Signal detection method, electronic device, storage medium and computer program product

Abstract

This application discloses a signal detection method, electronic device, storage medium, and computer program product. The method includes: receiving a speech signal to be detected; analyzing the speech signal to be detected to obtain an analysis result, the analysis result representing a first application scenario to which the speech signal to be detected belongs; extracting features from the speech signal to be detected to obtain a first feature vector, the first feature vector containing at least two-dimensional feature vectors; transforming the first feature vector based on the analysis result to obtain a second feature vector; and obtaining a detection result based on a first model and the second feature vector, the first model being used to detect the authenticity of the speech signal, and the detection result representing whether the speech signal to be detected is forged speech. The technical solution provided by this application optimizes the original multi-dimensional feature vector based on the recognition result of the application scenario, thereby improving the detection accuracy of speech signals.

Description

Technical Field

[0001] This application relates to the field of signal security technology, and in particular to a signal detection method, electronic device, storage medium, and computer program product. Background Technology

[0002] With the rapid development of speech synthesis and deep learning, speech fabrication detection technology has begun to receive widespread attention. Related technologies have proposed methods that rely on the spectral or speaker characteristics of speech signals to detect whether a speech signal is forged. This involves analyzing the frequency domain features of the extracted speech signal through Short-Time Fourier Transform (STFT) spectrogram analysis to classify these features. Furthermore, with the popularization of deep learning technology, related technologies have also proposed using Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs) to analyze the temporal and spectral characteristics of speech signals to detect whether a speech signal is forged.

[0003] However, in the face of complex and ever-changing acoustic environments, the relevant detection solutions suffer from low detection accuracy. Summary of the Invention

[0004] To address the related technical problems, embodiments of this application provide a signal detection method, an electronic device, a storage medium, and a computer program product.

[0005] The technical solution of this application embodiment is implemented as follows: This application provides a signal detection method applied to an electronic device, including: Receive the voice signal to be detected; The speech signal to be detected is analyzed to obtain an analysis result, and the analysis result represents the first application scenario to which the speech signal to be detected belongs; The speech signal to be detected is subjected to feature extraction to obtain a first feature vector, which contains feature vectors of at least two dimensions. Based on the analysis results, the first feature vector is transformed to obtain the second feature vector; Based on the first model and the second feature vector, a detection result is obtained. The first model is used to detect the authenticity of the speech signal, and the detection result indicates whether the speech signal to be detected is fake speech.

[0006] This application also provides an electronic device, including: a processor and a memory for storing a computer program that can run on the processor, wherein when the processor runs the computer program, it executes the steps of any of the above signal detection methods.

[0007] This application embodiment also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of any of the above signal detection methods.

[0008] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described signal detection methods.

[0009] The signal detection method, electronic device, storage medium, and computer program product provided in this application embodiment receive a speech signal to be detected; analyze the speech signal to be detected to obtain an analysis result, the analysis result representing a first application scenario to which the speech signal to be detected belongs; extract features from the speech signal to be detected to obtain a first feature vector, the first feature vector containing feature vectors of at least two dimensions; transform the first feature vector based on the analysis result to obtain a second feature vector; and obtain a detection result based on a first model and the second feature vector, the first model being used to detect the authenticity of the speech signal, the detection result representing whether the speech signal to be detected is fake speech. The technical solution provided in this application embodiment, after identifying the application scenario of the speech signal, optimizes the original multi-dimensional feature vector based on the identification result of the application scenario, so that the optimized multi-dimensional feature vector can adapt to the current application scenario. Thus, in complex and changing acoustic scenarios, the detection accuracy of speech signals using the model can be improved. Attached Figure Description

[0010] Figure 1 This is a schematic flowchart of a signal detection method according to an embodiment of this application; Figure 2 This is a schematic diagram of the detection process of the first model in the embodiments of this application; Figure 3 This is a schematic diagram illustrating a process for detecting fake voice as an application example of this application; Figure 4 This is a schematic diagram of the structure of a signal detection device according to an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0011] The present application will now be described in further detail with reference to the accompanying drawings and embodiments.

[0012] In related technologies, traditional detection schemes rely on the spectral or speaker characteristics of speech signals and machine learning models (such as Support Vector Machines (SVM)) to classify the frequency domain features of speech signals, thereby achieving speech signal detection. Deep learning-based detection schemes, on the other hand, rely on CNNs or RNNs to analyze the temporal and spectral characteristics of speech signals, thereby achieving speech signal detection.

[0013] However, the above detection scheme has the following problems: 1. Low detection performance: Due to the use of a fixed detection model, it cannot effectively distinguish between signal distortion and spoofing features in complex acoustic environments (such as telephone channels containing channel compression and background noise), resulting in low detection accuracy. 2. Insufficient generalization ability or robustness: Due to the single or incomplete dimension of feature analysis, the generalization ability to different spoofed speech signals is insufficient, that is, the detection accuracy of speech signals is low.

[0014] Based on this, in various embodiments of this application, the multi-dimensional feature vector of the speech signal is optimized according to the application scenario to which the speech signal belongs. This is done to enhance the key feature vectors associated with authenticity while suppressing interference caused by external environmental factors, thereby obtaining optimized feature vectors. In this way, under varying acoustic scenarios, it is beneficial to improve the accuracy of subsequent models in detecting the authenticity of speech, ensuring the robustness and generalization ability of the model, and enabling the model to have high stability and accuracy in different application scenarios.

[0015] This application provides a signal detection method applied to electronic devices, such as... Figure 1 As shown, the method includes: Step 101: Receive the voice signal to be detected; Step 102: Analyze the speech signal to be detected to obtain the analysis result, wherein the analysis result characterizes the first application scenario to which the speech signal to be detected belongs; Step 103: Extract features from the speech signal to be detected to obtain a first feature vector, wherein the first feature vector contains feature vectors of at least two dimensions; Step 104: Based on the analysis results, transform the first feature vector to obtain the second feature vector; Step 105: Based on the first model and the second feature vector, a detection result is obtained. The first model is used to detect the authenticity of the speech signal, and the detection result indicates whether the speech signal to be detected is fake speech.

