and formant are very similar.

[0066] S105, according to the calculated comprehensive similarity, a similarity threshold is set and a determination is made as to whether to record. When the comprehensive similarity value is greater than or equal to the set threshold, the current wake-up audio is the real user voice, otherwise, it is a recording.

[0067] Furthermore, based on the actual measurement data and user needs, the similarity threshold is preset to 0.85, and the calculated comprehensive similarity is compared with the preset similarity threshold of 0.85. When the comprehensive similarity value is greater than or equal to 0.85, the current wake-up audio is considered to be the real user voice, and the device is allowed to wake up or perform corresponding operations. When the comprehensive similarity value is less than 0.85, it is considered to be a recording or an unknown sound. The system refuses to wake up the device or perform the operation, and records the abnormal event for subsequent analysis.

[0068] The technical solutions in the above-mentioned embodiments of the present application have at least the following technical effects or advantages: by extracting multi-dimensional features, different sounds can be described and distinguished more comprehensively, thereby improving the accuracy of recognition; time dimension alignment eliminates the influence of time differences on the recognition results, thereby further improving the accuracy of recognition; the calculation of the comprehensive similarity value takes into account the contribution of multiple dimensions and weights them according to their importance, thereby making the recognition results more reliable; and by setting a similarity threshold, the real user voice and recording can be effectively distinguished, thereby reducing the misrecognition rate.

[0069] Embodiment 2: Based on the scatter plot in Embodiment 1, this embodiment connects the scattered points of the multi-dimensional features at the same time point, overlaps the vertical Y axis of the multi-dimensional features, and forms a ring structure through the connection.

[0070] like Figure 2 As shown in the figure, the specific steps of forming a multi-dimensional scatter ring based on a multi-dimensional feature scatter plot are:

[0071] S201, constructing a scatter ring based on the scatter plot of the multi-dimensional features constructed in step S103;

[0072] Furthermore, the scatter ring is a ring structure formed by connecting lines at a specific time point in which multiple feature dimensions of the audio (such as frequency, amplitude, timbre, etc.) are represented in the form of scattered points. The multi-dimensional features include timbre, pitch, speaking speed, resonance peaks, and other dimensional features. The Matplotlib drawing library is used for drawing, and the coordinate axes are set: for the X-axis, 8 scales are set in the circumferential direction, each scale represents a feature dimension, and the circumferential direction indicates that each feature dimension is arranged in a ring around a point (the point is the point formed by the overlap of the Y axes of the 8 feature dimensions), and the X-axis forms a circle; for the Y-axis , overlap the Y axes of all dimensions, indicating that the data points of different feature dimensions are represented in the same vertical space, and the value of the Y axis represents the size or strength of each feature dimension; use the loop structure for loop to traverse each time node, for each time node, obtain the feature dimension, extract the feature value, for each feature dimension, by calculating the angle of each feature dimension to determine the position of the feature dimension on the X-axis (circle), use connecting lines to connect the scattered points at the same position (angle) in sequence to form a scatter point ring, and ensure that the scatter point ring is closed, that is, the last scatter point should be connected to the first scatter point.

[0073] S202, constructing a scattered point ring group according to the scattered point ring obtained in step S201;

[0074] Specifically, the number of scatter rings is determined according to the total audio duration and the sampling rate. The sampling rate refers to the number of times samples are taken from a continuous signal per second. The time node interval is selected as per second according to actual conditions. The number of scatter rings = total audio duration ÷ time node interval. In one embodiment, the total audio duration is 10 seconds. According to the above formula for calculating the number of scatter rings, the number of scatter rings is 10; each scatter ring is equally spaced, and the interval between the scatter rings is the time node interval; the scatter rings on all time nodes are drawn on the same chart to form a scatter ring group, and the coordinate axis range of the chart is set so that all scatter rings are visible in the chart with appropriate margins, and finally the constructed scatter ring group is saved.

