Method for automatically generating parameters of audio equalizer based on deep learning

Abstract

The invention relates to the technical field of audio processing, in particular to a method for automatically generating parameters of an audio equalizer based on deep learning. The method comprises the following steps: collecting multi-source audio data and carrying out short-time Fourier transform and principal component analysis, generating a dimension reduction feature set, obtaining reference template data and constructing a parameter space model, carrying out corresponding relation analysis on the dimension reduction feature set, generating mapping relation data, and obtaining the multi-source audio data. After a deep learning network is built based on mapping relation data and interlayer links are optimized, optimization model structure data are generated and new audio samples are collected, iterative training is carried out by utilizing an optimization model to generate a deep audio equalization parameter model, a candidate equalizer parameter set is predicted, the candidate parameter set is processed through difference correction, feedback equalization optimization is carried out, and a deep audio equalization model is obtained. Accurate equalizer parameters are generated, and the automatic equalizer parameter generation process is completed. According to the invention, a more efficient equalizer parameter automatic generation method is realized.

Description

Technical Field

[0001] The present invention relates to the field of audio processing technology, and in particular to a method for automatically generating audio equalizer parameters through deep learning. Background Art

[0002] In the field of audio processing, the adjustment of equalizer parameters is of great significance for achieving sound quality optimization and personalized needs. However, the setting of equalizer parameters in existing methods often relies on manual adjustment based on the experience of professional audio engineers, which is inefficient and difficult to apply on a large scale. At the same time, due to the multi-dimensional characteristics and complexity of audio signals, traditional signal analysis and processing methods are difficult to fully capture their underlying laws, resulting in the lack of accuracy and robustness in the generation of equalization parameters. In addition, existing methods usually rely on fixed parameter templates or static rules, which makes it difficult to dynamically adapt the characteristics of new audio samples, and thus cannot meet the audio processing requirements under different scenarios and personalized needs. In addition, existing deep learning methods often ignore the nonlinear correlation between audio features in mapping relationship modeling, resulting in the generated parameter model being insufficient in complex audio samples. Summary of the invention

[0003] Based on this, it is necessary to provide a deep learning method for automatically generating audio equalizer parameters to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a deep learning method for automatically generating audio equalizer parameters includes the following steps:

[0005] Step S1: Collect multi-source audio data; perform short-time Fourier transform on the multi-source audio data to obtain an audio feature vector; perform principal component analysis on the audio feature vector to generate a dimensionality reduction feature set;

[0006] Step S2: obtaining reference template data; performing parameter space modeling on the reference template data to obtain a parameter space model; performing correspondence analysis on the dimension reduction feature set based on the parameter space model to generate mapping relationship data;

[0007] Step S3: constructing a deep learning network for the mapping relationship data to generate relationship model structure data; optimizing the inter-layer links of the relationship model structure data to generate optimized model structure data;

[0008] Step S4: collecting new audio samples; performing iterative model training based on the optimized model structure data to generate a deep audio equalization parameter model; using the deep audio equalization parameter model to perform parameter prediction on the new audio samples to obtain a candidate equalizer parameter set;

[0009] Step S5: performing difference correction processing on the candidate equalizer parameter set and the new audio sample to obtain parameter correction data; performing feedback equalization optimization on the candidate equalizer parameter set according to the parameter correction data to generate final equalizer parameters.

[0010] The present invention collects multi-source audio data and extracts audio feature vectors through short-time Fourier transform, so that the audio data is accurately quantified in the frequency domain characteristics. At the same time, the audio feature vector is reduced in dimension through principal component analysis to generate a reduced-dimensional feature set, which effectively reduces data redundancy and dimensional complexity, improves the calculation efficiency and analysis accuracy in the subsequent processing process, obtains reference template data and performs parameter space modeling, establishes a parameter space model, and realizes a systematic description of multi-dimensional audio features. At the same time, based on the parameter space model, the corresponding relationship analysis of the reduced-dimensional feature set is performed to generate mapping relationship data, so that a high-correlation mapping is formed between the multi-dimensional features and the template, and the logical coherence and accuracy of the mapping relationship data are ensured. By building a deep learning network for the mapping relationship data, relationship model structure data is generated, and a multi-layer neural network is established to deeply model the nonlinear relationship of the mapping data. The inter-layer links are optimized, and the optimized model structure data is generated, which significantly improves the convergence speed and parameter expression ability of the network model, ensures the efficiency and robustness of the deep model, collects new audio samples and performs iterative model training based on the optimized model structure data, generates a deep audio equalization parameter model, and realizes the adaptive ability of the equalization parameter model to new audio samples. Through the generation of candidate equalizer parameter sets, a parametric expression of the multi-dimensional features of audio samples is formed to ensure the diversity and comprehensiveness of the equalization parameters. The candidate equalizer parameter sets and new audio samples are differentially corrected to obtain parameter correction data, so that the deviation of the equalizer parameters is controlled within an acceptable range. At the same time, the final equalizer parameters are generated through feedback equalization optimization to ensure the precise adaptability of the equalizer parameters to the new audio samples, so that the finally generated parameters have high stability and consistency, and improve the delicacy of the audio equalization effect and the satisfaction of the user experience.

[0011] Preferably, step S1 comprises the following steps:

[0012] Step S11: collecting multi-source audio data; performing time-frequency domain conversion on the multi-source audio data to obtain an initial spectrum diagram;

[0013] Step S12: performing frame processing on the initial spectrum graph to generate a frame sequence; performing short-time Fourier transform on the frame sequence to obtain an audio feature vector;

[0014] Step S13: normalizing the audio feature vector to obtain a standardized feature matrix;

[0015] Step S14: Calculate statistical parameters of the standardized feature matrix to obtain a statistical feature set; perform principal component analysis on the statistical feature set to generate a reduced-dimensional feature set.

[0016] The present invention ensures the diversity of data sources by collecting multi-source audio data, and can capture audio information under different environments and conditions. The initial spectrum diagram obtained after time-frequency domain conversion provides rich spectrum features for subsequent analysis, ensuring that the frequency components of the audio signal can be fully resolved. The implementation of framing processing on the initial spectrum diagram can divide the audio signal into multiple manageable time periods, and the generated frame sequence provides fine-grained data support for subsequent feature extraction. The audio feature vector generated by short-time Fourier transform can effectively reflect the changing characteristics of the audio signal in time and frequency. The implementation of normalizing the audio feature vector can eliminate the scale differences between different audio signals. The generated standardized feature matrix provides a consistent feature representation for subsequent analysis, ensuring that various audio features can be compared and analyzed under the same standard. The implementation of statistical parameter calculation on the standardized feature matrix can extract the key statistical features of the audio signal. The generated statistical feature set provides a general description of the overall features of the audio signal. The application of principal component analysis can significantly reduce the feature dimension. The obtained dimensionality reduction feature set provides concise and effective feature information for the input of the subsequent deep learning model, ensuring the training efficiency and performance optimization of the model.

[0017] Preferably, step S2 comprises the following steps:

[0018] Step S21: Acquire reference template data; identify the data type of the reference template data to obtain template type data;

[0019] Step S22: performing parameter space modeling on the reference template data according to the template type data to obtain a parameter space model;

[0020] Step S23: performing virtual space projection on the dimension reduction feature set based on the parameter space model to obtain feature projection data; performing feature matching analysis on the feature projection data to generate a feature parameter matching set;

[0021] Step S24: capturing the change trend of the feature parameter matching set to obtain the feature-parameter change law; performing mapping relationship evolution on the feature-parameter change law to generate mapping relationship data.

[0022] The present invention ensures the basic data source in the process of generating equalizer parameters by obtaining reference template data. The template type data obtained after data type identification provides a clear classification basis for subsequent modeling, ensuring the pertinence and effectiveness of parameter space modeling. The implementation of parameter space modeling of the reference template based on the template type data can form a clear parameter space model, providing a multi-dimensional parameter perspective for subsequent analysis, thereby improving the adaptability of the model under different audio conditions. The implementation of virtual space projection of the reduced dimension feature set based on the parameter space model can present complex audio features in a more intuitive way. The generated feature projection data provides an effective basis for feature matching analysis. The feature parameter matching set generated after feature matching analysis can reveal the relationship between audio features and equalizer parameters, and promote the precise adjustment of parameters. The implementation of capturing the change trend of the feature parameter matching set can identify the dynamic relationship between parameters and features. The obtained feature-parameter change law provides an important basis for the subsequent evolution of the mapping relationship. The generation of mapping relationship data ensures the intelligence and automation of the equalizer parameter generation process, and improves the practicality and efficiency of the deep learning model in audio processing.