[0016] In practical applications, the electronic device may include a device capable of performing steps 101 to 105, such as a server, laptop, tablet, or desktop computer. This application does not limit the type of electronic device, as long as its function is implemented. Furthermore, the first model may be called a speech detection model, a fake speech detection model, or a deep learning-based fake speech detection model. The first model may include a model with a Transformer architecture at its core, capable of efficiently and accurately detecting the authenticity of speech signals, i.e., detecting whether the speech signal is fake speech. This application does not limit the type of the first model, as long as its function is implemented.

[0017] In practical applications, in step 102, the electronic device can comprehensively analyze the voice signal to be detected from different angles in order to make an accurate judgment on the first application scenario.

[0018] Specifically, in some optional embodiments, the analysis of the speech signal to be detected to obtain the analysis result includes: The speech signal to be detected is analyzed to obtain a first analysis result, which characterizes the channel characteristics of the speech signal to be detected during transmission. The speech signal to be detected is analyzed to obtain a second analysis result, which characterizes the degree to which the interaction mode of the speech signal to be detected conforms to the interaction mode of one or more application scenarios. The analysis results are obtained based on the first analysis results and the second analysis results.

[0019] In practical applications, the first application scenario may include telephone communication, voice payment, etc., and this application embodiment does not limit this. Furthermore, the first analysis result can be understood as a channel fingerprint score of the voice signal to be detected, which at least reflects the characteristics generated by the voice signal to be detected during transmission due to the influence of device or channel characteristics; the second analysis result can be understood as a conversation dynamics score of the voice signal to be detected, which at least reflects the degree to which the voice interaction of the voice signal to be detected conforms to the patterns of different application scenarios.

[0020] In practical applications, based on the first analysis result, the electronic device can perform channel fingerprinting through energy detection and artifact detection to obtain the first analysis result.

[0021] Specifically, in some optional embodiments, the analysis of the speech signal to be detected to obtain a first analysis result includes: Energy distribution detection is performed on the speech signal to be detected to obtain a third analysis result, which characterizes whether the speech signal to be detected is a narrowband signal. The speech signal to be detected is subjected to vocoder detection to obtain a fourth analysis result, which represents the probability of the presence of the first vocoder. The third and fourth analysis results are fused together to obtain the first analysis result.

[0022] In practical applications, the third analysis result can be understood as the detection result of narrowband energy distribution. This third analysis result contains Boolean values ​​and reflects whether the detected speech signal is a narrowband signal. The fourth analysis result can be understood as the artifact detection result of the vocoder (which can be expressed as a vocoder in English), reflecting the probability of the existence of a specific first vocoder (which can also be understood as a codec). The corresponding value range can be [0, 1] (which can be represented as...). ).

[0023] In practical applications, the electronic device can perform energy distribution detection on the speech signal to be detected, and obtain the corresponding long-term average power spectral density (which can be expressed as...). Based on the long-term average power spectral density and the first threshold (which can be expressed as...). The third analysis result is obtained by using the following formula: The first threshold can be understood as a pre-configured in-band / out-of-band energy ratio threshold (e.g., 40 dB). (1) It should be noted that if the long-term average power spectral density satisfies the above formula, the speech signal to be detected is determined to be a narrowband signal, that is, the third analysis result can be true, and the corresponding value is 1; if the long-term average power spectral density does not satisfy the above formula, the speech signal to be detected is determined to be a non-narrowband signal, that is, the third analysis result can be false, and the corresponding value is 0.

[0024] In practical applications, the electronic device can also convert the speech signal to be detected into a Mel spectrogram, and then use a lightweight neural network (such as a CNN) to perform vocoder detection on the speech signal to obtain the fourth analysis result. In this way, it is possible to identify periodic holes or regular noise in the spectrum generated by a specific vocoder during signal processing, thereby determining whether the first vocoder exists.

[0025] In practical applications, after obtaining the third and fourth analysis results, the electronic device can fuse the third and fourth analysis results into the first analysis result; wherein, the first analysis result can be obtained through the following formula: (2) in, This is represented as the first analysis result, namely the channel fingerprint score; This is represented as the third analysis result; This is represented as the fourth analysis result; This represents the preset weights of the third analysis results. This represents the preset weight of the fourth analysis result, and + =1.

[0026] In practical applications, based on the second analysis result, the electronic device can use a Voice Activity Detection (VAD) algorithm to segment the speech stream to which the speech signal to be detected belongs into speech segments and silence segments; count the number of speaker rotations per unit time in the speech segments (which can be understood as the number of times different subjects take turns speaking), and count the average silence duration in the silence segments; by normalizing the number of speaker rotations and the average silence duration per unit time, the second analysis result is obtained (which can be expressed as, and the corresponding value range can be [0, 1]).

[0027] In practical applications, after obtaining the first analysis result and the second analysis result, the electronic device can perform a fusion process on the first analysis result and the second analysis result to obtain a fused logical value; wherein, the fused logical value can be obtained through the following formula: (3) in, Represented as the merged logical value; This represents the weight of the first analysis result, and the value can be set according to the importance of the feature in the scene judgment; This represents the weight of the second analysis result. The value can be set according to the importance of the feature in the scene judgment. The value is greater than The value.

[0028] In practical applications, to obtain the analysis results, the electronic device can activate the fused logical values ​​using a relevant function (such as the Sigmoid function) to obtain the analysis results. These results reflect the probability that the detected speech signal belongs to one or more application scenarios. The analysis results can be obtained using the following formula (which can be expressed as follows): ): (4) It should be noted that, for the one or more application scenarios, a corresponding second threshold can be preset to determine the first application scenario to which the voice signal to be detected belongs; wherein, the second threshold corresponding to different application scenarios can be different.

[0029] For example, in a scenario where the application is telephone communication, the corresponding second threshold can be set to 0.8. In this case, the electronic device can compare the probability of telephone communication with the corresponding second threshold. If the probability of telephone communication (which can be expressed as...) If the probability of the detected voice signal is greater than the corresponding second threshold, then the first application scenario to which the detected voice signal belongs is considered to be telephone communication. In the scenario of secure payment, the corresponding second threshold can be set to 0.9. In this case, the electronic device can compare the probability of telephone communication with the corresponding second threshold. If the probability of telephone communication (which can be expressed as...) is greater than the corresponding second threshold, then the detected voice signal belongs to the first application scenario of telephone communication. If the value is greater than the corresponding second threshold, then the first application scenario to which the voice signal to be detected belongs is considered to be secure payment.

