A voiceprint collection and analysis method and system based on artificial intelligence

Abstract

The present invention discloses a voiceprint collection and analysis method and system based on artificial intelligence, which relates to the field of artificial intelligence and voiceprint recognition technology, including collecting voiceprint data through a 128MIC microphone array, performing STFT time domain signal conversion on the voiceprint data to obtain the time-frequency domain and extracting noise features from it, using CNN to identify the noise type, dynamically adjusting the parameters of the MMSE algorithm and performing adaptive gain control of the sound level, building a deep voiceprint feature learning model that integrates time and space domains, identifying and analyzing the voiceprint, building an environmental voiceprint map, and performing real-time scene matching and noise reduction strategy selection. The method described in the present invention collects voiceprints by using a 128MIC array, combined with AI technology: STFT analyzes noise, CNN identifies the type and dynamically adjusts the noise reduction algorithm; builds a deep learning model to integrate time and space features to improve recognition accuracy; creates a voiceprint map to intelligently match scenes and noise reduction strategies, forming an efficient and highly adaptable system, and significantly improving the voiceprint recognition effect in complex environments.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and voiceprint recognition technology, and in particular to a voiceprint collection and analysis method and system based on artificial intelligence. Background Art

[0002] In recent years, with the rapid development of artificial intelligence technology, voiceprint recognition technology has achieved a rapid transformation from theoretical exploration to widespread application. In the early days, this field mainly relied on classic signal processing algorithms to extract and match voiceprint features. Faced with increasingly complex and changeable practical application scenarios, such as environmental noise interference and pronunciation changes caused by the speaker's emotional fluctuations, the performance of traditional methods has dropped sharply, highlighting their poor environmental adaptability and lack of robustness.

[0003] To overcome these limitations, researchers have gradually turned to models based on deep learning, such as convolutional neural networks (CNNs), and have used their powerful automatic feature learning capabilities to significantly improve the accuracy and noise resistance of voiceprint recognition. Despite this, existing technologies still have obvious shortcomings in the effective integration of spatiotemporal features and adaptive strategies for different noise environments. Specifically, they have limited environmental perception capabilities and find it difficult to accurately adapt to and eliminate complex background noise in real time; feature extraction is mostly concentrated in a single field, ignoring the complementary advantages of time domain and frequency domain information; there is a lack of intelligent noise classification and personalized noise reduction strategies, which affects the versatility and reliability of the recognition system. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a voiceprint collection and analysis method and system based on artificial intelligence to solve the problems of large noise interference, unstable data quality and poor environmental adaptability in voiceprint recognition.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, an embodiment of the present invention provides a voiceprint collection and analysis method based on artificial intelligence, which includes collecting voiceprint data through a 128MIC microphone array, performing STFT time domain signal conversion on the voiceprint data to obtain a time-frequency domain and extracting noise features therefrom, using CNN to identify the noise type, dynamically adjusting the parameters of the MMSE algorithm based on the noise type and performing sound level adaptive gain control, optimizing the voiceprint data, and using the optimized voiceprint data to construct a deep voiceprint feature learning model that integrates time and space domains, identifying and analyzing the voiceprint, and based on the voiceprint analysis results, constructing an environmental voiceprint map to perform real-time scene matching and noise reduction strategy selection.

[0008] As a preferred solution of the voiceprint collection and analysis method based on artificial intelligence described in the present invention, the voiceprint data is collected through a 128MIC microphone array, the voiceprint data is converted from the STFT time domain signal to the time-frequency domain and noise features are extracted from it, and the noise type is identified using CNN. The specific steps are as follows:

[0009] Generate a detailed time-frequency spectrum diagram, the expression is as follows:

[0010]

[0011] Where S(t, f) represents the amplitude of the time-spectrum diagram at time t and frequency f;

[0012] From the time-frequency spectrum obtained by STFT, the Mel frequency cepstrum coefficients are extracted as the main frequency domain features of the voiceprint, and the extracted main frequency domain features are processed by normalization to obtain the noise features;

[0013] For the extracted noise features, CNN is used to identify the noise type and obtain the noise type classification result.