[0075] S203, calculating the morphological features of the scattered point ring group and the relationship features of adjacent scattered point rings through the constructed scattered point ring group, and calculating the similarity of the scattered point ring group according to the calculated morphological features of the scattered point ring group and the relationship features of adjacent scattered point rings;

[0076] Furthermore, the morphological characteristics of the scattered point ring group are calculated by calculating the area and perimeter of the scattered point ring group. The area of ​​the scattered point ring group is calculated using the polygon area calculation formula:

[0077] ,in,

[0078] The summation part of the formula traverses all the edges of the polygon, and each pair of adjacent points and Both calculate the difference of a cross product , the last item It is the difference of the cross product of the edge formed by connecting the last point with the first point. The absolute value of the whole summation result is taken to ensure that the area is a positive number, and finally multiplied by 1 / 2 to get the area of ​​the polygon; for the perimeter of the scattered point ring group, the adjacent scattered points are directly connected, and the total length of the line segment is calculated as the perimeter of the scattered point ring group; the relationship characteristics of adjacent scattered point rings are calculated by calculating the Euclidean distance of adjacent scattered point ring groups. For each adjacent scattered point ring, the Euclidean distance between adjacent scattered point rings is calculated. Suppose two points A on adjacent scattered point rings and B , the Euclidean distance between A and B .

[0079] The similarity of the scattered point ring group is calculated based on the weighted sum of the morphological features of the scattered point ring group and the relationship features of the adjacent scattered point rings. Suppose the morphological feature vector of the scattered point ring group is , the eigenvector of the adjacent scattered point ring relationship is , and their corresponding weights are and , weight and According to expert experience and experimental data, for the weighted sum similarity algorithm, the similarity values ​​of morphological features and adjacent ring relationship features are calculated and recorded as and , based on the calculated and Calculate the similarity S of the scattered ring group. The specific formula is as follows:

[0080] , the similarity of the scattered ring group is calculated according to the formula.

[0081] S204, based on the calculated scatter point ring group similarity, set a scatter point ring group similarity threshold to judge the audio.

[0082] Specifically, a scatter ring group similarity threshold is preset based on the scatter ring group data of historical real speech, and the calculated scatter ring group similarity is compared with the scatter ring group similarity threshold. If the calculated scatter ring group similarity is greater than the scatter ring group similarity threshold, the current audio is considered to be a real user voice. If the calculated scatter ring group similarity is less than or equal to the scatter ring group similarity threshold, a judgment is made in combination with the comprehensive similarity in Example 1, and a weighted calculation is performed. If the calculated value is less than or equal to the scatter ring group similarity threshold, it is not a real user voice.

[0083] The technical solutions in the above-mentioned embodiments of the present application have at least the following technical effects or advantages: by forming scatter rings and scatter ring groups, different sound features can be described and distinguished more comprehensively, especially for the recognition of audio with lengthened or shortened time axis, the calculation of ring group morphological features and adjacent ring relationship features further refines the description of sound features and improves recognition accuracy, and by setting the ring group similarity threshold and combining the comprehensive similarity of the scatter diagram in Embodiment 1 for judgment, real user voices and recorded attacks can be more effectively identified, reducing the misrecognition rate.

[0084] Embodiment 3: Based on the scattered point ring group in Embodiment 2, this embodiment generates a feature combination of historical scattered point ring groups through a model to form a new model, which can quickly identify whether the current wake-up audio is authentic.

[0085] like Figure 3 As shown in the figure, the steps of forming a new model by combining the features of the historical scattered ring group are:

[0086] S301, collecting historical scattered point ring group features to construct historical scattered point ring group feature combinations;

[0087] Specifically, all scattered ring group features extracted from the speech data are summarized and combined, the summarized feature set is optimized, and the features are normalized to eliminate the dimension effect.