[0030] In this embodiment of the application, the application scenario of the speech signal is judged to quickly and accurately identify the application scenario to which the speech signal belongs, so as to lay the foundation for subsequent feature vector processing.

[0031] In practical applications, to counter diverse voice spoofing techniques, the electronic device can deeply mine the features of the voice signal to be detected from two orthogonal dimensions: physiological characteristics and signal structure, to construct a complete and complementary feature set. This dual-track parallel feature extraction method enables the electronic device to cross-verify the spoofing traces of the voice signal from different levels, thereby forming a more robust detection foundation.

[0032] Specifically, in some optional embodiments, the step of extracting features from the speech signal to be detected to obtain a first feature vector includes: Feature extraction is performed on the speech signal to be detected to obtain a third feature vector, which represents the defects of the speech signal to be detected in the physiological feature dimension. Feature extraction is performed on the speech signal to be detected to obtain a fourth feature vector, which represents the defects of the speech signal to be detected in the signal structure dimension. The first feature vector is obtained based on the third feature vector and the fourth feature vector.

[0033] In practical applications, the first feature vector can be understood as a multi-dimensional feature vector (e.g., 120-dimensional) of the speech signal to be detected. The first feature vector contains at least a third feature vector of physiological feature dimension and a fourth feature vector of signal structure dimension. The third feature vector can reflect the physical randomness of the speech signal to be detected when simulating human vocal organs, and the fourth feature vector can reflect the mathematical structural artifacts generated by the speech signal to be detected when reconstructing the signal waveform.

[0034] In practical applications, the electronic device can incorporate the analysis of micro-jitter changes based on the physiological characteristics of natural language to identify minute irregular fluctuations, thereby obtaining the third feature vector. This is because micro-jitter changes are determined by the naturalness of physiological vibrations and exhibit a high degree of randomness and non-linear distribution in real speech, while forged speech usually relies on algorithm generation and lacks natural physiological perturbations (which can also be understood as the degree of micro-jitter changes being lower or the changes tending to be uniform).

[0035] In practical applications, to identify the aforementioned differences, the electronic device can determine the adjacent fundamental frequency periods corresponding to the speech signal to be detected using the zero-crossing point method or the autocorrelation method; extract the fundamental frequency jitter features and a first statistic (such as variance, maximum or minimum values ​​associated with the fundamental frequency jitter features) of the adjacent fundamental frequency periods to quantify artificial defects in the fundamental frequency modulation. The fundamental frequency period can be understood as the time required for the vocal organ (such as the vocal cords) to complete one vibration; that is, the third feature vector can include the fundamental frequency jitter features (which can be expressed as jitter in English) and the first statistic. The fundamental frequency jitter features can be understood as the time difference features of adjacent fundamental frequency periods (such as the average value of the duration difference), which can reflect the irregular fluctuations in the period length of the speech signal to be detected; wherein, the fundamental frequency jitter features of adjacent fundamental frequency periods can be extracted using the following formula: (5) in, This is represented as the fundamental frequency jitter characteristic. The larger the value, the more unstable the fundamental frequency (which can also be understood as pitch) of the speech signal to be detected is at the microscopic level. It is represented as the total number of fundamental frequency cycles within a fixed-duration analysis window (such as a 20ms or 30ms speech stream); Represented as a counting variable, with a value range of [1, ..., ...]. -1]; Represented as the first The duration or length of each fundamental frequency period, the reciprocal of which is the fundamental frequency (can be expressed as ); Represented as the first The first fundamental frequency cycle after the +1 fundamental frequency cycle duration or length.

[0036] In practical applications, the electronic device can also extract amplitude jitter features and a second statistic (such as variance, maximum or minimum values ​​associated with amplitude jitter features) from adjacent fundamental frequency cycles to identify regular defects in the dynamic amplitude changes of the speech signal to be detected. That is, the third feature vector can also include amplitude jitter features (which can be expressed as "shimmer") and the second statistic. The amplitude jitter features can be understood as the amplitude difference features of adjacent fundamental frequency cycles (i.e., the average amplitude difference), which can reflect the natural fluctuation of the amplitude of the speech signal to be detected over time. The amplitude jitter features of adjacent fundamental frequency cycles can be calculated using the following formula: (6) in, This is represented as amplitude jitter characteristic. The larger the value, the more unstable the amplitude (or loudness) of the speech signal to be detected is at the microscopic level. Represented as the first The amplitude of the fundamental frequency cycle is usually the amplitude of the first cycle. The peak value (i.e., maximum value) of the signal waveform within one fundamental frequency cycle. Represented as the first The first fundamental frequency cycle after the Amplitude of +1 fundamental frequency cycle.

[0037] In practical applications, because the modeling process of fake speech focuses more on fitting the spectral amplitude (which can be expressed as Magnitude Spectrum) that the human ear is more sensitive to, while ignoring the phase spectrum (which can be expressed as Phase Spectrum) of the signal structure and waveform morphology, the underlying signal structure of fake speech will have unnatural regularity. Therefore, the electronic device can quantify the features of the speech signal to be detected in the signal dimension to obtain the fourth feature vector.

[0038] Specifically, in some optional embodiments, the step of extracting features from the speech signal to be detected to obtain a fourth feature vector includes: The speech signal to be detected is transformed to obtain the transformed speech signal; Based on the transformed speech signal, a sixth analysis result is obtained, which includes the instantaneous phase and instantaneous frequency of the transformed speech signal. Based on the sixth analysis result, the fourth feature vector is obtained.