[0014] As a preferred solution of the voiceprint collection and analysis method based on artificial intelligence described in the present invention, the specific steps of dynamically adjusting the parameters of the MMSE algorithm and performing adaptive gain control of the sound level based on the noise type to optimize the voiceprint data are as follows:

[0015] Based on the noise types identified by CNN, a mapping relationship between noise types and MMSE parameter preset sets is established. Each type of noise corresponds to a set of optimal MMSE parameter initial values, as shown in the following expression:

[0016]

[0017] in, is the observed signal, y[n] is the noise prediction value, and are the power spectral densities of signal and noise respectively;

[0018] The signal gain is dynamically adjusted through the sound level adaptive gain control algorithm to maintain a stable output performance. The expression is as follows:

[0019]

[0020] Where G[n] is the gain adjustment factor of the current frame, γ is the scaling factor, TargetLevel is the target sound level, CurrentLevel[n] is the sound level estimate of the previous frame, MinGain is the minimum gain limit, η adjusts the factor affected by the noise type, and TypeFactor[n] is the factor adjusted based on the noise type.

[0021] As a preferred solution of the voiceprint collection and analysis method based on artificial intelligence described in the present invention, the optimized voiceprint data is used to construct a deep voiceprint feature learning model integrating time and space domains, extract and analyze voiceprint features and time and space domain features, and identify and analyze the voiceprint. The specific steps are as follows:

[0022] The optimized time domain signal is spliced ​​with the frequency domain features to form a spatiotemporal feature matrix, which provides an efficient and unified input representation for the model. The expression is as follows:

[0023] F st (t, f)=α·S′(t, f)+(1-α)·MFCC′(f);

[0024] Among them, S′(t, f) is the time-frequency spectrum after noise reduction, MFCC′(f) is the frequency domain feature adjusted to match the time dimension of the time-frequency feature, α is the fusion weight, t is the time variable, and f is the frequency variable;

[0025] Build a deep learning model that integrates time and space domains, input the time and space matrix into the model, and perform voiceprint recognition. The expression is as follows:

[0026]

[0027] Where * represents the convolution operation, which combines filtering in the time and frequency dimensions, K is the number of convolutional layers, and W k is the convolution kernel weight matrix of the kth layer, reflecting the mode of feature extraction, b k is the corresponding bias term, σ is an activation function used to introduce nonlinearity, is a function determined by the parameter set Θ, and y is the result of voiceprint recognition;

[0028] The loss function L is defined to measure the difference between the model prediction result and the true label. The expression is as follows:

[0029]

[0030] Among them, L is the value of the loss function, i is an index variable, and y i is an independent encoding vector, p(y i ) is the probability distribution of the model predicting that sample i belongs to each category;

[0031] Based on the loss function L, the Adam optimizer is used to update the parameters. The expression is as follows:

[0032]

[0033] Among them, Θ new is the updated new parameter value, Θ oldis the parameter used by the model in the current iteration step, η is the learning rate, is a vector loss function consisting of partial derivatives with respect to each parameter.

[0034] As a preferred solution of the voiceprint collection and analysis method based on artificial intelligence described in the present invention, wherein: the environmental voiceprint map is constructed according to the voiceprint analysis results, and the specific steps are as follows:

[0035] The voiceprint recognition result y obtained by the spatiotemporal fusion deep learning model is mapped back to the feature vector space. The expression is as follows:

[0036] V i =φ(y i );

[0037] Among them, V i represents the feature vector mapped to the i-th voiceprint sample, φ is the mapping function, (y i ) is the category probability distribution of the model output or the directly associated feature representation;

[0038] Collect and integrate voiceprint samples in different environments, and use feature vector space mapping to form an environmental voiceprint feature library, which is expressed as follows:

[0039] ε={V1,V2,...,V M};

[0040] Among them, ε represents the environmental voiceprint feature library, and M is the total number of feature vectors.

[0041] As a preferred solution of the voiceprint collection and analysis method based on artificial intelligence described in the present invention, the specific steps of performing real-time scene matching are as follows:

[0042] Using cosine similarity as the basis and combining it with the environmental noise model, the environmental matching function is generated. The expression is as follows:

[0043]

[0044] Among them, V test is the real-time voiceprint feature vector to be matched, V j is the jth voiceprint feature vector in the library, Γ(V test , V j ) is the similarity between the real-time voiceprint feature and a certain environment voiceprint feature in the library, λ j It is the weight that is dynamically adjusted according to the environmental noise model;

[0045] The voiceprint data collected in real time is processed through the spatiotemporal fusion model to obtain the feature vector, and then the feature library ε is traversed to calculate the correlation between each feature vector and V test, and find the environment with the highest matching degree. The expression is as follows:

[0046]

[0047] in, is the environment index with the highest matching degree, Solve for the maximum value.