[0023] Preferably, step S24 comprises the following steps:

[0024] Step S241: performing time series decomposition on the feature parameter matching set to obtain a time series feature sequence; performing cluster analysis on the time series feature sequence to generate trend category data;

[0025] Step S242: performing pattern recognition on the trend category data to obtain basic pattern data; performing similarity calculation based on the basic pattern data to generate inter-pattern association data;

[0026] Step S243: Summarize the regularity of the inter-pattern correlation data to obtain regularity index data; capture the change trend based on the regularity index data to generate the feature-parameter change regularity;

[0027] Step S244: Analyze the corresponding relationship of the feature-parameter variation rule to obtain mapping corresponding data; and evolve the mapping relationship according to the mapping corresponding data to generate mapping relationship data.

[0028] The present invention can extract the dynamic characteristics of parameter changes by implementing the time series decomposition of the feature parameter matching set, and the generated time series feature sequence provides basic data for trend analysis. The trend category data obtained after cluster analysis can effectively identify different parameter change patterns, ensuring the pertinence and accuracy of subsequent analysis. The implementation of pattern recognition of trend category data can reveal the potential basic patterns in the audio processing process, and the generated basic pattern data provides a basis for similarity calculation, which promotes the correlation analysis between patterns. The implementation of regular induction of inter-pattern correlation data can identify the inherent law of parameter changes, and the obtained regular index data provides a reference for parameter adjustment. The implementation of change trend capture based on regular index data can capture the change relationship between features and parameters. The generated feature-parameter change law lays the foundation for subsequent mapping analysis. The implementation of corresponding relationship analysis of feature-parameter change law can clarify the mapping relationship, and the obtained mapping corresponding data provides effective support for the evolution of the mapping relationship. The mapping relationship data finally generated ensures the effectiveness and intelligence of the deep learning model in the automatic generation of equalizer parameters, and improves the accuracy and practicality of audio processing.

[0029] Preferably, step S244 includes the following steps:

[0030] Extract key transformation points of feature-parameter variation law to obtain key transformation points of variation law;

[0031] Based on the key transformation points of the change law, the time series relationship of the feature-parameter change law is established to obtain the law decomposition set;

[0032] Identify the corresponding points of feature-parameter conversion on the regular decomposition set to obtain mapping corresponding data; perform conversion function mapping on the mapping corresponding data to generate a mapping conversion function;

[0033] The mapping relationship is deconstructed based on the mapping conversion function to generate mapping relationship data.

[0034] The present invention can identify important nodes in the change process by extracting key conversion points of feature-parameter change rules. The generated key conversion points of change rules provide key basis for the subsequent establishment of time series relationship. The implementation of establishing time series relationship of feature-parameter change rules based on key conversion points of change rules can construct a complete change process. The obtained rule decomposition set provides structured data support for further analysis. The implementation of feature-parameter conversion corresponding point identification of rule decomposition set can clarify the specific correspondence between different features and parameters. The generated mapping corresponding data provides a basis for subsequent conversion function mapping. The implementation of conversion function mapping of mapping corresponding data can express complex change rules in the form of mathematical functions. The generated mapping conversion function provides clear operational guidance for parameter adjustment and optimization. The implementation of mapping relationship deconstruction based on mapping conversion function can deeply analyze the internal structure of mapping relationship. The generated mapping relationship data ensures the accuracy and effectiveness of deep learning model in automatic generation of equalizer parameters, and improves the intelligent level and application value of audio processing.

[0035] Preferably, step S3 comprises the following steps:

[0036] Step S31: performing dimension deconstruction processing on the mapping relationship data to obtain input and output dimension data; constructing a network structure for the mapping relationship data based on the input and output dimension data to obtain an initial network structure model;

[0037] Step S32: Analyze the data flow characteristics of the initial network structure model to obtain the data flow characteristic distribution of each layer; optimize the inter-layer data transmission of the initial network structure model according to the data flow characteristic distribution of each layer to generate the relationship model structure data;

[0038] Step S33: evaluating the inter-layer transfer efficiency of the relational model structure data to obtain inter-layer transfer efficiency data; performing skip connection optimization simulation on the relational model structure data based on the inter-layer transfer efficiency data to obtain optimized skip connection data;

[0039] Step S34: performing residual structure optimization simulation on the relationship model structure data based on the inter-layer transfer efficiency data to generate optimized residual structure data;

[0040] Step S35: Update the inter-layer links of the relational model structure data according to the optimized jump connection data and the optimized residual structure data to generate optimized model structure data.

[0041] The present invention can effectively extract the key features between input and output by implementing dimensional deconstruction processing on the mapping relationship data. The generated input and output dimension data provides a clear basis for the subsequent network structure construction. The implementation of the network structure construction of the mapping relationship data based on the input and output dimension data can construct a preliminary network structure model, which lays a framework for the optimization of the deep learning model. The implementation of data flow feature analysis on the initial network structure model can identify the data flow characteristics between layers. The distribution of data flow characteristics of each layer provides data support for subsequent optimization. The implementation of inter-layer data transmission optimization of the initial network structure model according to the distribution of data flow characteristics of each layer can achieve efficient data flow. The generated relationship model structure data provides a guarantee for the performance improvement of the deep learning model. The implementation of inter-layer transmission efficiency evaluation of the relationship model structure data can quantify the data transmission efficiency between different layers. The obtained inter-layer transmission efficiency data provides a basis for the formulation of optimization strategies. The implementation of jump connection optimization simulation of the relationship model structure data based on the inter-layer transmission efficiency data can improve the flow path of information. The generated optimized jump connection data improves the expression ability of the model. The implementation of residual structure optimization simulation of relational model structure data based on inter-layer transmission efficiency data can reduce information loss. The generated optimized residual structure data enhances the robustness of the model. The implementation of inter-layer link update of relational model structure data according to optimized jump connection data and optimized residual structure data can ultimately form efficient optimized model structure data, thereby improving the overall performance and practicality of deep learning in the automatic generation of equalizer parameters.

[0042] Preferably, step S33 includes the following steps:

[0043] Performing hierarchical division on the relational model structure data to obtain hierarchical feature data; performing transmission path analysis on the hierarchical feature data to generate transmission path data;

[0044] Performing connected bottleneck identification on the transmission path data to obtain bottleneck feature data; performing gradient flow analysis on the bottleneck feature data to generate a gradient flow graph;

[0045] Calculate the transmission delay of the gradient flow graph to generate the transmission delay matrix; perform bandwidth utilization analysis on the transmission delay matrix to generate inter-layer transmission efficiency data;

[0046] Identify the key paths of the inter-layer transfer efficiency data to obtain path feature data; screen the connection points of the path feature data to generate a candidate connection point set;

[0047] Performing skip pattern design on the candidate connection point set to obtain candidate skip connection data; performing connection simulation on the candidate skip connection data to generate simulated skip connection results;

[0048] The relational model structure data is optimized and integrated based on the simulated skip connection results to generate optimized skip connection data.

[0049] The present invention can clearly identify the features of different levels by implementing hierarchical division of relational model structure data, and the generated hierarchical feature data provides a structured basis for subsequent analysis. The implementation of transmission path analysis on hierarchical feature data can depict the specific path of information transmission, and the generated transmission path data provides an important basis for optimization. The implementation of connected bottleneck identification on transmission path data can reveal the bottleneck position in the network, and the obtained bottleneck feature data points out the direction for subsequent optimization. The implementation of gradient flow analysis on bottleneck feature data can understand the flow of gradients in the network, and the generated gradient flow graph provides visual support for optimization adjustment. The implementation of transmission delay calculation on the gradient flow graph can quantify the delay in the information transmission process, and the generated transmission delay matrix provides necessary data for analyzing bandwidth utilization. The implementation of bandwidth utilization analysis on the transmission delay matrix can identify the efficiency differences between different layers, and the generated inter-layer transmission efficiency data provides a basis for optimization decision-making. The implementation of critical path identification on inter-layer transmission efficiency data can determine the critical path that affects efficiency, and the obtained path feature data provides a reference for subsequent connection point screening. The implementation of connection point screening on path feature data can identify potential optimization points. The generated candidate connection point set provides a basis for jump connection design. The implementation of jump mode design on the candidate connection point set can create an efficient information flow plan. The obtained candidate jump connection data provides the necessary conditions for connection simulation. The implementation of connection simulation on the candidate jump connection data can verify the effectiveness of jump connections. The generated simulated jump connection results provide an empirical basis for optimization and integration. The implementation of optimizing and integrating relational model structure data based on simulated jump connection results can form efficient optimized jump connection data, thereby improving the performance and efficiency of deep learning models in automatic generation of equalizer parameters.