[0039] In practical applications, the electronic device can perform Hilbert Transform processing on the speech signal to be detected to obtain the transformed speech signal (which can be represented as...). Based on the transformed speech signal and the speech signal to be detected, an analytical signal is constructed (which can be represented as...). , (represented as an imaginary unit); based on the analytic signal, the sixth analysis result is obtained, which may include the instantaneous phase corresponding to the analytic signal (which may be represented as...). ) and instantaneous frequency (which can be expressed as Based on the sixth analysis result, the fourth feature vector is obtained. The fourth feature vector may include one or more of the following (or at least one): the standard deviation of the instantaneous frequency associated with the speech signal to be detected, the modified group delay feature (MODGDF), and the relative phase shift.

[0040] For example, the electronic device can calculate the standard deviation of the instantaneous frequency based on the instantaneous frequency in the sixth analysis result. The standard deviation of the instantaneous frequency can reflect the defects of the detected speech signal in the physiological feature dimension. This is because the fluctuation of the instantaneous frequency of real speech is relatively smooth, while fake speech will produce unnatural, violent jitters or overly flat instantaneous frequencies, resulting in abnormal standard deviation of the instantaneous frequency.

[0041] For example, the electronic device can calculate the average phase difference between specific frequency bands (such as two adjacent harmonic frequency bands) based on the instantaneous gastrointestinal data in the sixth analysis result, thereby obtaining the relative phase shift. The relative phase shift can reflect the defects of the detected speech signal in the physiological feature dimension. This is because the phase relationship between the harmonic frequency bands of real speech usually follows certain physical laws, while forged speech usually disrupts these physical laws, resulting in an abnormal distribution of the relative phase shift.

[0042] For example, the electronic device can calculate the MODGDF from the cepstral smoothed amplitude spectrum of the speech signal to be detected, thereby enhancing the robustness of the features of the speech signal to be detected and identifying smooth or disordered features in the signal dimension; wherein, the group delay can be understood as the negative derivative of the phase with respect to the angular frequency, which can reflect the time delay of different frequency components. For the speech signal to be detected (represented as...) MODGDF can be represented as Among them, the MODGDF of real speech has a clear peak structure near the formant.

[0043] In practical applications, the electronic device can also extract features from the speech signal to be detected to obtain a fifth feature vector. The fifth feature vector represents the defects or anomalies of the speech signal to be detected in the frequency domain space (such as being too smooth or lacking high-frequency details in real speech). The fifth feature vector can include the spectral features of the speech signal to be detected; that is, the first feature vector can also include the fifth feature vector.

[0044] For example, the electronic device can use an STFT to convert the speech signal to be detected from the time domain to the frequency domain to obtain a Mel spectrogram. Then, the Mel spectrogram is mapped onto the Mel frequency dimension that the human ear can perceive in order to perform feature extraction, thereby obtaining the fifth feature vector.

[0045] In practical applications, since real speech is naturally generated by the vocal organs, the corresponding temporal features are random and natural, while forged speech will have consistency or lack dynamic changes. Therefore, the electronic device can also extract features from the speech signal to be detected to obtain a sixth feature vector. The sixth feature vector represents the defects or anomalies of the speech signal to be detected in the temporal space; that is, the first feature vector can also include the sixth feature vector.

[0046] For example, the electronic device can extract the formant trajectory of the speech signal to be detected using Linear Predictive Coding (LPC). The formant trajectory reflects the variation of the formants over time; where the formants are the main frequency components of the speech signal to be detected, formed by the vocal tract. Additionally, the electronic device also extracts the energy distribution and zero-crossing rate of the speech signal to be detected. The energy distribution reflects the intensity variation of the speech signal to be detected, and the zero-crossing rate reflects the waveform oscillation characteristics of the speech signal to be detected. Based on the formant trajectory, energy distribution, and zero-crossing rate, the sixth feature vector is obtained.

[0047] In this embodiment, to achieve deep recognition of forged speech, feature analysis of the signal structure dimension is introduced as a new dimension for forgery traces, in addition to physical dimension feature analysis of the speech signal. Since the features of the signal structure dimension can reveal the unnaturalness in the temporal structure of the signal introduced by the speech synthesis algorithm, electronic devices can accurately identify the features of forged speech generated by different technical principles, thereby gaining generalization ability against unknown types of attacks and helping to improve the subsequent detection accuracy.

[0048] In practical applications, after obtaining the analysis results and the first feature vector, the electronic device can intelligently transform the first feature vector based on the analysis results to optimize it. This can suppress interference introduced by environmental noise and channel distortion in the first feature vector and enhance key features related to forged speech.

[0049] Specifically, in some optional embodiments, the transformation of the first feature vector based on the analysis results to obtain the second feature vector includes: Based on the analysis results, a first weight is obtained, which represents the importance of the first feature vector in the first application scenario; Based on the analysis results, a second weight is obtained, which includes the weight of the transformation network to which the first feature vector belongs; Based on the first weight and the second weight, the first feature vector is subjected to a nonlinear transformation to obtain the second feature vector.

[0050] In practical applications, the first weight can be called the scenario condition weight (which can be expressed as...). The second weight can be understood as a transformation network specific to the first feature vector (which can be represented as...). The weights of a two-layer perceptron (such as a two-layer perceptron) can be expressed as In this application, the names of the first weight and the second weight are not limited, as long as their functions are implemented.

[0051] In practical applications, during the scene embedding stage, the electronic device can use the embedding layer of the Scene Conditional Adaptive Feature Transform Network (SCAFT-Net) to map the discrete identifier information (such as ID) corresponding to the first application scene into a low-dimensional and dense scene vector (which can be represented as...). The scene vector can be understood as the mathematical representation of the first application scenario; the scene vector can be obtained by mapping through a pre-configured table in the embedding layer.

[0052] In practical applications, during the scene condition modulation stage, the electronic device can group the first feature vector based on the scene vector to obtain the first weight.

[0053] Specifically, in some optional embodiments, obtaining the first weight based on the analysis results includes: The first feature vector is grouped to obtain at least two feature groups; Based on the analysis results and the at least two feature groups, at least two fifth analysis results are generated, each fifth analysis result representing the attention score of the corresponding feature group in the first application scenario; The first weight is obtained based on the at least two fifth analysis results.