[0088] S302, training the historical scattered point ring group feature combination, and identifying the historical scattered point ring group feature combination model based on the trained one;

[0089] Furthermore, the scattered ring group feature combinations of historical real speech are used as positive sample data sets, and the scattered ring group feature combinations of non-real speech (such as splicing of multiple recordings, synthesized speech, etc.) are collected as negative sample data sets, so that the positive and negative sample data sets are consistent in quantity. The model is trained using the machine learning algorithm support vector machine (SVM) and the positive and negative sample data sets. While training the model, the model parameters (learning rate and number of iterations) are adjusted to optimize the performance of the model, and the trained model is saved.

[0090] Preprocess the current wake-up audio. The preprocessing may include denoising, filtering, volume normalization and other operations to ensure the quality and consistency of the audio data. Extract scattered ring group features from the preprocessed audio to obtain a set of features to be identified. Input the set of features to be identified into the trained scattered ring group recognition model. The model quickly identifies the input feature set based on machine learning and makes a preliminary judgment on the authenticity of the audio. After the model recognition is completed, the recognition result is output. The result includes a judgment on whether the current wake-up audio is authentic and a confidence level. The confidence level is used to indicate the degree of certainty of the model on the recognition result. A threshold is set according to the confidence level of the recognition result. The confidence level of the output recognition result is compared with the preset threshold level. If the confidence level of the recognition result exceeds the threshold level, it indicates that the current audio is authentic. If the recognition result does not exceed the threshold level, it indicates that the audio is forged or there are abnormal conditions such as splicing attacks.

[0091] S303, setting a model performance monitoring mechanism, and optimizing and updating the model according to the monitoring results;

[0092] Furthermore, the performance of the model is monitored. Performance monitoring includes key indicators such as accuracy, recall, and F1 score. The monitoring system can automatically detect and report abnormal recognition situations, such as a sudden drop in recognition rate, frequent misidentification in specific scenarios, etc. The feature selection strategy is adjusted according to the identified abnormal situations, and more diverse training data is collected to enhance the generalization ability of the model, and the optimized model replaces the old model.

[0093] The technical solutions in the above-mentioned embodiments of the present application have at least the following technical effects or advantages: by constructing a historical scattered ring group feature combination and a scattered ring group recognition model, it is possible to more quickly identify whether the current wake-up audio is real, especially for splicing attacks of multiple recordings. It has a stronger recognition ability, making the recognition result more objective and accurate, reducing the subjectivity of human judgment, and by setting the threshold, the recognition accuracy and false recognition rate can be further controlled to meet the needs of practical applications.

[0094] Embodiment 4: Based on the recognition model of embodiment 3, in this embodiment, the audio is recognized by using a cutting and splicing recognition model, so that forged recordings formed by cutting and splicing can be recognized.

[0095] like Figure 4 As shown in the figure, the specific steps of constructing the cutting and splicing recognition model are:

[0096] S401, collecting real cut and spliced ​​recordings and un-spliced ​​real speech, and using the collected real cut and spliced ​​recordings and un-spliced ​​real speech as training data;

[0097] Furthermore, real cut and spliced ​​recording materials are obtained from multiple channels, including but not limited to public voice databases, volunteer recordings, public resources on the Internet, etc., and the collected recordings are preliminarily screened. From the screened recording library, recordings containing relevant voice segments are selected according to the needs of the target voice instructions. If the goal is to construct a spliced ​​recording containing the "Xiao Ai classmate" instruction, it is necessary to find a clear voice segment containing this instruction from multiple recordings; for the collection of unspliced ​​real voice, select a variety of recording environments (such as indoors, outdoors, in noisy environments, quiet environments, etc.), use different types of recording equipment (such as mobile phones, professional voice recorders, computer microphones, etc.), obtain voice data of different sound quality and formats, the recorder conducts a natural conversation, or reads aloud according to a preset script, organizes the collected unspliced ​​real voice, and converts the collected recordings into a unified audio format.