[0050] Preferably, step S4 comprises the following steps:

[0051] Step S41: collecting new audio samples; extracting features from the new audio samples to obtain new audio feature data;

[0052] Step S42: performing a loading simulation on the optimization model structure data to obtain a simulated loading optimization model; performing a back propagation calculation on the simulated loading optimization model to obtain a predicted loss value;

[0053] Step S43: performing gradient update processing on the optimized model structure data according to the predicted loss value to generate a deep audio equalization parameter model;

[0054] Step S44: Use the deep audio equalization parameter model to perform equalization parameter inference and prediction on the new audio sample to obtain a candidate equalizer parameter set.

[0055] The present invention ensures the timeliness and relevance of data by collecting new audio samples. The new audio feature data obtained after feature extraction provides a specific basis for the application of the model. The implementation of loading simulation on the optimized model structure data can effectively combine the new audio features with the existing model. The generated simulated loading optimization model lays the foundation for subsequent reasoning calculations. The implementation of back propagation calculation on the simulated loading optimization model can evaluate the performance of the model. The obtained predicted loss value provides a quantitative indicator for the optimization direction. The implementation of gradient updating processing on the optimized model structure data according to the predicted loss value can adjust the model parameters. The generated deep audio equalization parameter model provides precise adjustment for audio processing to adapt to different audio sample characteristics. The implementation of equalization parameter reasoning prediction of new audio samples using the deep audio equalization parameter model can generate a candidate equalizer parameter set, ensuring that the processing of new audio data is highly adaptable and flexible, improving the application effect of deep learning in the field of audio equalization, and enhancing the practicality and effectiveness of the model in practical applications.

[0056] Preferably, step S5 comprises the following steps:

[0057] Step S51: performing correspondence analysis on the candidate equalizer parameter set and the new audio sample to obtain parameter correspondence data; performing deviation calculation on the candidate equalizer parameter set and the new audio sample according to the correspondence data to obtain parameter difference data;

[0058] Step S52: identifying the deviation distribution law of the parameter difference data and generating the deviation distribution law;

[0059] Step S53: performing difference correction processing based on the deviation distribution law to obtain parameter correction data;

[0060] Step S54: performing feedback compensation on the candidate equalizer parameter set according to the parameter correction data to generate final equalizer parameters.

[0061] The present invention can clarify the relationship between different parameters and audio samples by implementing the correspondence analysis on the candidate equalizer parameter set and the new audio sample, and the generated parameter correspondence data provides a basis for the subsequent deviation calculation. The implementation of the deviation calculation on the candidate equalizer parameter set and the new audio sample according to the correspondence data can quantify the difference between the parameters. The obtained parameter difference data lays the foundation for the subsequent correction processing. The implementation of the deviation distribution law identification on the parameter difference data can reveal the distribution characteristics of the parameter difference, and the generated deviation distribution law provides a scientific basis for the difference correction. The implementation of the difference correction processing based on the deviation distribution law can effectively adjust the parameters to make them more in line with the actual needs of the audio sample. The obtained parameter correction data provides optimization support for the final parameter generation. The implementation of feedback compensation for the candidate equalizer parameter set according to the parameter correction data can ensure that the generated final equalizer parameters have higher adaptability and accuracy, improve the practicality and effect of the deep learning model in the field of audio equalization, and ensure that the final output equalizer parameters can better meet the actual audio processing needs.

[0062] Preferably, step S54 includes the following steps:

[0063] Perform frequency band decomposition on the parameter correction data to obtain a frequency band feature set; perform parameter mapping on the frequency band feature set to generate a frequency band mapping matrix;

[0064] Perform compensation calculation on the frequency band mapping matrix to obtain a compensation coefficient set;

[0065] Perform boundary constraint processing on the compensation coefficient set to generate a constraint parameter set; perform gradient calculation on the constraint parameter set to generate a gradient feature set;

[0066] Optimizing parameters of the gradient feature set is calculated to obtain an optimized parameter set; and adjusting parameters of the candidate equalizer parameter set is performed based on the optimized parameter set to generate final equalizer parameters.

[0067] The present invention can separate different frequency components of the audio signal by implementing frequency band decomposition on parameter correction data. The obtained frequency band feature set provides detailed frequency information for subsequent parameter mapping. The implementation of parameter mapping on the frequency band feature set can associate the characteristics of different frequency bands with equalizer parameters. The generated frequency band mapping matrix lays the foundation for compensation calculation. The implementation of compensation calculation on the frequency band mapping matrix can quantify the influence of each frequency band on the equalizer parameters. The obtained compensation coefficient set provides an important basis for subsequent constraint processing. The implementation of boundary constraint processing on the compensation coefficient set can ensure that the generated parameters are within a reasonable range. The generated constraint parameter set provides a guarantee for the stability of the model. The implementation of gradient calculation on the constraint parameter set can identify the direction and amplitude of parameter adjustment. The generated gradient feature set provides key data for subsequent optimization. The implementation of optimized parameter calculation on the gradient feature set can achieve fine adjustment of parameters. The obtained optimized parameter set provides a scientific basis for the adjustment of the candidate equalizer parameter set. The implementation of parameter adjustment of the candidate equalizer parameter set based on the optimized parameter set can ensure that the final equalizer parameters generated are improved in performance and adaptability, enhance the application effect and practicality of the deep learning model in the field of audio equalization, and ensure that the final output equalizer parameters can better meet user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A schematic diagram of the steps of a method for automatically generating audio equalizer parameters through deep learning;

[0069] Figure 2 for Figure 1 Detailed implementation steps of step S2 in the flowchart;

[0070] Figure 3 for Figure 1 Detailed implementation steps of step S3 in FIG.

[0071] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0072] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0073] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0074] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0075] To achieve this, please refer to Figures 1 to 3 , a deep learning method for automatically generating audio equalizer parameters, comprising the following steps:

[0076] Step S1: Collect multi-source audio data; perform short-time Fourier transform on the multi-source audio data to obtain an audio feature vector; perform principal component analysis on the audio feature vector to generate a dimensionality reduction feature set;

[0077] Step S2: obtaining reference template data; performing parameter space modeling on the reference template data to obtain a parameter space model; performing correspondence analysis on the dimension reduction feature set based on the parameter space model to generate mapping relationship data;

[0078] Step S3: constructing a deep learning network for the mapping relationship data to generate relationship model structure data; optimizing the inter-layer links of the relationship model structure data to generate optimized model structure data;

[0079] Step S4: collecting new audio samples; performing iterative model training based on the optimized model structure data to generate a deep audio equalization parameter model; using the deep audio equalization parameter model to perform parameter prediction on the new audio samples to obtain a candidate equalizer parameter set;

[0080] Step S5: performing difference correction processing on the candidate equalizer parameter set and the new audio sample to obtain parameter correction data; performing feedback equalization optimization on the candidate equalizer parameter set according to the parameter correction data to generate final equalizer parameters.

[0081] The present invention collects multi-source audio data and extracts audio feature vectors through short-time Fourier transform, so that the audio data is accurately quantified in the frequency domain characteristics. At the same time, the audio feature vector is reduced in dimension through principal component analysis to generate a reduced-dimensional feature set, which effectively reduces data redundancy and dimensional complexity, improves the calculation efficiency and analysis accuracy in the subsequent processing process, obtains reference template data and performs parameter space modeling, establishes a parameter space model, and realizes a systematic description of multi-dimensional audio features. At the same time, based on the parameter space model, the corresponding relationship analysis of the reduced-dimensional feature set is performed to generate mapping relationship data, so that a high-correlation mapping is formed between the multi-dimensional features and the template, and the logical coherence and accuracy of the mapping relationship data are ensured. By building a deep learning network for the mapping relationship data, relationship model structure data is generated, and a multi-layer neural network is established to deeply model the nonlinear relationship of the mapping data. The inter-layer links are optimized, and the optimized model structure data is generated, which significantly improves the convergence speed and parameter expression ability of the network model, ensures the efficiency and robustness of the deep model, collects new audio samples and performs iterative model training based on the optimized model structure data, generates a deep audio equalization parameter model, and realizes the adaptive ability of the equalization parameter model to new audio samples. Through the generation of candidate equalizer parameter sets, a parametric expression of the multi-dimensional features of audio samples is formed to ensure the diversity and comprehensiveness of the equalization parameters. The candidate equalizer parameter sets and new audio samples are differentially corrected to obtain parameter correction data, so that the deviation of the equalizer parameters is controlled within an acceptable range. At the same time, the final equalizer parameters are generated through feedback equalization optimization to ensure the precise adaptability of the equalizer parameters to the new audio samples, so that the finally generated parameters have high stability and consistency, and improve the delicacy of the audio equalization effect and the satisfaction of the user experience.