[0054] In practical applications, each of the at least two feature groups contains a feature vector of one dimension. That is, in the first feature vector (which can be represented as...) In the case of a feature vector containing two dimensions, it can be divided into two feature groups according to physical meaning: the feature group corresponding to the physiological feature dimension (which can be represented as...). The feature set corresponding to the signal structure dimension (which can be represented as) and the signal structure dimension. When the first feature vector contains feature vectors of four dimensions, it can be divided into four feature groups according to physical meaning: the feature group corresponding to the physiological feature dimension, the feature group corresponding to the signal structure dimension, and the feature group corresponding to the frequency domain dimension (which can be represented as...). ) and the feature groups corresponding to the time domain dimension (which can be represented as ).

[0055] In practical applications, for each feature group (which can be represented as...) The electronic device can use the attention mechanism of the scene-condition adaptive feature transformation network to perform average pooling on the feature group to obtain the corresponding summary vector representation (which can be represented as...). The summary vector representation reflects the lowest scalar score of the feature group; after concatenating the summary vector representation and the scene vector, the concatenated result (which can be represented as...) The input is fed into the FNN to obtain a fifth analysis result (which can be represented as...). , The first weight is obtained by normalizing the at least two fifth analysis results.

[0056] In practical applications, for the second weight, the electronic device can introduce a hypernetwork to assign a dedicated transformation network (also called a transformation function) and corresponding second weights and biases (which can be expressed as...) to the at least two feature groups. This allows for subsequent nonlinear deep reshaping of the first feature vector based on the first weight, the second weight, and the bias.

[0057] For example, the electronic device can generate a second weight and bias of a dedicated transformation network for each feature group through a hypernetwork; based on the second weight and bias, a customized nonlinear transformation process is performed on the feature group to obtain the processed feature group (which can be represented as...). Based on the first weight, modulation is applied to the processed feature set (which can also be understood as feature set selection) to obtain the final feature set (which can be represented as...). By concatenating the at least two feature groups, the electronic device can obtain an optimized second feature vector for the first application scenario.

[0058] In this embodiment, based on the scene analysis results, the original multidimensional feature vector is subjected to a nonlinear transformation to suppress interference introduced by environmental noise and channel distortion, while enhancing the feature recognition accuracy associated with forgery traces, thereby generating an optimized feature vector for subsequent models. Since the optimized feature vector contains richer prior scene information and has a higher signal-to-noise ratio, it is better suited to the current scene, thus improving the accuracy of subsequent model detection.

[0059] In practical applications, in step 105, the electronic device can preprocess the second feature vector (such as time alignment and normalization) and then input the preprocessed second feature vector into the first model to perform in-depth analysis and judgment through the first model, thereby outputting the detection result. Among them, a sliding window-based inference mechanism can be introduced into the first model to optimize the temporal dependency characteristics of the second feature vector, so that the first model can better capture traces of forged speech, while meeting the real-time and adaptability requirements in complex scenarios.

[0060] For example, such as Figure 2As shown, the second feature vector comprises four transformed feature vectors: a feature vector representing physiological characteristics (also known as micro-jitter characteristics), a feature vector representing signal structure (also known as phase distortion characteristics), a feature vector representing frequency domain characteristics (also known as spectral characteristics), and a feature vector representing time domain characteristics (also known as temporal characteristics). The electronic device can perform normalization and positional encoding on the second feature vector to preprocess it. Positional encoding can be understood as adding temporal position information (such as a timestamp) to each frame of the feature sequence in the second feature vector, so that the first model can identify the dynamic characteristics of the feature vector over time while capturing the static correlation between feature vectors. Then, the electronic device can input the preprocessed second feature vector into the first model to learn the interaction or dependency relationships between feature vectors of different dimensions through the multi-head attention mechanism of the first model, thereby efficiently capturing the global correlation (or global pattern) of the feature vectors. Furthermore, in the aforementioned recognition process, the sliding window-based inference mechanism in the first model can divide the preprocessed second feature vector into multiple analysis windows of fixed length (e.g., 100ms), and slide them forward with small step sizes (e.g., 15ms or 30ms, which can be understood as the detection period). As the analysis window moves forward, the electronic device can extract features from the feature vectors within the analysis window and update the feature representation corresponding to the analysis window. Leveraging the parallel processing capability of the first model, the electronic device can perform fusion analysis and inference on all feature vectors within the current analysis window (specifically, it can include historical feature vectors and newly input feature vectors). Thus, by optimizing and reusing the computational scope of the multi-head attention mechanism, only the newly input feature vectors and contextual information need to be calculated to output the detection result after each sliding analysis window, without recalculating all feature vectors in the entire analysis window. This shortens the processing delay of a single detection without reducing the analysis accuracy of the first model, achieving real-time monitoring of the speech signal.

[0061] In practical applications, to enhance the adaptability of the first model to novel forged speech, the electronic device also incorporates an online learning mechanism for the first model, enabling dynamic updates. Specifically, for detected forged speech, the electronic device can annotate it and update the incremental dataset with the annotated forged speech. This allows the first model to update its relevant parameters and dynamically adjust its classification threshold through incremental training to adapt to the characteristics of new forged speech. In this way, the first model can quickly adapt to new attack patterns, avoiding large-scale, time-consuming offline retraining due to model aging or decreased recognition rates, thus ensuring the model's long-term processing performance. In other words, this rapid adaptability ensures that in dynamically evolving threat scenarios, it can maintain efficient and accurate real-time detection performance without frequent service interruptions for model updates.

[0062] In the application example of this application, the multi-head self-attention mechanism of the first model can efficiently capture the global correlation between multi-dimensional feature vectors, thereby achieving global modeling of long-distance dependencies between multi-dimensional feature vectors (which can also be understood as better fusion of feature vectors). Furthermore, by combining a sliding window inference mechanism, the second feature vector can be divided into multiple overlapping analysis windows and processed in parallel while ensuring the integrity of contextual information. Through this inference mechanism, the first model can complete the analysis of the current analysis window and output the detection result within each sliding step, thus achieving real-time monitoring of the second feature vector. In this way, the processing latency of a single detection can be reduced to the millisecond level without reducing the analysis accuracy of the first model.