[0048] As a preferred solution of the voiceprint collection and analysis method based on artificial intelligence of the present invention, wherein: the noise reduction strategy selection is based on the matched environment The complexity index of the environmental noise model and the clarity index of the voiceprint feature Ω(V test ) and select the most appropriate strategy from the preset noise reduction strategy library. The expression is as follows:

[0049]

[0050] in, is the complexity index of the environmental noise model, Ω(V test ) is the clarity index of the voiceprint feature, f is the strategy selection function, is the most appropriate noise reduction strategy selected.

[0051] In the second aspect, the present invention provides a voiceprint collection and analysis system based on artificial intelligence, including an environment perception and voiceprint collection module, a time-frequency conversion and noise feature extraction module, a noise type recognition and MMSE parameter adjustment module, a deep voiceprint feature learning and recognition module, an environment voiceprint map construction and real-time strategy selection module, the environment perception and voiceprint collection module is used to collect voiceprint data with high precision through a 128MIC microphone array, synchronously monitor the environmental status, and provide high-quality input for subsequent processing, the time-frequency conversion and noise feature extraction module is used to convert the collected voiceprint signal from the time domain to the time-frequency domain, and carefully extract the noise feature. , laying the foundation for noise type identification. The noise type identification and MMSE parameter adjustment module is used to dynamically adjust the MMSE algorithm parameters according to the noise type, implement sound level adaptive gain control, and optimize the voiceprint data quality. The deep voiceprint feature learning and recognition module is used to build a deep learning model that integrates time and space domains, and perform deep feature extraction and analysis on the optimized voiceprint data to achieve efficient voiceprint recognition. The environmental voiceprint map construction and real-time strategy selection module is used to dynamically construct an environmental voiceprint map based on the voiceprint analysis results, intelligently match the real-time scene, select the best noise reduction strategy, and improve the system adaptability and recognition accuracy.

[0052] In a third aspect, an embodiment of the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the voiceprint collection and analysis method based on artificial intelligence as described in the first aspect of the present invention is implemented.

[0053] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the voiceprint collection and analysis method based on artificial intelligence as described in the first aspect of the present invention is implemented.

[0054] The beneficial effects of the present invention are as follows: the present invention collects voiceprint data with high precision through a 128MIC array, and uses AI technology for in-depth analysis and processing, including STFT conversion to extract noise features, CNN intelligent identification of noise types to dynamically adjust noise reduction algorithm parameters, construction of a spatiotemporal fusion deep learning model to improve recognition accuracy, and creation of an environmental voiceprint map to achieve intelligent scene matching and noise reduction strategy selection. Ultimately, a set of efficient and adaptable voiceprint collection and analysis systems is formed, which significantly improves the accuracy and practicality of voiceprint recognition in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0056] Figure 1 This is a flow chart of the voiceprint collection and analysis method based on artificial intelligence in Example 1. DETAILED DESCRIPTION

[0057] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0058] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0059] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0060] Example 1, reference Figure 1 , which is the first embodiment of the present invention, provides a voiceprint collection and analysis method based on artificial intelligence, comprising the following steps:

[0061] S1. Collect voiceprint data through a 128MIC microphone array, perform STFT time domain signal conversion on the voiceprint data to obtain the time-frequency domain and extract noise features from it, and use CNN to identify the noise type. The specific steps are as follows:

[0062] Furthermore, a detailed time-frequency spectrum is generated, and the expression is as follows:

[0063]

[0064] Where S(t,f) represents the amplitude of the time-frequency spectrum at time t and frequency f.

[0065] STFT is a signal analysis method that divides the signal in the time domain into a series of shorter time segments and applies Fourier transform to each time segment to convert it into frequency domain representation.

[0066] From the time-frequency spectrum obtained by STFT, the Mel frequency cepstrum coefficients are extracted as the main frequency domain features of the voiceprint. The extracted main frequency domain features are processed by normalization to eliminate the amplitude changes caused by factors such as speaker volume differences and microphone sensitivity, ensure the comparability between features, and finally obtain the noise features;

[0067] The extracted noise features are classified using deep learning technology, especially convolutional neural networks, to identify the type of noise (such as environmental noise, speech noise, electronic equipment noise, etc.). The noise features are input into a pre-trained or de novo trained CNN model. The model will automatically learn the feature representations of different noise types and output the corresponding noise type classification results. This process not only improves the recognition accuracy, but also provides important prior information for subsequent noise suppression or voiceprint recognition algorithms.