[0082] In the embodiment of the present invention, refer to Figure 1 , is a schematic diagram of the steps of a method for automatically generating audio equalizer parameters by deep learning of the present invention. In this example, the method for automatically generating audio equalizer parameters by deep learning includes the following steps:

[0083] Step S1: Collect multi-source audio data; perform short-time Fourier transform on the multi-source audio data to obtain an audio feature vector; perform principal component analysis on the audio feature vector to generate a dimensionality reduction feature set;

[0084] In this embodiment, multi-source audio data is collected, a high-sensitivity microphone array is used, and a device with a bandwidth ranging from 20 Hz to 20 kHz is selected to synchronously collect data from multiple sound sources. A high-precision analog-to-digital converter (ADC) is used to convert the analog signal into a digital signal. The sampling rate is set to 48 kHz and the quantization accuracy is 24 bits. Then, the collected digital audio data is subjected to frequency domain analysis by applying a short-time Fourier transform (STFT). The window length is set to 1024, the window type is a Hanning window, and the overlap rate is 50%. The frequency distribution is calculated, and the amplitude spectrum and phase spectrum of the frequency component are extracted to obtain an audio feature vector. The feature vector is subjected to dimensionality reduction processing by a principal component analysis (PCA) method, and a cumulative contribution rate of 99% is retained to generate a dimensionality reduction feature set.

[0085] Step S2: obtaining reference template data; performing parameter space modeling on the reference template data to obtain a parameter space model; performing correspondence analysis on the dimension reduction feature set based on the parameter space model to generate mapping relationship data;

[0086] In this embodiment, reference template data is selected through an audio database, a standardized audio sample set is selected, and the reference data is grouped using the K-Means Clustering method to generate a preliminary template set. Subsequently, parameter space modeling is used to construct a three-dimensional parameter space model including frequency, time and amplitude for each template sample. The parameter space modeling process uses a support vector machine (SVM) algorithm to establish a classification hyperplane based on the frequency domain features of each sample. Based on the model, a correspondence analysis is performed on the dimensionality reduction feature set generated in step S1, and the cosine similarity is used to calculate the degree of matching between the feature set and the template data to generate mapping relationship data.

[0087] Step S3: constructing a deep learning network for the mapping relationship data to generate relationship model structure data; optimizing the inter-layer links of the relationship model structure data to generate optimized model structure data;

[0088] In this embodiment, the mapping relationship data is used as the training set input to construct a deep learning network model, and a convolutional neural network (CNN) structure is selected, which includes 3 convolutional layers, 2 fully connected layers and 1 output layer. The kernel size of the convolutional layer is set to 3×3, the step size is 1, and the activation function uses ReLU (Rectified Linear Unit). The network parameters are initialized and the learning rate is set to 0.001. Then, the links between the network layers are optimized, and the back propagation algorithm (BP) is used to adjust the weights and bias values. The model is optimized by minimizing the mean square error (MSE) loss function, and finally the optimized model structure data is generated.

[0089] Step S4: collecting new audio samples; performing iterative model training based on the optimized model structure data to generate a deep audio equalization parameter model; using the deep audio equalization parameter model to perform parameter prediction on the new audio samples to obtain a candidate equalizer parameter set;

[0090] In this embodiment, new audio samples are collected, and the same collection equipment and processing parameters are used to input the samples into the optimized model structure data for iterative model training. During the training process, stochastic gradient descent (SGD) is used for optimization. The number of training rounds is set to 100, and the batch size of each round is set to 32. The convergence is ensured by dynamically adjusting the learning rate to generate a deep audio equalization parameter model. The model is then used to predict new audio samples, and preliminary equalization parameters are obtained based on the model output. The weight distribution is adjusted in combination with the optimal parameter search strategy (such as the grid search method) to obtain a candidate equalizer parameter set.

[0091] Step S5: performing difference correction processing on the candidate equalizer parameter set and the new audio sample to obtain parameter correction data; performing feedback equalization optimization on the candidate equalizer parameter set according to the parameter correction data to generate final equalizer parameters.

[0092] In this embodiment, a difference correction process is performed on the candidate equalizer parameter set and the new audio sample, and an audio feature alignment algorithm based on Lagrange Interpolation is adopted to perform linear interpolation adjustment on the equalization parameter set to correct errors and ensure continuity. After processing, parameter correction data is generated, and feedback equalization optimization is performed on the candidate equalizer parameter set based on the parameter correction data. The particle swarm optimization (PSO) method is adopted to perform multi-dimensional parameter search by adjusting the particle position and speed to obtain the globally optimal equalization parameter combination, and finally the equalizer parameters are generated.

[0093] Preferably, step S1 comprises the following steps:

[0094] Step S11: collecting multi-source audio data; performing time-frequency domain conversion on the multi-source audio data to obtain an initial spectrum diagram;

[0095] Step S12: performing frame processing on the initial spectrum graph to generate a frame sequence; performing short-time Fourier transform on the frame sequence to obtain an audio feature vector;

[0096] Step S13: normalizing the audio feature vector to obtain a standardized feature matrix;

[0097] Step S14: Calculate statistical parameters of the standardized feature matrix to obtain a statistical feature set; perform principal component analysis on the statistical feature set to generate a reduced-dimensional feature set.

[0098] In this embodiment, a multi-channel microphone array is used to simultaneously collect audio signals from multiple sound sources. A high-sensitivity device with a frequency response range of 20 Hz to 20 kHz is selected, the sampling rate is set to 48 kHz, and the quantization accuracy is set to 24 bits. The collected audio signals are converted into digital signals using an analog-to-digital converter (ADC). After the collection is completed, the digital signals are converted into time-frequency domains and wavelet transform is used. Transform) method, set the mother wavelet function to Daubechies wavelet, decompose the number of layers to 6, generate the time-frequency spectrum of the corresponding frequency components, save the obtained spectrum in a two-dimensional matrix data format, divide the time-frequency spectrum data into frames, set the length of each frame to 1024 points, and the overlap rate between frames to 50%. Use the Hanning window to weight each frame of data to reduce the influence of spectrum leakage, generate a frame sequence, perform short-time Fourier transform on the frame sequence, set the Fourier transform length to 1024 points, extract the spectrum amplitude and phase information, and combine the results into an audio feature vector. The feature vector is stored in a matrix form, each column represents a frequency component, and each row represents a time frame sequence. The audio feature vector is normalized, and the average value of each column feature is calculated. And standard deviation, each feature is standardized and adjusted to make its distribution range relatively balanced, and the normalized data is saved as a standardized feature matrix. The matrix adopts a sparse storage format to improve the subsequent calculation efficiency and reduce the storage space occupancy, laying a data foundation for the next step of operation. The statistical parameters of the standardized feature matrix, including mean, variance, skewness and kurtosis, are calculated to describe the main characteristics of the feature distribution. The statistical results are saved in the form of a feature set, and the statistical feature set is input into the principal component analysis tool. By analyzing the covariance matrix of the feature matrix, the most representative principal component is extracted. The selection of the principal component is based on the cumulative contribution rate meeting the set threshold, and finally a reduced dimensionality feature set is generated. The feature data after dimensionality reduction is saved in the form of a structured file for subsequent deep learning modeling and training.

[0099] Preferably, step S2 comprises the following steps:

[0100] Step S21: Acquire reference template data; identify the data type of the reference template data to obtain template type data;

[0101] Step S22: performing parameter space modeling on the reference template data according to the template type data to obtain a parameter space model;

[0102] Step S23: performing virtual space projection on the dimension reduction feature set based on the parameter space model to obtain feature projection data; performing feature matching analysis on the feature projection data to generate a feature parameter matching set;

[0103] Step S24: capturing the change trend of the feature parameter matching set to obtain the feature-parameter change law; performing mapping relationship evolution on the feature-parameter change law to generate mapping relationship data.