[0063] The signal detection method provided in this application embodiment receives a speech signal to be detected; analyzes the speech signal to be detected to obtain an analysis result, the analysis result representing a first application scenario to which the speech signal to be detected belongs; extracts features from the speech signal to be detected to obtain a first feature vector, the first feature vector containing feature vectors of at least two dimensions; transforms the first feature vector based on the analysis result to obtain a second feature vector; and obtains a detection result based on a first model and the second feature vector, the first model being used to detect the authenticity of the speech signal, the detection result representing whether the speech signal to be detected is fake speech. The technical solution provided in this application embodiment, after identifying the application scenario of the speech signal, optimizes the original multi-dimensional feature vector based on the identification result of the application scenario, so that the optimized multi-dimensional feature vector can adapt to the current application scenario. Thus, in complex and changing acoustic scenarios, the accuracy of speech signal detection using the model can be improved.

[0064] The following section provides a more detailed description of this application with reference to application examples.

[0065] In the application example of this application, a forged speech detection scheme with scene adaptability is proposed, which is suitable for complex communication scenarios with high noise and channel distortion. Specifically, the detection system dynamically adjusts the feature analysis strategy and detection model parameters according to the acoustic environment and application scenario of the input speech, so as to achieve high-precision and high-robust real-time detection in complex and ever-changing environments. The detection system includes four core modules: a scene intelligent recognition module, a dual-track feature extraction module, a scene conditional adaptive feature transform network (SCAFT-Net) module, and a real-time processing module.

[0066] Here, as Figure 3 As shown, the process of detecting fake voice includes the following steps: Step 301: Input the speech signal to be tested (i.e., the speech signal to be detected mentioned above). Step 302: The scene intelligent recognition module performs scene intelligent recognition; Here, the scene intelligent recognition module performs channel fingerprint recognition through narrowband energy distribution detection and vocoder artifact detection to obtain a channel fingerprint score (i.e., the first analysis result mentioned above); it performs conversation dynamics detection on the voice signal under test through a voice activity detection algorithm to obtain a conversation dynamics score (i.e., the second analysis result mentioned above); it performs weighted processing on the channel fingerprint score and the conversation dynamics score to obtain a probability value of the scene; and it determines the application scene (such as telephone communication, voice payment, etc.) to which the current voice signal under test belongs based on the probability value.

[0067] For example, taking a telephone communication scenario, in order to identify whether the voice has been processed by a telecommunications network, the scene intelligent recognition module performs channel fingerprint recognition through narrowband energy distribution detection and vocoder artifact detection to obtain a channel fingerprint score; it performs conversation dynamics detection on the voice signal under test through a voice activity detection algorithm to obtain a conversation dynamics score; it performs weighted fusion processing on the channel fingerprint score and the conversation dynamics score, and after activation through the Sigmoid function, it obtains the final probability of determining it as a telephone communication scenario; if the final probability is greater than a preset decision threshold, the application scenario is considered to be telephone communication.

[0068] Step 303: The scene intelligent recognition module outputs the scene determination result (i.e., the analysis result mentioned above); Step 304: The dual-track feature extraction module performs dual-track feature extraction; Here, the dual-track feature extraction module detects the fundamental frequency and amplitude of the speech signal under test in order to extract the feature vector (i.e., the third feature vector mentioned above) from the physiological feature dimension, and then judges the physical randomness of the speech signal under test; it also analyzes the phase distortion of the speech signal under test in order to extract the feature vector (i.e., the fourth feature vector mentioned above) from the signal structure dimension, and then judges the structural defects of the speech signal under test.

[0069] In addition, the dual-track feature extraction module also analyzes the spectral and temporal features of the speech signal under test in order to extract feature vectors from the time domain and frequency domain dimensions.

[0070] Step 305: Output the original feature vector (i.e., the first feature vector mentioned above); The original feature vector contains feature vectors of different dimensions, such as frequency domain, time domain, physiological characteristics, and signal structure.

[0071] Step 306: The scene condition adaptive feature transformation network module performs nonlinear feature transformation and reshaping on the original feature vector based on the scene determination result to generate an optimized feature vector (i.e., the second feature vector mentioned above). Here, the scene-conditional adaptive feature transformation network module maps discrete scene IDs into a dense scene vector based on the scene determination result. Next, the module divides the original feature vector into four feature groups. Based on the scene vector, weights (i.e., the first weights mentioned above) are assigned to different feature groups through the attention mechanism of the scene-conditional adaptive feature transformation network. Furthermore, a hypernetwork technique is used to dynamically generate a dedicated transformation function and corresponding weights (i.e., the second weights mentioned above) for each feature group to achieve non-linear feature reshaping. Finally, the module fuses the weighted and transformed feature groups to obtain an optimized feature vector for the current scene.

[0072] Step 307: Based on the real-time processing architecture of Transformer and sliding window, perform in-depth analysis on the optimized feature vector; Here, after normalization and time alignment, the optimized feature vector is input into the Transformer model, so that the Transformer model (i.e. the first model mentioned above) learns the complex interaction relationship between each dimension and other dimensions based on multiple sets of attention weight matrices corresponding to the multi-head self-attention mechanism, thereby extracting the global pattern of fake speech features for deep analysis.

[0073] Step 308: Output the final detection result.

[0074] Here, the Transformer model is used to perform in-depth analysis on the optimized feature vectors and output the final detection results to reflect whether the test speech signal is real speech or fake speech.

[0075] In the application example of this application, based on the scene judgment result of the input speech, a nonlinear transformation is performed on the original multidimensional features to optimize them. This optimizes the multidimensional features to better fit the current scene, thereby improving the accuracy of model detection. Secondly, by introducing phase distortion analysis, unnaturalness in the signal's temporal structure can be identified, thus gaining generalization ability for different types of speech signals and improving the detection range. Finally, through the multi-head self-attention mechanism of the Transformer model, global modeling of long-distance dependencies between multidimensional features can be achieved. By combining a sliding window inference mechanism, continuous speech streams can be segmented into overlapping analysis windows, allowing for parallel processing while ensuring the integrity of contextual information. In this way, the processing latency of a single detection can be reduced to the millisecond level without reducing the model's analysis accuracy.