[0098] S402, selecting and training a cutting and splicing recognition model based on the training data collected in step S402;

[0099] Specifically, a machine learning algorithm, convolutional neural network (CNN), is selected, and the basic architecture of the model is determined. The basic architecture includes an input layer, a hidden layer, and an output layer. The input layer is responsible for receiving preprocessed audio features (such as MFCC, pitch, etc.); the hidden layer is composed of multiple CNN layers, which is responsible for extracting and representing high-level features of the audio; the output layer outputs classification results (real or fake) or confidence scores according to task requirements, and divides the training data into 70% training set, 20% validation set, and 10% test set. The training set is used to train the model; the validation set is used to evaluate the performance of the model during the training process and adjust the model parameters; the test set is used to finally evaluate the generalization ability of the model. The training set is input into the cut and splice recognition model. After the training is completed, the test set is used to perform a final evaluation of the cut and splice recognition model.

[0100] S403, recognizing the audio based on the trained cutting and splicing recognition model;

[0101] Furthermore, the scattered ring features in the audio are extracted, including the features of the splicing points (such as spectral discontinuity, volume mutation, etc.) and the morphological features of the scatter rings (such as phoneme distribution, intonation changes, etc.), and the extracted features are input into the trained cut-and-splice recognition model. The cut-and-splice recognition model recognizes the audio according to the internal convolutional neural network (CNN) to obtain the recognition result and a confidence level, and a threshold is preset. When the confidence level exceeds the preset threshold, the system determines that the audio is cut and spliced; otherwise, it is judged to be real audio.

[0102] Feature extraction is performed on the audio segment near the splicing point, such as spectral characteristics, time domain characteristics, etc. The similarity between the extracted features is calculated using Euclidean distance to obtain a similarity value. A similarity threshold is preset and the similarity value is compared with the similarity threshold. If the similarity value is higher than the threshold, it is considered that the features at the splicing point are similar and may be part of the same recording; otherwise, it is considered that there are traces of splicing, that is, the audio may be forged or edited.

[0103] The technical solutions in the above-mentioned embodiments of the present application have at least the following technical effects or advantages: by constructing a cutting and splicing recognition model, the technical solution can more accurately identify forged recordings formed by cutting and splicing, especially for splicing attacks on specific segments such as unlocking voice. It has stronger recognition capabilities, and the quantitative relationship of the model recognition process makes the recognition results more objective and accurate, reducing the false recognition rate. By setting the threshold and parameter range, the accuracy and stability of the recognition can be further controlled to meet the needs of practical applications.

[0104] Embodiment 5: Figure 5 As shown, this embodiment provides a voice recognition wake-up device for anti-recording detection, which is used to execute the above-mentioned voice recognition wake-up method for anti-recording detection. The anti-recording detection device includes an audio acquisition module, a feature extraction and processing unit, a scatter point construction and visualization module, a similarity calculation and judgment module, a model recognition and optimization module, and an output and feedback module, wherein the audio acquisition module uses a high-sensitivity microphone array to capture sound signals in the surrounding environment; the feature extraction and processing unit includes an integrated deep neural network processor and a time dimension alignment module, and the time dimension alignment module is used to align the current audio with the historical real voice on the time axis; the scatter point construction and visualization module includes a data structure storage unit, and the data structure storage unit is used to store the feature value of each sampling point and the corresponding time point; the output and feedback module includes a recognition result output unit and an abnormal event recording and feedback unit, and the abnormal event recording and feedback unit is used to record abnormal events in the recognition process for subsequent analysis and optimization.