[0104] In this embodiment, a database management system is used to read the stored reference template data, the template data includes an audio feature template and a spectrum response template, a data analysis tool is used to identify the data type of the read reference template data, the data is classified and identified according to the structural characteristics and content characteristics of the data, the identification method adopts a classification algorithm based on regularized feature matching, the identification result includes the category identification and specific parameter attributes of the template data, the identified template type data is stored as a structured record file, the file contains the name of the template, the classification label and the main parameter value, according to the identified template type data, the reference template data is parameter space modeled by a parametric modeling tool, a custom function fitting method is used to build a parameter space model, the template parameter data is fitted and solved by the least squares method during the model building process, the dimension of the parameter space is automatically adjusted according to the category label of the template, the low-dimensional parameter template is represented by a two-dimensional matrix, the high-dimensional parameter template is stored using a multi-dimensional tensor, the modeling result is stored in the form of a binary file for subsequent virtual space projection processing, the virtual projection algorithm is used to interact with the dimension reduction feature set and the parameter space model, the projection result of the dimension reduction feature set in the virtual space is calculated based on the parameter space model, and the virtual space The dimension setting between the two is consistent with the parameter space model. The linear mapping method is used in the projection operation. The coordinates of the feature data in the virtual space are calculated by matrix multiplication to generate feature projection data. The feature projection data is stored in the form of a sparse matrix. Then, feature matching analysis is performed on the feature projection data and the original data in the parameter space model. The feature matching analysis uses the cosine similarity calculation method to sort the similarity between the projection points and the reference template points to generate a feature parameter matching set. The matching set is saved in a table form, including the projection point index, the reference template point index and the similarity score. A time series analysis is performed on the generated feature parameter matching set. A sliding window is used to capture the change trend of the parameters in the matching set. The sliding window length is set to 10 data points. The second-order difference method is used to calculate the change rate of the parameter value. The feature-parameter change law is generated according to the change trend. The change law is stored in the form of a line chart data. Then, the mapping relationship evolution of the feature-parameter change law is performed. A mapping algorithm based on time weight is used. The weight value is allocated in descending order according to time. The mapping algorithm associates the historical change law with the current data to generate mapping relationship data. The mapping relationship data is stored in the form of a nested dictionary, including a timestamp, a feature parameter change curve and its weight associated value.

[0105] Preferably, step S24 comprises the following steps:

[0106] Step S241: performing time series decomposition on the feature parameter matching set to obtain a time series feature sequence; performing cluster analysis on the time series feature sequence to generate trend category data;

[0107] Step S242: performing pattern recognition on the trend category data to obtain basic pattern data; performing similarity calculation based on the basic pattern data to generate inter-pattern association data;

[0108] Step S243: Summarize the regularity of the inter-pattern correlation data to obtain regularity index data; capture the change trend based on the regularity index data to generate the feature-parameter change regularity;

[0109] Step S244: Analyze the corresponding relationship of the feature-parameter variation rule to obtain mapping corresponding data; and evolve the mapping relationship according to the mapping corresponding data to generate mapping relationship data.

[0110] In this embodiment, a time series analysis tool is used to perform time series decomposition on a feature parameter matching set, and the data in the matching set is arranged in chronological order. The data is converted from time domain to frequency domain through Fourier transform to obtain a time series feature sequence. The time series feature sequence is subdivided through a time-frequency analysis module, and the decomposition results are stored as multiple frequency bands, each frequency band corresponding to a different frequency range. Next, a clustering algorithm is applied to the time series feature sequence. The clustering method uses a K-means clustering algorithm, and the number of clusters is set to 5. In the clustering analysis, clustering is performed according to the similarity between the feature values ​​to generate trend category data. The data structure is a cluster number and an index of each type of data point. The trend category data is used for subsequent pattern recognition and similarity calculation. Based on the trend category data, a pattern recognition algorithm based on template matching is used for analysis. First, a pattern recognition model is constructed. The model uses an adaptive gradient down-conversion algorithm. The algorithm is optimized by training the historical pattern data to obtain the basic pattern data. The basic pattern data contains the spectrum characteristics and change rules of each category of data. The basic pattern data structure is in vector form. Each vector contains frequency components, amplitude changes and time series. Then the similarity between the pattern data of each category is calculated. The cosine similarity algorithm is used to calculate the similarity between each pair of pattern data to generate pattern association data. The associated data is stored in the form of key-value pairs, where the key is the pattern identifier and the value is the similarity score with other patterns. The data is smoothed and the sliding average algorithm is used for time window processing. The window size is 15 data points. The smoothed data is used for pattern recognition. Next, the smoothed data is summarized by an algorithm based on heuristic rules to summarize the regular pattern. The regular pattern adopts a method based on dynamic time warping (Dynamic The similarity matching technology of Time Warping (DTW) is used for multiple iterative optimizations to finally obtain the regularity index data. The timestamp and change trend of each regularity in the regularity index data are marked, and groups are generated according to the change trend. Based on the regularity index data, the trend capture model is used to analyze the change trend of the data. The trend capture model adopts the support vector machine (SVM) algorithm. Historical data is used as the training set during the training process. The features defined in the model include the time span of the regularity, the volatility of the frequency response, etc. The capture process is updated in real time through a dynamic window with a time resolution of 10 seconds. The captured change trend is stored as feature-parameter change regularity data, which is used for subsequent corresponding relationship analysis. Next, according to the feature-parameter change regularity data, the mapping relationship analysis algorithm based on graph theory is applied. By analyzing the relationship between the rules, the mapping corresponding data is obtained. The mapping relationship data is each rule and its corresponding parameter change path, presented in the form of a chart, and the mapping relationship data is used for the training of the deep learning model.

[0111] Preferably, step S244 includes the following steps:

[0112] Extract key transformation points of feature-parameter variation law to obtain key transformation points of variation law;

[0113] Based on the key transformation points of the change law, the time series relationship of the feature-parameter change law is established to obtain the law decomposition set;

[0114] Identify the corresponding points of feature-parameter conversion on the regular decomposition set to obtain mapping corresponding data; perform conversion function mapping on the mapping corresponding data to generate a mapping conversion function;

[0115] The mapping relationship is deconstructed based on the mapping conversion function to generate mapping relationship data.

[0116] In this embodiment, the change law data is discretized on the time axis, a fixed time window of 20ms is selected, and the change law data is divided into multiple time periods. In each time period, a peak detection algorithm is applied, and the peak threshold is set to twice the standard deviation, and the key conversion point of the change law is extracted. The conversion point refers to the time point when the characteristic changes suddenly. The steep point of the characteristic change is detected by derivative calculation to obtain the key conversion point data. The conversion point data is stored as a timestamp and its corresponding characteristic value change position. According to the extracted key conversion point, the change law data is reorganized. In the process of reorganization, the time series analysis method is used, and ARIMA (autoregressive integrated moving average model) is selected to model the time series. The previously trained data set is used as the initial condition. The change trend of each time point is predicted by the model, and the correlation between the time points is established, and a time series relationship diagram is generated. The time series relationship diagram shows the flow of the law on the time axis. The result of the time series relationship analysis output is a law decomposition set, which contains the evolution process of each law from one time point to another. Based on the change trend of the known law, the corresponding points of the parameters of each decomposed law sequence are matched, and dynamic time warping (Dynamic Time Warping) is used. The DTW algorithm is used for matching. The DTW algorithm calculates the similarity of two time series and identifies the corresponding points between the change patterns. Furthermore, the similarity of the conversion points is screened during the identification process, and the points with an error less than 0.05 are selected as the final corresponding points. The obtained mapping data is stored in the form of coordinates, recording the corresponding relationship between feature changes and parameter changes. The mapping function is established using the identified mapping corresponding points. The mapping function is fitted using the least squares method, and a polynomial regression model is selected with a fitting degree of 95%. The parameter changes corresponding to the feature changes are mapped to mathematical functions. The conversion function uses first-order, second-order, and third-order polynomials for fitting in this process to ensure the accuracy and effectiveness of the conversion. The converted function data is stored in the form of a formula. Each mapping corresponding point will be processed by the conversion function to generate mapping conversion function data. The generated mapping conversion function data is structured and analyzed. The conversion function is deconstructed using the graph structure of graph theory to analyze the key variables and parameter relationships in the conversion function. The graph convolutional network (Graph Convolutional Network) in deep learning is used. The GCN model performs convolution operations on each node through the adjacency matrix to extract the feature information of the mapping relationship and finally generates mapping relationship data. The mapping relationship data is structured and stored as a mapping matrix, which records the association between features and parameters.

[0117] Preferably, step S3 comprises the following steps:

[0118] Step S31: performing dimension deconstruction processing on the mapping relationship data to obtain input and output dimension data; constructing a network structure for the mapping relationship data based on the input and output dimension data to obtain an initial network structure model;

[0119] Step S32: Analyze the data flow characteristics of the initial network structure model to obtain the data flow characteristic distribution of each layer; optimize the inter-layer data transmission of the initial network structure model according to the data flow characteristic distribution of each layer to generate the relationship model structure data;

[0120] Step S33: evaluating the inter-layer transfer efficiency of the relational model structure data to obtain inter-layer transfer efficiency data; performing skip connection optimization simulation on the relational model structure data based on the inter-layer transfer efficiency data to obtain optimized skip connection data;

[0121] Step S34: performing residual structure optimization simulation on the relationship model structure data based on the inter-layer transfer efficiency data to generate optimized residual structure data;

[0122] Step S35: Update the inter-layer links of the relational model structure data according to the optimized jump connection data and the optimized residual structure data to generate optimized model structure data.