[0068] S2. Based on the noise type, dynamically adjust the parameters of the MMSE algorithm and perform adaptive gain control of the sound level to optimize the voiceprint data;

[0069] Furthermore, based on the noise type, the parameters of the MMSE algorithm are dynamically adjusted and the sound level adaptive gain control is performed to optimize the voiceprint data. The specific steps are as follows:

[0070] Based on the noise types identified by CNN, a mapping relationship between noise types and MMSE parameter preset sets is established. Each type of noise corresponds to a set of optimal MMSE parameter initial values, as shown in the following expression:

[0071]

[0072] in, is the observed signal, y[n] is the noise prediction value, and are the power spectral densities of signal and noise respectively;

[0073] The signal gain is dynamically adjusted through the sound level adaptive gain control algorithm to maintain a stable output performance. The expression is as follows:

[0074]

[0075] Where G[n] is the gain adjustment factor for the current frame, γ is the scaling factor, TargetLevel is the target sound level, CurrentLevel[n] is the sound level estimate for the previous frame, MinGain is the minimum gain limit to prevent over-amplification of noise, TypeFactor[n] is a factor adjusted based on the noise type, and η is a factor that adjusts the influence of the noise type.

[0076] The MMSE (Mininum Mean Square Error) algorithm is a statistical signal processing technique used to estimate the optimal value of a signal in the presence of noise. In voiceprint recognition, it is usually used to separate pure voiceprint signals from noisy observation signals. Sound level adaptive gain control aims to automatically adjust the gain of the signal to achieve a stable and appropriate output sound level while suppressing noise and maintaining the naturalness and clarity of the speech. This process relies on real-time analysis of the current signal frame and dynamic adjustment based on historical sound level information. Although these two steps are not directly related, they both optimize the data in different ways, providing more accurate voiceprint data for the subsequent deep voiceprint feature learning model with spatiotemporal fusion.

[0077] S3. Use the optimized voiceprint data to build a deep voiceprint feature learning model that integrates time and space domains to identify and analyze voiceprints;

[0078] Furthermore, the optimized voiceprint data is used to build a deep voiceprint feature learning model that integrates time and space domains, extract and analyze voiceprint features and time and space domain features, and identify voiceprints. The specific steps are as follows:

[0079] The time domain signal is divided into time frames to form a one-dimensional vector sequence. After applying STFT to each time frame, the spectrum is flattened into a one-dimensional vector. The time domain feature vector and the corresponding frequency domain feature vector are spliced ​​by column to form a wide vector. The optimized time domain signal and the frequency domain feature are spliced ​​to form a spatiotemporal feature matrix, which provides an efficient and unified input representation for the model. The expression is as follows:

[0080] F st (t, f)=α·S′(t, f)+(1-α)·MFCC′(f);

[0081] Among them, S′(t, f) is the time-frequency spectrum after noise reduction, MFCC′(f) is the frequency domain feature adjusted to match the time dimension of the time-frequency feature, α is the fusion weight, t is the time variable, which represents the time point or temporal position of the sample in the signal analysis, and f is the frequency variable, which refers to the specific frequency component analyzed in the signal;

[0082] Build a deep learning model that integrates time and space domains, input the time and space matrix into the model, and perform voiceprint recognition. The expression is as follows:

[0083]

[0084] Where * represents the convolution operation, which combines filtering in the time and frequency dimensions, K is the number of convolutional layers, and W k is the convolution kernel weight matrix of the kth layer, reflecting the mode of feature extraction, b k is the corresponding bias term, σ is an activation function used to introduce nonlinearity, is a function or model determined by a parameter set Θ, and y is the result of voiceprint recognition;

[0085] The loss function L is defined to measure the difference between the model prediction result and the true label. The expression is as follows:

[0086]

[0087] Among them, L is the value of the loss function, i is an index variable, and y i is an independent encoding vector, p(y i ) is the probability distribution of the model predicting that sample i belongs to each category;

[0088] Based on the loss function L, the Adam optimizer is used to update the parameters. The expression is as follows:

[0089]

[0090] Among them, Θ new is the updated new parameter value, Θ oldis the parameter used by the model in the current iteration step, η is the learning rate, which determines the step size of each parameter update, is a vector of partial derivatives of the loss function with respect to each parameter.