[0105] Example 6: Figure 6As shown, this embodiment provides a voice recognition wake-up system for anti-recording detection, including an audio collection layer, a data processing layer, a model recognition layer, a decision and output layer, and an optimization and update layer, wherein the audio collection layer is deployed on a microphone array of various smart devices (such as smart phones, smart home devices, smart security devices, etc.), which is responsible for collecting user voices and surrounding environmental sounds; the data processing layer includes a feature extraction module, a time alignment module, and a scatter plot construction module. The feature extraction module receives the audio data transmitted by the audio collection layer, performs preprocessing and multi-dimensional feature extraction, and the time alignment module uses the DTW algorithm to achieve the time dimension alignment of the current audio and the historical real voice. The scatter plot construction module constructs a multi-dimensional scatter plot and a scatter ring group; the model recognition layer includes a comprehensive similarity calculation module, a scatter ring group recognition module, and a cutting and splicing recognition module. The comprehensive similarity calculation module calculates the current audio based on cosine similarity and weight distribution. The comprehensive similarity with the historical real speech, the scattered ring group recognition module uses the trained scattered ring group recognition model to quickly recognize the audio, and the cutting and splicing recognition module integrates the cutting and splicing recognition model to specifically identify forged recordings; the decision and output layer includes a similarity judgment module, a recognition result output module and an abnormal event processing module. The similarity judgment module judges the calculated similarity according to the preset similarity threshold, and the recognition result output module outputs the recognition result to the user or related system in an intuitive manner. The abnormal event processing module records and processes abnormal events in the recognition process, such as recording attacks, equipment failures, etc.; the optimization and update layer includes a model performance monitoring module and a model optimization and update module. The model performance monitoring module monitors the recognition performance of the model in real time, including key indicators such as accuracy and recall rate. The model optimization and update module optimizes and updates the model according to the monitoring results to ensure long-term stable operation.

[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A voice recognition wake-up method for anti-recording detection, characterized in that: include: S101, collecting current audio and extracting multi-dimensional features in the current audio; S102, based on the multi-dimensional features extracted in step S101, setting the time dimension, and aligning the current audio with the historical real speech on the time axis; S103, constructing a multi-dimensional feature scatter plot based on the aligned current audio and historical real speech, constructing a scatter ring based on the constructed multi-dimensional feature scatter plot, constructing a scatter ring group based on the constructed scatter ring, collecting historical scatter ring group features to construct a historical scatter ring group feature combination, training the historical scatter ring group feature combination, based on the trained historical scatter ring group feature combination recognition model, and constructing a shearing and splicing recognition model based on the recognition model; S104, calculating the comprehensive similarity based on comparing the multi-dimensional scatter plot of the aligned audio generated in step S103 with the multi-dimensional scatter plot of the historical real speech; S105, according to the calculated comprehensive similarity, a similarity threshold is set and a determination is made as to whether to record. When the comprehensive similarity value is higher than the set threshold, the current wake-up audio is the real user voice, otherwise, it is a recording.

2. The voice recognition wake-up method for anti-recording detection according to claim 1, characterized in that: The steps to align the current audio with the historical real speech on the time axis are as follows: create an empty data structure, extract the timbre, pitch, speaking speed and resonance peak feature sampling points of the aligned audio signal; store the feature value and the corresponding time point of each feature sampling point in the data structure, and use a visualization tool to draw the data in the data structure, with the X-axis representing the time dimension and the Y-axis representing the value of the feature dimension. For each feature dimension, construct a corresponding scatter plot, overlay and display the generated multiple scatter plots, and overlap the multiple scatter plots in the same coordinate system.

3. The voice recognition wake-up method for anti-recording detection according to claim 1, characterized in that: Comprehensive similarity = , where Σ represents the summation operation, dimension similarity represents the similarity value calculated on the current dimension, and weight represents the contribution of the dimension in the comprehensive similarity calculation.

4. The voice recognition wake-up method for anti-recording detection according to claim 1, characterized in that: The specific steps to form a scatter ring based on a multi-dimensional feature scatter plot are: S201, constructing a scatter ring based on the scatter plot of the multi-dimensional features constructed in step S103; S202, constructing a scattered point ring group according to the scattered point ring obtained in step S201; S203, calculating the morphological features of the scattered point ring group and the relationship features of adjacent scattered point rings through the constructed scattered point ring group, and calculating the similarity of the scattered point ring group according to the calculated morphological features of the scattered point ring group and the relationship features of adjacent scattered point rings; S204, based on the calculated scatter point ring group similarity, set a scatter point ring group similarity threshold to judge the audio.