[0123] In this embodiment, mapping relationship data is obtained, which includes multidimensional relationships between features and parameters, and principal component analysis (Principal Component Analysis,The PCA algorithm is used to reduce the dimension of the data, and the first 80% of the variance is selected as the retained data. The data after dimensionality reduction is divided into input and output dimensions, where the input data dimension represents the feature space, and the output data dimension represents the parameter space, to obtain input and output dimension data. Based on these data, a preliminary framework of the neural network model is constructed, where the number of input layer nodes is equal to the dimension of the input data, and the number of output layer nodes is equal to the dimension of the output data, to obtain the initial network structure model. The initial network includes an input layer, a hidden layer, and an output layer. The data flow characteristics of each layer are defined as the ratio of the output to the input of the layer. The neural network back propagation algorithm (Backpropagation) is used to train and test each layer, and the data flow characteristic distribution of each layer is analyzed. Specifically, the relationship between the activation function output of each layer and the input layer nodes is calculated. Based on the data, a data flow characteristic graph is drawn to obtain a flow characteristic distribution graph of each layer. Based on the graph , determine the feature transfer relationship between each layer, and further optimize the inter-layer data transfer based on this result. The optimization goal is to reduce the redundant information transmission of the middle layer and improve the data transmission efficiency. The relational model structure data is obtained, and the transfer efficiency of each layer is calculated, that is, the effective amount of information input and output of each layer. By analyzing the activation function output, gradient transfer and training error of each layer, the efficiency of each layer in the overall model is evaluated. The information entropy (Entropy) theory is used to calculate the amount of information transmitted in each layer to obtain the inter-layer transfer efficiency data. The inter-layer transfer efficiency data contains the transfer efficiency value of each layer. Based on this data, the relational model structure is optimized by skip connection simulation, using the skip connection technology (SkipConnection), that is, directly connecting some layers to bypass the middle layer and reduce information loss. The optimized skip connection data is obtained. The residual network structure is improved by introducing the residual connection (Residual Connection) solves the degradation problem that occurs during information transmission. Residual connections are introduced in each layer of the model so that the input signal directly bypasses certain layers and is added to the output of the subsequent layers after the residual connection to optimize the training efficiency of the model. In specific implementation, the residual points of each layer are identified based on the transmission efficiency data, and the layers with lower transmission efficiency are selected for residual connection. The specific implementation of the residual connection adopts the addition operation, and the new output is formed by adding it to the original input signal to obtain the optimized residual structure data. Combined with the optimized skip connection and residual structure, the network is updated with the deep learning framework (such as TensorFlow or PyTorch) to update the inter-layer connection. Skip connections and residual connections are introduced in each layer, and the data is locally weighted by convolution operation. Global optimization is performed in the network, the inter-layer link data is updated, and the optimized model structure data is generated. The structure contains skip connections, residual structures and updated information of each layer. Through the updated network structure, the automatic generation capability of equalizer parameters is further improved. ,

[0124] Preferably, step S33 includes the following steps:

[0125] Performing hierarchical division on the relational model structure data to obtain hierarchical feature data; performing transmission path analysis on the hierarchical feature data to generate transmission path data;

[0126] Performing connected bottleneck identification on the transmission path data to obtain bottleneck feature data; performing gradient flow analysis on the bottleneck feature data to generate a gradient flow graph;

[0127] Calculate the transmission delay of the gradient flow graph to generate the transmission delay matrix; perform bandwidth utilization analysis on the transmission delay matrix to generate inter-layer transmission efficiency data;

[0128] Identify the key paths of the inter-layer transfer efficiency data to obtain path feature data; screen the connection points of the path feature data to generate a candidate connection point set;

[0129] Performing skip pattern design on the candidate connection point set to obtain candidate skip connection data; performing connection simulation on the candidate skip connection data to generate simulated skip connection results;

[0130] The relational model structure data is optimized and integrated based on the simulated skip connection results to generate optimized skip connection data.

[0131] In this embodiment, the relational model structure data including the hierarchical information of the network is obtained, and the parameter weights and node connection relationships of each layer are analyzed, and the hierarchical clustering algorithm is used to divide the layers in the network into several levels. When dividing, the network layer is divided into several levels according to the input-output relationship, weight distribution and information transmission dependency between the layers. Each level contains multiple network layers with similar data transmission characteristics, and finally the hierarchical feature data is obtained. These data reflect the hierarchical relationship and feature distribution between the layers. Using graph theory, The shortest path algorithm in the Theory (such as Dijkstra algorithm) analyzes the hierarchical feature data and identifies the transmission path from the input layer to the output layer. The path analysis is based on the characteristic flow information between the network layers, analyzes the connection mode between the network layers, and obtains the data of each transmission path. The transmission path data not only includes the node connection information of the path, but also includes the strength of information transmission in each path and its transmission efficiency. The transmission path data is graphically analyzed to find the bottleneck points in the information transmission process. These bottleneck points represent the places where information flows slowly, bandwidth is limited, or node processing capacity is insufficient. The Max-Flow Min-Cut Theorem is used to analyze the bottleneck points. Theorem) algorithm is used to identify the nodes and connections that are most likely to cause blockage or delay in the transmission path, and bottleneck feature data is obtained. The bottleneck feature data includes the key bottleneck nodes that affect network performance and their corresponding transmission paths. The gradient flow of each layer is tracked based on the backpropagation algorithm (Backpropagation), the gradient change of each layer is calculated, and the gradient information is used to further analyze the impact of the bottleneck node on the gradient flow. In the gradient flow analysis, the gradient value change of each layer of nodes is calculated, and the heat map (Heatmap) is used to show the gradient flow trend and bottleneck area to obtain the gradient flow graph. The gradient flow graph reflects the gradient changes between layers and the impact of bottleneck nodes on the overall gradient transmission. The delay value of each layer and the bottleneck point is marked on the gradient flow graph. The delay calculation is based on According to the transmission speed and processing time of each layer, the transmission delay of each layer is estimated by time complexity analysis. The calculation formula is: delay = transmission time + calculation time, where the transmission time is calculated based on the data volume and bandwidth, and the calculation time is evaluated based on the node processing capacity. The transmission delay matrix is ​​generated, which contains the delay data of each layer and bottleneck node. The bandwidth utilization of each layer is analyzed by the bandwidth utilization calculation formula (bandwidth utilization = actual transmission data volume / maximum bandwidth). The delay data in the delay matrix is ​​used to evaluate the bandwidth occupancy of each layer, and the layer with low bandwidth utilization is identified. The model is optimized by bandwidth utilization to obtain the inter-layer transmission efficiency data. The transmission efficiency data includes the bandwidth utilization, delay and information transmission efficiency of each layer. According to the transmission efficiency data,A path selection strategy based on the maximum flow algorithm is used to identify the key paths that are critical to network performance. Key paths refer to those paths with high transmission efficiency and that affect the transmission performance of the entire network. Path feature data is obtained. The path feature data includes the node information, transmission efficiency, and contribution to the overall network performance of the key paths. Based on the path feature data, strong dependencies between connection points are identified. These connection points are key nodes in the network that can accelerate information transmission. The association rules algorithm is used to analyze the relationship between each connection point to obtain a set of candidate connection points. The set of candidate connection points includes all connection points that can improve network performance. Based on the set of candidate connection points, a jump connection pattern in the network is designed using jump connection technology. Jump connection technology directly connects distant nodes to reduce redundant information transmission in the middle layer and constructs multiple jump connection patterns. By using graph search algorithms (such as breadth-first search, Search) to generate different skip connection configurations, obtain candidate skip connection data, and use deep learning frameworks (such as TensorFlow or PyTorch) to simulate skip connections based on the candidate skip connection data. During the simulation, the network structure is adjusted according to the parameters of the skip connection. By training the network, the impact of each skip connection configuration on the model performance is evaluated to obtain the simulated skip connection results. The simulation results show the effect of different skip connection configurations on improving network performance. According to the simulated skip connection results, the optimal skip connection configuration is determined, and the relationship model structure data is adjusted and optimized using the configuration. By updating the inter-layer connections, the optimized network structure is formed. The optimized skip connection data reflects the improvement effect of the network after skip connection optimization.

[0132] Preferably, step S4 comprises the following steps:

[0133] Step S41: collecting new audio samples; extracting features from the new audio samples to obtain new audio feature data;

[0134] Step S42: performing a loading simulation on the optimization model structure data to obtain a simulated loading optimization model; performing a back propagation calculation on the simulated loading optimization model to obtain a predicted loss value;

[0135] Step S43: performing gradient update processing on the optimized model structure data according to the predicted loss value to generate a deep audio equalization parameter model;

[0136] Step S44: Use the deep audio equalization parameter model to perform equalization parameter inference and prediction on the new audio sample to obtain a candidate equalizer parameter set.