[0076] To implement the signal detection method of this application embodiment, this application embodiment also provides a signal detection device, which is installed on an electronic device, such as... Figure 4 As shown, the device includes: The receiving unit 401 is used to receive the voice signal to be detected; Analysis unit 402 is used to analyze the speech signal to be detected and obtain analysis results, wherein the analysis results characterize the first application scenario to which the speech signal to be detected belongs; The extraction unit 403 is used to extract features from the speech signal to be detected to obtain a first feature vector, wherein the first feature vector contains feature vectors of at least two dimensions. The first processing unit 404 is used to transform the first feature vector based on the analysis result to obtain the second feature vector; The second processing unit 405 obtains a detection result based on the first model and the second feature vector. The first model is used to detect the authenticity of the speech signal, and the detection result indicates whether the speech signal to be detected is fake speech.

[0077] In some optional embodiments, the analysis unit 402 is configured to analyze the speech signal to be detected to obtain a first analysis result, the first analysis result representing the channel characteristics of the speech signal to be detected during transmission; analyze the speech signal to be detected to obtain a second analysis result, the second analysis result representing the degree to which the interaction mode of the speech signal to be detected conforms to the interaction mode of one or more application scenarios; and obtain the analysis result based on the first analysis result and the second analysis result.

[0078] In some optional embodiments, the analysis unit 402 is configured to perform energy distribution detection on the speech signal to be detected to obtain a third analysis result, the third analysis result being characterized as to whether the speech signal to be detected is a narrowband signal; perform vocoder detection on the speech signal to be detected to obtain a fourth analysis result, the fourth analysis result being characterized as to the probability of the presence of a first vocoder; and perform fusion processing on the third analysis result and the fourth analysis result to obtain the first analysis result.

[0079] In some optional embodiments, the first processing unit 404 is configured to obtain a first weight based on the analysis results, the first weight representing the importance of the first feature vector in the first application scenario; obtain a second weight based on the analysis results, the second weight including the weight of the transformation network to which the first feature vector belongs; and perform nonlinear transformation processing on the first feature vector based on the first weight and the second weight to obtain the second feature vector.

[0080] In some optional embodiments, the first processing unit 404 is configured to group the first feature vector to obtain at least two feature groups; generate at least two fifth analysis results based on the analysis results and the at least two feature groups, each fifth analysis result representing the attention score of the corresponding feature group in the first application scenario; and obtain the first weight based on the at least two fifth analysis results.

[0081] In some optional embodiments, the extraction unit 403 is used to extract features from the speech signal to be detected to obtain a third feature vector, the third feature vector representing the defects of the speech signal to be detected in the physiological feature dimension; to extract features from the speech signal to be detected to obtain a fourth feature vector, the fourth feature vector representing the defects of the speech signal to be detected in the signal structure dimension; and to obtain a first feature vector based on the third feature vector and the fourth feature vector.

[0082] In some optional embodiments, the extraction unit 403 is used to transform the speech signal to be detected to obtain a transformed speech signal; based on the transformed speech signal, a sixth analysis result is obtained, the sixth analysis result including the instantaneous phase and instantaneous frequency corresponding to the transformed speech signal; based on the sixth analysis result, the fourth feature vector is obtained.

[0083] It should be noted that the signal detection device provided in the above embodiments is only illustrated by the division of the above-described program units. In practical applications, the above processing can be assigned to different program units as needed, that is, the internal structure of the device can be divided into different program units to complete all or part of the processing described above. In addition, the signal detection device and the signal detection method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0084] Based on the hardware implementation of the above program modules, and in order to implement the signal detection method of the embodiments of this application, the embodiments of this application also provide an electronic device, such as... Figure 5 As shown, the electronic device 500 includes: The communication interface 501 enables interaction with other devices; The processor 502 is connected to the communication interface 501 to enable interaction with other devices and to execute the methods provided by one or more of the above-mentioned technical solutions when running a computer program; The computer program is stored in memory 503.

[0085] Specifically, the communication interface 501 is used to receive the voice signal to be detected; The processor 502 is configured to analyze the speech signal to be detected and obtain an analysis result, the analysis result representing a first application scenario to which the speech signal to be detected belongs; extract features from the speech signal to be detected to obtain a first feature vector, the first feature vector containing feature vectors of at least two dimensions; transform the first feature vector based on the analysis result to obtain a second feature vector; and obtain a detection result based on a first model and the second feature vector, the first model being used to detect the authenticity of the speech signal, the detection result representing whether the speech signal to be detected is fake speech.

[0086] In some optional embodiments, the processor 502 is configured to analyze the speech signal to be detected to obtain a first analysis result, the first analysis result being characterized by the channel characteristics of the speech signal to be detected during transmission; analyze the speech signal to be detected to obtain a second analysis result, the second analysis result being characterized by the degree to which the interaction mode of the speech signal to be detected conforms to the interaction mode of one or more application scenarios; and obtain the analysis result based on the first analysis result and the second analysis result.

[0087] In some optional embodiments, the processor 502 is configured to perform energy distribution detection on the speech signal to be detected to obtain a third analysis result, the third analysis result being characterized as to whether the speech signal to be detected is a narrowband signal; perform vocoder detection on the speech signal to be detected to obtain a fourth analysis result, the fourth analysis result being characterized as to the probability of the presence of a first vocoder; and perform fusion processing on the third analysis result and the fourth analysis result to obtain the first analysis result.

[0088] In some optional embodiments, the processor 502 is configured to obtain a first weight based on the analysis results, the first weight representing the importance of the first feature vector in the first application scenario; obtain a second weight based on the analysis results, the second weight including the weight of the transformation network to which the first feature vector belongs; and perform nonlinear transformation processing on the first feature vector based on the first weight and the second weight to obtain the second feature vector.

[0089] In some optional embodiments, the processor 502 is configured to group the first feature vector to obtain at least two feature groups; generate at least two fifth analysis results based on the analysis results and the at least two feature groups, each fifth analysis result representing the attention score of the corresponding feature group in the first application scenario; and obtain the first weight based on the at least two fifth analysis results.