[0091] The loss function L defines the goal of model training, which is to minimize the prediction error, and the Adam optimizer achieves this goal by adjusting the parameter Θ. It utilizes the gradient information of the loss function and dynamically adjusts the learning rate to accelerate the training process and improve the stability of model convergence. These two expressions work together to form the core mechanism of deep learning model training.

[0092] S4. Based on the voiceprint analysis results, build an environmental voiceprint map to perform real-time scene matching and noise reduction strategy selection.

[0093] Furthermore, based on the voiceprint analysis results, an environmental voiceprint map is constructed. The specific steps are as follows:

[0094] The voiceprint recognition result y obtained by the spatiotemporal fusion deep learning model is mapped back to the feature vector space. The model output is the probability distribution of the category. The feature vector corresponding to the highest probability is selected as the representative of the voiceprint. If the model output is directly associated with the feature space, the expression is directly used as follows:

[0095] V i =φ(y i );

[0096] Among them, V i represents the feature vector mapped to the i-th voiceprint sample, φ is the mapping function, (y i ) is the category probability distribution of the model output or the directly associated feature representation;

[0097] Collect and integrate voiceprint samples in different environments, and use feature vector space mapping to form an environmental voiceprint feature library. This will contain a set of feature vectors of voiceprints in various environments to construct an environmental voiceprint map. The expression is as follows:

[0098] ε={V1,V2,...,V M};

[0099] Among them, ε represents the environmental voiceprint feature library, and M is the total number of feature vectors.

[0100] To perform real-time scene matching, the specific steps are as follows:

[0101] An environmental matching function is designed to measure the similarity between the real-time voiceprint features and the features in the environmental voiceprint feature library. The cosine similarity is used as the basis, combined with the environmental noise model, and the dynamic adjustment factor λ considering the influence of noise is expressed as follows:

[0102]

[0103] Among them, V test is the real-time voiceprint feature vector to be matched, V j is the jth voiceprint feature vector in the library, Γ(V test , V j ) is the similarity between the real-time voiceprint feature and a certain environment voiceprint feature in the library, λ j It is a weight dynamically adjusted according to the environmental noise model, which is used to reflect the impact of noise on the matching degree;

[0104] V test The source is that when the real-time collected voiceprint data is processed by the spatiotemporal fusion model to obtain the feature vector y, it actually means using the model Processing the input spatiotemporal feature matrix According to the model design and task requirements, the feature vector V for subsequent matching operations is obtained in some form (such as direct output, probability to feature mapping, etc.) test , that is, V test It is one of the output results y of the deep learning model fusion in time and space domains;

[0105] The voiceprint data collected in real time is processed through the spatiotemporal fusion model to obtain the feature vector, and then the feature library ε is traversed to calculate the correlation between each feature vector and V test , and find the environment with the highest matching degree. The expression is as follows:

[0106]

[0107] in, is the environment index with the highest matching degree, Point out which of all possible values ​​of j maximizes the value of the matching function Γ.

[0108] Select a noise reduction strategy based on the matched environment The complexity index of the environmental noise model and the clarity index of the voiceprint feature Ω(V test ) and select the most appropriate strategy from the preset noise reduction strategy library. The expression is as follows:

[0109]

[0110] in, is the complexity index of the environmental noise model, Ω(V test ) is the clarity index of the voiceprint feature, f is the strategy selection function, is the most appropriate noise reduction strategy selected.

[0111] This process achieves dynamic optimization of the voiceprint recognition system in complex and changing environments through real-time voiceprint feature extraction and environment matching, as well as intelligent noise reduction strategy selection based on the matching results, ensuring maximum recognition performance.