5. The voice recognition wake-up method for anti-recording detection according to claim 4, characterized in that: The scatter ring is a ring structure that represents multiple feature dimensions of audio in the form of scattered points at a specific time point and is formed by connecting lines. The scatter ring is constructed by using a drawing library to draw and set the coordinate axis: for the X-axis, several scales are set in the circumferential direction, each scale represents a feature dimension, and the circumferential direction indicates that each feature dimension is arranged in a ring around a point, and the X-axis forms a circle; for the Y-axis, the Y axes of all dimensions are overlapped, indicating that data points of different feature dimensions are represented in the same vertical space, and the value of the Y-axis represents the size or intensity of each feature dimension; for each time node, the feature dimension is obtained and the feature value is extracted. For each feature dimension, the position of the feature dimension on the X-axis is determined by calculating the angle of each feature dimension, and the scattered points at the same position are connected in sequence using connecting lines to form a scatter ring.

6. The voice recognition wake-up method for anti-recording detection according to claim 4, characterized in that: A scatter point ring group is constructed by determining the number of scatter point rings, where the number of scatter point rings = total audio duration ÷ time node interval, where the total audio duration is the total duration of the current audio collected, and the time node interval is the time interval between sampling points.

7. The voice recognition wake-up method for anti-recording detection according to claim 4, characterized in that: Steps to form a new model through scattered ring groups: S301, collecting historical scattered point ring group features to construct historical scattered point ring group feature combinations; S302, training the historical scattered point ring group feature combination, based on the trained historical scattered point ring group feature combination recognition model, building a cutting and splicing recognition model based on the recognition model; S303, setting up a model performance monitoring mechanism, and optimizing and updating the model according to the monitoring results.

8. The voice recognition wake-up method for anti-recording detection according to claim 7, characterized in that: Specific steps for building a cutting and splicing recognition model: S401, collecting real cut and spliced ​​recordings and un-spliced ​​real speech, and using the collected real cut and spliced ​​recordings and un-spliced ​​real speech as training data; S402, selecting and training a cutting and splicing recognition model based on the training data collected in step S402; S403: Recognize the audio based on the trained cutting and splicing recognition model.

9. A voice recognition wake-up device for preventing recording detection, applied to the voice recognition wake-up method for preventing recording detection as claimed in any one of claims 1 to 8, characterized in that: It includes an audio acquisition module, a feature extraction and processing unit, a scatter point construction and visualization module, a similarity calculation and judgment module, a model recognition and optimization module, and an output and feedback module. The audio acquisition module uses a high-sensitivity microphone array to capture sound signals in the surrounding environment; the feature extraction and processing unit includes an integrated deep neural network processor and a time dimension alignment module, and the time dimension alignment module is used to align the current audio with the historical real speech on the time axis; the scatter point construction and visualization module includes a data structure storage unit; the output and feedback module includes a recognition result output unit and an abnormal event recording and feedback unit.

10. A voice recognition wake-up system for preventing recording detection, applied to the voice recognition wake-up method for preventing recording detection as claimed in any one of claims 1 to 8, characterized in that: It includes audio collection layer, data processing layer, model recognition layer, decision and output layer, and optimization and update layer. Among them, the audio collection layer is deployed on the microphone array of various smart devices; the data processing layer includes feature extraction module, time alignment module and scatter plot construction module; the model recognition layer includes comprehensive similarity calculation module, scatter ring group recognition module and cutting and splicing recognition module; the optimization and update layer includes model performance monitoring module and model optimization and update module. The model performance monitoring module monitors the recognition performance of the model in real time.

Images