[0137] In this embodiment, a high-quality audio acquisition device is used, such as a 24-bit / 192kHz audio interface, to ensure the acquisition accuracy of the audio sample. The acquisition frequency of the audio sample is set to 48kHz to capture sufficient frequency information, and a microphone is used to record a specific audio source. Different audio scenes, such as music, voice, etc., are selected during the acquisition process to ensure the diversity of the audio samples. The environmental noise is processed during the recording process to ensure that the quality of the collected audio samples meets the subsequent processing requirements. After the acquisition is completed, the audio samples are saved in wav format for subsequent feature extraction and analysis. The audio samples are processed using the Mel Frequency Cepstral Coefficients (MFCC) algorithm, and 32 Mel frequency channels are used to decompose the audio signal. The audio signal is divided into 20 millisecond frames, and the MFCC features of each frame are calculated. The frequency domain analysis of each frame of audio is performed through Fourier Transform to extract feature data such as frequency, amplitude, and phase. At the same time, the Short-Time Fourier Transform (STFT) is used FourierTransform) is used to further extract the time-frequency information of the audio, and the new audio feature data is organized into a two-dimensional matrix form as input data for subsequent model training and prediction. The TensorFlow framework is used to load the trained deep learning model, which has been trained with a large amount of audio data. The structure of the model contains multiple convolutional layers and fully connected layers. The pre-trained weight data is used during loading to ensure that the loaded optimization model has good performance. During the loading process, batch normalization is used to standardize the data to ensure that the input data of each layer has zero mean and unit variance, further improving the stability of the model. The process of simulating the loading of the optimization model is completed at this stage, which provides a basis for subsequent back propagation calculations. The loss function of the model is defined as the mean square error (MSE,Mean Squared Error), calculate the error between the prediction result and the target audio feature, use the backpropagation algorithm (Backpropagation) to calculate the gradient of the error, update the weight parameters in the model layer by layer, and use the chain rule (Chain Rule) to calculate the gradient. Starting from the output layer, propagate the error to the input layer layer by layer, and calculate the gradient of each layer. Use the gradient descent method (Gradient Descent) to update the weight. After multiple iterations, the prediction loss value of the model is calculated. The prediction loss value reflects the prediction accuracy of the current model on the audio sample. Based on the gradient calculated by backpropagation, the Adam optimization algorithm (Adaptive MomentEstimation) is used to update the optimized model structure data. The Adam algorithm combines the momentum of the gradient and the square value of the gradient, and can adaptively adjust the learning rate of each parameter. The weight parameters in the model are updated through iterative calculation. The optimized model will adjust the model parameters according to the predicted loss value, thereby reducing the error, so that the optimized deep audio equalization parameter model can more accurately predict the equalizer parameters of the audio sample. The updated optimized model structure data provides accurate parameters for subsequent inference prediction. Using the optimized deep audio equalization parameter model, the new audio feature data is input into the model, and the prediction results of the equalizer parameters are obtained through forward propagation calculation. During the inference process, the model will automatically adjust its internal weight parameters according to the input new audio feature data, and calculate the candidate equalizer parameter set. The candidate equalizer parameter set includes the gain value of each frequency band, frequency response and other information as the basis for the final audio equalization processing. ,

[0138] Preferably, step S5 comprises the following steps:

[0139] Step S51: performing correspondence analysis on the candidate equalizer parameter set and the new audio sample to obtain parameter correspondence data; performing deviation calculation on the candidate equalizer parameter set and the new audio sample according to the correspondence data to obtain parameter difference data;

[0140] Step S52: identifying the deviation distribution law of the parameter difference data and generating the deviation distribution law;

[0141] Step S53: performing difference correction processing based on the deviation distribution law to obtain parameter correction data;

[0142] Step S54: performing feedback compensation on the candidate equalizer parameter set according to the parameter correction data to generate final equalizer parameters.

[0143] In this embodiment, feature data of the new audio sample is matched with the candidate equalizer parameter set, and a dynamic time warping (DTW) algorithm is used for timing alignment. The spectrum feature of the audio sample is compared with the frequency response feature of the candidate equalizer. The DTW algorithm will gradually compare the spectrum feature sequence in the audio sample with the frequency response feature sequence in the equalizer parameter set, calculate the corresponding minimum distance, and determine the equalizer parameters corresponding to each frequency band. Through this process, parameter correspondence data is obtained, which represents the relationship between the feature of the new audio sample and the candidate equalizer parameters. The difference between the equalizer parameters corresponding to each frequency band and the actual audio sample is calculated. The mean square error (MSE) is used as a deviation metric to compare the equalizer parameters of each frequency band with the actual frequency response of the frequency band in the audio sample, and parameter difference data is calculated. The parameter difference data includes the deviation of the gain, frequency, Q value, etc. of each frequency band. The candidate equalizer parameter set is corrected using the calculated error value to reflect the difference with the new audio sample. The Gaussian mixture model (MSE) is used to calculate the parameter difference data. Model, GMM) to cluster the parameter difference data, divide the difference data of each frequency band into multiple categories, identify the distribution law of the difference data, and use the probability distribution function (PDF,Probability Density Function) is used to fit the difference data, analyze the variation trend of the deviation data in each frequency band, and generate the deviation distribution law. The deviation distribution law reflects the typical distribution form of parameter differences in different frequency bands, and helps determine which frequency bands' equalizer parameters need more adjustment. According to the deviation distribution law generated above, a difference correction algorithm is designed. The algorithm makes directional adjustments to the equalizer parameters of each frequency band according to the distribution pattern of the deviations of each frequency band, smoothes the differences between the frequency bands, and uses the weighted average method to correct the difference value, normalize the deviation value, and increase the compensation weight for the frequency bands with larger deviations, so as to realize the correction of the parameters, so as to reduce the difference between the sample and the candidate equalizer parameter set, and generate parameter correction data. The parameter correction data includes the corrected equalizer parameters. According to the corrected parameter data, a feedback compensation algorithm is used. The compensation algorithm adjusts the candidate equalizer parameters according to the error compensation ratio of each frequency band, and adjusts the gain, frequency and Q value of each frequency band accordingly. It makes stronger compensation for the frequency band with larger deviation, and makes slight adjustment for the frequency band with small deviation. It uses the compensation coefficient to make weighted adjustment for each frequency band, and outputs the equalizer parameter set obtained by the final compensation as the final equalizer parameter. Through this process, the equalizer parameters that meet the characteristics of the new audio sample are generated.

[0144] Preferably, step S54 includes the following steps:

[0145] Perform frequency band decomposition on the parameter correction data to obtain a frequency band feature set; perform parameter mapping on the frequency band feature set to generate a frequency band mapping matrix;

[0146] Perform compensation calculation on the frequency band mapping matrix to obtain a compensation coefficient set;

[0147] Perform boundary constraint processing on the compensation coefficient set to generate a constraint parameter set; perform gradient calculation on the constraint parameter set to generate a gradient feature set;

[0148] Optimizing parameters of the gradient feature set is calculated to obtain an optimized parameter set; and adjusting parameters of the candidate equalizer parameter set is performed based on the optimized parameter set to generate final equalizer parameters.