[0112] This embodiment also provides a voiceprint collection and analysis system based on artificial intelligence, including an environment perception and voiceprint collection module, a time-frequency conversion and noise feature extraction module, a noise type recognition and MMSE parameter adjustment module, a deep voiceprint feature learning and recognition module, and an environment voiceprint map construction and real-time strategy selection module;

[0113] The environment perception and voiceprint collection module is used to collect voiceprint data with high precision through a 128MIC microphone array, synchronously monitor the environment status, and provide high-quality input for subsequent processing;

[0114] The time-frequency conversion and noise feature extraction module is used to convert the collected voiceprint signal from the time domain to the time-frequency domain, and carefully extract the noise features to lay the foundation for noise type identification;

[0115] The noise type identification and MMSE parameter adjustment module is used to dynamically adjust the MMSE algorithm parameters according to the noise type, implement sound level adaptive gain control, and optimize the voiceprint data quality;

[0116] The deep voiceprint feature learning and recognition module is used to build a deep learning model that integrates time and space domains, perform deep feature extraction and analysis on the optimized voiceprint data, and achieve efficient voiceprint recognition;

[0117] The environmental voiceprint map construction and real-time strategy selection module is used to dynamically construct an environmental voiceprint map based on the voiceprint analysis results, intelligently match the real-time scene, select the best noise reduction strategy, and improve the system adaptability and recognition accuracy.

[0118] This embodiment also provides a computer device, which is suitable for the case of a voiceprint collection and analysis method based on artificial intelligence, including: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the voiceprint collection and analysis method based on artificial intelligence proposed in the above embodiment.

[0119] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0120] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by the processor, the voiceprint collection and analysis method based on artificial intelligence as proposed in the above embodiment is implemented; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic storage, flash memory, disk or optical disk.

[0121] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A voiceprint collection and analysis method based on artificial intelligence, characterized by: The voiceprint data is collected through a 128MIC microphone array, and the STFT time domain signal is converted into the time-frequency domain. Noise features are extracted from the time domain, and CNN is used to identify the noise type. Based on the noise type, the parameters of the MMSE algorithm are dynamically adjusted and adaptive gain control of the sound level is performed to optimize the voiceprint data; Using the optimized voiceprint data, a deep voiceprint feature learning model integrating time and space domains is constructed to identify and analyze voiceprints. According to the voiceprint analysis results, an environmental voiceprint map is constructed to perform real-time scene matching and noise reduction strategy selection; The voiceprint data is collected by a 128MIC microphone array, the voiceprint data is converted from the STFT time domain signal to the time-frequency domain, and noise features are extracted from it, and the noise type is identified by using CNN. The specific steps are as follows: Generate a detailed time-frequency spectrum diagram, the expression is as follows: ； in, Represents the time spectrum at time and frequency The amplitude of From the time-frequency spectrum obtained by STFT, the Mel frequency cepstrum coefficients are extracted as the main frequency domain features of the voiceprint, and the extracted main frequency domain features are processed by normalization to obtain the noise features; For the extracted noise features, CNN is used to identify the noise type and obtain the noise type classification result; Based on the noise type, the parameters of the MMSE algorithm are dynamically adjusted and the sound level adaptive gain control is performed to optimize the voiceprint data. The specific steps are as follows: Based on the noise type identified by CNN, a mapping relationship between the noise type and the preset set of MMSE parameters is established. Each type of noise corresponds to a set of optimal MMSE parameter initial values. The denoised observation signal is calculated using the following expression: ； in, is the observed signal after denoising, is the observed signal, is the noise prediction value, and are the power spectral densities of signal and noise respectively; The signal gain is dynamically adjusted through the sound level adaptive gain control algorithm to maintain a stable output performance. The expression is as follows: ； in, is the gain adjustment factor of the current frame, is the proportionality coefficient, is the target sound level, is the sound level estimate of the previous frame, is the minimum gain limit, Factors that adjust the effect of noise type, is a factor adjusted based on the noise type; The optimized voiceprint data is used to construct a deep voiceprint feature learning model that integrates time and space domains, extract and analyze voiceprint features and time and space domain features, and identify and analyze voiceprints. The specific steps are as follows: The optimized time domain signal is spliced ​​with the frequency domain features to form a spatiotemporal feature matrix, which provides an efficient and unified input representation for the model. The expression is as follows: ； in, is the time-frequency spectrum after noise reduction, is the frequency domain feature adjusted to match the time dimension of the time-frequency feature, 𝛼 is the fusion weight, 𝑡 is the time variable, and 𝑓 is the frequency variable; Build a deep learning model that integrates time and space domains, input the time and space matrix into the model, and perform voiceprint recognition. The expression is as follows: ； Among them, ∗ represents the convolution operation, which combines filtering in the time and frequency dimensions. is the number of convolutional layers, is the convolution kernel weight matrix of the 𝑘th layer, reflecting the mode of feature extraction, is the corresponding bias term, is an activation function used to introduce nonlinearity. is a function determined by the parameter set Θ, It is the result of voiceprint recognition; Define the loss function 𝐿 to measure the difference between the model prediction result and the true label. The expression is as follows: in, is the value of the loss function, 𝑖 is an index variable, is an independent encoding vector, It is the probability distribution of the model predicting that sample 𝑖 belongs to each category; Based on the loss function Use the Adam optimizer to update the parameters. The expression is as follows: ； in, is the new parameter value after update, are the parameters used by the model in the current iteration step, is the learning rate, is a vector loss function consisting of partial derivatives with respect to each parameter.