[0090] In some optional embodiments, the processor 502 is configured to perform feature extraction on the speech signal to be detected to obtain a third feature vector, the third feature vector representing the defects of the speech signal to be detected in the physiological feature dimension; perform feature extraction on the speech signal to be detected to obtain a fourth feature vector, the fourth feature vector representing the defects of the speech signal to be detected in the signal structure dimension; and obtain a first feature vector based on the third feature vector and the fourth feature vector.

[0091] In some optional embodiments, the processor 502 is configured to perform transformation processing on the speech signal to be detected to obtain a transformed speech signal; based on the transformed speech signal, obtain a sixth analysis result, the sixth analysis result including the instantaneous phase and instantaneous frequency corresponding to the transformed speech signal; and based on the sixth analysis result, obtain the fourth feature vector.

[0092] It should be noted that the specific processing procedures of the communication interface 501 and the processor 502 can be understood with reference to the above method.

[0093] Of course, in practical applications, the various components in electronic device 500 are coupled together through bus system 504. It can be understood that bus system 504 is used to realize the connection and communication between these components. In addition to a data bus, bus system 504 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, in... Figure 5 The general designated all buses as Bus System 504.

[0094] The memory 503 in this embodiment is used to store various types of data to support the operation of the electronic device 500. Examples of such data include any computer program used to operate on the electronic device 500.

[0095] The methods disclosed in the embodiments of this application can be applied to the processor 502, or implemented by the processor 502. The processor 502 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor 502 or by instructions in the form of software. The processor 502 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 502 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software modules may be located in a storage medium, which is located in the memory 503. The processor 502 reads the information in the memory 503 and combines its hardware to complete the steps of the aforementioned method.

[0096] In an exemplary embodiment, the electronic device 500 may be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers (MCUs), microprocessors, or other electronic components to perform the aforementioned method.

[0097] It is understood that the memory (memory 503) in this embodiment of the application can be volatile memory or non-volatile memory, or both. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); the magnetic surface memory can be disk storage or magnetic tape storage. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM).The memories described in the embodiments of this application are intended to include, but are not limited to, these and any other suitable types of memories.

[0098] In an exemplary embodiment, this application also provides a storage medium, namely a computer storage medium, specifically a computer-readable storage medium, such as a memory 503 storing a computer program, which can be executed by the processor 502 of the electronic device 500 to complete the steps of the aforementioned signal detection method. The computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, Flash Memory, magnetic surface memory, optical disc, or CD-ROM.

[0099] In an exemplary embodiment, this application also provides a computer program product, including a computer program that can be executed by a processor 502 of an electronic device 500 to complete the steps of the aforementioned signal detection method.

[0100] It should be noted that terms such as "first" and "second" are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0101] Furthermore, the technical solutions described in the embodiments of this application can be combined arbitrarily without conflict.

[0102] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application.

Claims

1. A signal detection method, characterized in that, The method includes: Receive the voice signal to be detected; The speech signal to be detected is analyzed to obtain an analysis result, and the analysis result represents the first application scenario to which the speech signal to be detected belongs; The speech signal to be detected is subjected to feature extraction to obtain a first feature vector, which contains feature vectors of at least two dimensions. Based on the analysis results, the first feature vector is transformed to obtain the second feature vector; Based on the first model and the second feature vector, a detection result is obtained. The first model is used to detect the authenticity of the speech signal, and the detection result indicates whether the speech signal to be detected is fake speech.

2. The method according to claim 1, characterized in that, The analysis of the speech signal to be detected, to obtain the analysis result, includes: The speech signal to be detected is analyzed to obtain a first analysis result, which characterizes the channel characteristics of the speech signal to be detected during transmission. The speech signal to be detected is analyzed to obtain a second analysis result, which characterizes the degree to which the interaction mode of the speech signal to be detected conforms to the interaction mode of one or more application scenarios. The analysis results are obtained based on the first analysis results and the second analysis results.

3. The method according to claim 2, characterized in that, The analysis of the speech signal to be detected to obtain a first analysis result includes: Energy distribution detection is performed on the speech signal to be detected to obtain a third analysis result, which characterizes whether the speech signal to be detected is a narrowband signal. The speech signal to be detected is subjected to vocoder detection to obtain a fourth analysis result, which represents the probability of the presence of the first vocoder. The third and fourth analysis results are fused together to obtain the first analysis result.

4. The method according to claim 1, characterized in that, Based on the analysis results, the first feature vector is transformed to obtain the second feature vector, including: Based on the analysis results, a first weight is obtained, which represents the importance of the first feature vector in the first application scenario; Based on the analysis results, a second weight is obtained, which includes the weight of the transformation network to which the first feature vector belongs; Based on the first weight and the second weight, the first feature vector is subjected to a nonlinear transformation to obtain the second feature vector.

5. The method according to claim 4, characterized in that, The first weight obtained based on the analysis results includes: The first feature vector is grouped to obtain at least two feature groups; Based on the analysis results and the at least two feature groups, at least two fifth analysis results are generated, each fifth analysis result representing the attention score of the corresponding feature group in the first application scenario; The first weight is obtained based on the at least two fifth analysis results.

6. The method according to claim 1, characterized in that, The step of extracting features from the speech signal to be detected to obtain a first feature vector includes: Feature extraction is performed on the speech signal to be detected to obtain a third feature vector, which represents the defects of the speech signal to be detected in the physiological feature dimension. Feature extraction is performed on the speech signal to be detected to obtain a fourth feature vector, which represents the defects of the speech signal to be detected in the signal structure dimension. The first feature vector is obtained based on the third feature vector and the fourth feature vector.

7. The method according to claim 6, characterized in that, The step of extracting features from the speech signal to be detected to obtain a fourth feature vector includes: The speech signal to be detected is transformed to obtain the transformed speech signal; Based on the transformed speech signal, a sixth analysis result is obtained, which includes the instantaneous phase and instantaneous frequency of the transformed speech signal; Based on the sixth analysis result, the fourth feature vector is obtained.

8. An electronic device, characterized in that, include: The processor and the memory used to store computer programs that can run on the processor. When the processor is used to run the computer program, it performs the steps of the method according to any one of claims 1 to 7.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images