2. The voiceprint collection and analysis method based on artificial intelligence according to claim 1, characterized in that: According to the voiceprint analysis results, the environmental voiceprint map is constructed, and the specific steps are as follows: The voiceprint recognition result 𝑦 obtained by the spatiotemporal fusion deep learning model is mapped back to the feature vector space, and the expression is as follows: ； in, represents the feature vector mapped to the 𝑖th voiceprint sample, is the mapping function, It is the category probability distribution of the model output or the directly associated feature representation; Collect and integrate voiceprint samples in different environments, and use feature vector space mapping to form an environmental voiceprint feature library, which is expressed as follows: ； in, Represents the environmental voiceprint feature library, is the total number of eigenvectors.

3. The voiceprint collection and analysis method based on artificial intelligence as claimed in claim 2, characterized in that: The specific steps of performing real-time scene matching are as follows: Using cosine similarity as the basis and combining it with the environmental noise model, the environmental matching function is generated. The expression is as follows: ； in, is the real-time voiceprint feature vector to be matched, is the 𝑗th voiceprint feature vector in the library, It is the similarity between the real-time voiceprint feature and a certain environment voiceprint feature in the library. It is the weight that is dynamically adjusted according to the environmental noise model; The voiceprint data collected in real time is processed through the spatiotemporal fusion model to obtain the feature vector, and then traverses the feature library , calculate each eigenvector and , and find the environment with the highest matching degree. The expression is as follows: ； in, is the environment index with the highest matching degree, Solve for the maximum value.

4. The voiceprint collection and analysis method based on artificial intelligence as described in claim 3 is characterized by: The noise reduction strategy selection refers to: Based on the most matching environment index , the complexity index of the environmental noise model and the clarity index of the voiceprint feature Combined, select the most appropriate strategy from the preset noise reduction strategy library, the expression is as follows: ； in, is the complexity index of the environmental noise model, It is the clarity index of the voiceprint feature. is the strategy selection function, is the most appropriate noise reduction strategy selected.

5. A voiceprint collection and analysis system based on artificial intelligence, based on the voiceprint collection and analysis method based on artificial intelligence according to any one of claims 1 to 4, characterized in that: It includes environment perception and voiceprint collection module, time-frequency conversion and noise feature extraction module, noise type identification and MMSE parameter adjustment module, deep voiceprint feature learning and identification module, environment voiceprint map construction and real-time strategy selection module; The environment perception and voiceprint collection module is used to collect voiceprint data with high precision through a 128MIC microphone array, synchronously monitor the environment status, and provide high-quality input for subsequent processing; The time-frequency conversion and noise feature extraction module is used to convert the collected voiceprint signal from the time domain to the time-frequency domain, and carefully extract the noise features to lay the foundation for noise type identification; The noise type identification and MMSE parameter adjustment module is used to dynamically adjust the MMSE algorithm parameters according to the noise type, implement sound level adaptive gain control, and optimize the voiceprint data quality; The deep voiceprint feature learning and recognition module is used to build a deep learning model that integrates time and space domains, perform deep feature extraction and analysis on the optimized voiceprint data, and achieve efficient voiceprint recognition; The environmental voiceprint map construction and real-time strategy selection module is used to dynamically construct an environmental voiceprint map based on the voiceprint analysis results, intelligently match the real-time scene, select the best noise reduction strategy, and improve the system adaptability and recognition accuracy.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the voiceprint collection and analysis method based on artificial intelligence described in any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a voiceprint collection and analysis method based on artificial intelligence as described in any one of claims 1 to 4 are implemented.

Images