[0149] In this embodiment, the parameter correction data is subjected to a fast Fourier transform (FFT, Fast Fourier Transform) is used to convert the signal from the time domain to the frequency domain to obtain the amplitude and phase information of each frequency band. The signal is decomposed by FFT to obtain a set of frequency band feature sets. Each frequency band corresponds to a specific frequency range, in which the low frequency, medium frequency and high frequency each correspond to different audio features. The selection of the frequency band is set according to the bandwidth of the audio signal and the frequency band setting of the target equalizer. After the frequency band is decomposed, the obtained frequency band feature set contains the amplitude, phase and other parameters of each frequency band. These features will be used for the subsequent equalizer parameter adjustment. Based on each frequency band feature in the frequency band feature set, the linear regression algorithm is used for parameter mapping to map the amplitude and phase information of each frequency band to the corresponding frequency band of the equalizer to generate a frequency band mapping matrix. Each element in the matrix represents the mapping relationship between a specific frequency band feature and an equalizer parameter. During the mapping process, different mapping weights are used for different frequency bands to ensure the accuracy of the frequency band feature and the effectiveness of the equalizer output. The mapped matrix reflects the relationship between the characteristics of each frequency band and the equalizer setting. Based on each frequency band feature in the frequency band mapping matrix, the compensation coefficient is calculated. The calculation is performed by the least squares method. SquaresMethod) is used to compare the characteristics of each frequency band with the expected target to obtain the compensation factor, and weighted compensation is performed on each frequency band. The compensation coefficient set contains the compensation parameters of each frequency band. These compensation coefficients can adjust the output of the equalizer according to the deviation of the current frequency band to correct the difference in the frequency band, so that the audio output is closer to the target effect. The L1 regularization method is used to perform boundary constraints on the compensation coefficient set. This method can ensure that each parameter in the compensation coefficient set does not exceed the set threshold, thereby avoiding over-compensation or unstable parameter values. After the boundary constraint processing, the parameters in the compensation coefficient set will be normalized to generate a constrained parameter set. Based on the constrained parameter set, the backpropagation algorithm is used to calculate the gradient. The gradient of each parameter with respect to the objective function is calculated to determine the influence of each parameter on the final audio output. During the backpropagation process, the gradient information will be propagated through each layer of the neural network and according to the chain rule (Chain Rule) accumulates gradients and finally generates a gradient feature set. The gradient feature set contains the gradient information of each parameter. This information will indicate which parameters need to be further optimized. The gradient feature set is optimized using the gradient descent method. The gradient descent method adjusts the value of each parameter so that the loss function gradually decreases, and finally obtains a set of optimized parameter sets. During the optimization process, the learning rate is used to control the step size of the parameter update to ensure the stability of the parameter update. After the optimization parameter calculation is completed,The obtained optimized parameter set is the equalizer parameter that best suits the current audio sample. According to each optimized value in the optimized parameter set, the candidate equalizer parameters are adjusted. During the adjustment process, the optimal adjustment ratio is calculated by comparing each frequency band feature in the optimized parameter set with the candidate equalizer parameter set. The gain, frequency, Q value and other parameters of the equalizer are adjusted proportionally, and the final equalizer parameters that meet the characteristics of the new audio sample are finally generated. These adjusted parameters will effectively improve the audio output effect.

[0150] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0151] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A deep learning method for automatically generating audio equalizer parameters, characterized in that: The following steps are involved: Step S1: Collect multi-source audio data; perform short-time Fourier transform on the multi-source audio data to obtain an audio feature vector; perform principal component analysis on the audio feature vector to generate a dimensionality reduction feature set; Step S2: Obtain reference template data; Perform parameter space modeling on the reference template data to obtain a parameter space model; perform correspondence analysis on the dimension reduction feature set based on the parameter space model to generate mapping relationship data; Step S3: constructing a deep learning network for the mapping relationship data to generate relationship model structure data; Optimize the inter-layer links of the relational model structure data to generate optimized model structure data; Step S4: Collect new audio samples; Iterative model training is performed based on the optimized model structure data to generate a deep audio equalization parameter model; the deep audio equalization parameter model is used to predict parameters of new audio samples to obtain a candidate equalizer parameter set; Step S5: performing difference correction processing on the candidate equalizer parameter set and the new audio sample to obtain parameter correction data; Feedback equalization optimization is performed on the candidate equalizer parameter set according to the parameter correction data to generate the final equalizer parameters.

2. The method for automatically generating audio equalizer parameters by deep learning according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: collecting multi-source audio data; performing time-frequency domain conversion on the multi-source audio data to obtain an initial spectrum diagram; Step S12: performing frame processing on the initial spectrum graph to generate a frame sequence; performing short-time Fourier transform on the frame sequence to obtain an audio feature vector; Step S13: normalizing the audio feature vector to obtain a standardized feature matrix; Step S14: Calculate statistical parameters of the standardized feature matrix to obtain a statistical feature set; perform principal component analysis on the statistical feature set to generate a reduced-dimensional feature set.

3. The method for automatically generating audio equalizer parameters by deep learning according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: Acquire reference template data; identify the data type of the reference template data to obtain template type data; Step S22: performing parameter space modeling on the reference template data according to the template type data to obtain a parameter space model; Step S23: performing virtual space projection on the dimension reduction feature set based on the parameter space model to obtain feature projection data; performing feature matching analysis on the feature projection data to generate a feature parameter matching set; Step S24: capturing the change trend of the feature parameter matching set to obtain the feature-parameter change law; performing mapping relationship evolution on the feature-parameter change law to generate mapping relationship data.

4. The method for automatically generating audio equalizer parameters by deep learning according to claim 3, characterized in that: Step S24 includes the following steps: Step S241: performing time series decomposition on the feature parameter matching set to obtain a time series feature sequence; performing cluster analysis on the time series feature sequence to generate trend category data; Step S242: performing pattern recognition on the trend category data to obtain basic pattern data; performing similarity calculation based on the basic pattern data to generate inter-pattern association data; Step S243: Summarize the regularity of the inter-pattern correlation data to obtain regularity index data; capture the change trend based on the regularity index data to generate the feature-parameter change regularity; Step S244: Analyze the corresponding relationship of the feature-parameter variation law to obtain mapping corresponding data; and evolve the mapping relationship according to the mapping corresponding data to generate mapping relationship data.

5. The method for automatically generating audio equalizer parameters by deep learning according to claim 4, characterized in that: Step S244 includes the following steps: Extract key transformation points of feature-parameter variation law to obtain key transformation points of variation law; Based on the key transformation points of the change law, the time series relationship of the feature-parameter change law is established to obtain the law decomposition set; Identify the corresponding points of feature-parameter conversion on the regular decomposition set to obtain mapping corresponding data; map the mapping corresponding data with a conversion function to generate a mapping conversion function; The mapping relationship is deconstructed based on the mapping conversion function to generate mapping relationship data.

6. The method for automatically generating audio equalizer parameters by deep learning according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: performing dimension deconstruction processing on the mapping relationship data to obtain input and output dimension data; constructing a network structure for the mapping relationship data based on the input and output dimension data to obtain an initial network structure model; Step S32: Analyze the data flow characteristics of the initial network structure model to obtain the data flow characteristic distribution of each layer; optimize the inter-layer data transmission of the initial network structure model according to the data flow characteristic distribution of each layer to generate the relationship model structure data; Step S33: evaluating the inter-layer transfer efficiency of the relational model structure data to obtain inter-layer transfer efficiency data; performing skip connection optimization simulation on the relational model structure data based on the inter-layer transfer efficiency data to obtain optimized skip connection data; Step S34: performing residual structure optimization simulation on the relationship model structure data based on the inter-layer transfer efficiency data to generate optimized residual structure data; Step S35: Update the inter-layer links of the relational model structure data according to the optimized jump connection data and the optimized residual structure data to generate optimized model structure data.

7. The method for automatically generating audio equalizer parameters by deep learning according to claim 6, characterized in that: Step S33 includes the following steps: Performing hierarchical division on the relational model structure data to obtain hierarchical feature data; performing transmission path analysis on the hierarchical feature data to generate transmission path data; Performing connected bottleneck identification on the transmission path data to obtain bottleneck feature data; performing gradient flow analysis on the bottleneck feature data to generate a gradient flow graph; Calculate the transmission delay of the gradient flow graph to generate the transmission delay matrix; perform bandwidth utilization analysis on the transmission delay matrix to generate inter-layer transmission efficiency data; Identify the key paths of the inter-layer transfer efficiency data to obtain path feature data; filter the connection points of the path feature data to generate a candidate connection point set; Performing skip pattern design on the candidate connection point set to obtain candidate skip connection data; performing connection simulation on the candidate skip connection data to generate simulated skip connection results; The relational model structure data is optimized and integrated based on the simulated skip connection results to generate optimized skip connection data.

8. The method for automatically generating audio equalizer parameters by deep learning according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: collecting new audio samples; extracting features from the new audio samples to obtain new audio feature data; Step S42: performing a loading simulation on the optimization model structure data to obtain a simulated loading optimization model; performing a back propagation calculation on the simulated loading optimization model to obtain a predicted loss value; Step S43: performing gradient update processing on the optimized model structure data according to the predicted loss value to generate a deep audio equalization parameter model; Step S44: Use the deep audio equalization parameter model to perform equalization parameter inference and prediction on the new audio sample to obtain a candidate equalizer parameter set.

9. The method for automatically generating audio equalizer parameters by deep learning according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: performing correspondence analysis on the candidate equalizer parameter set and the new audio sample to obtain parameter correspondence data; performing deviation calculation on the candidate equalizer parameter set and the new audio sample according to the correspondence data to obtain parameter difference data; Step S52: identifying the deviation distribution law of the parameter difference data and generating the deviation distribution law; Step S53: performing difference correction processing based on the deviation distribution law to obtain parameter correction data; Step S54: performing feedback compensation on the candidate equalizer parameter set according to the parameter correction data to generate final equalizer parameters.

10. The method for automatically generating audio equalizer parameters by deep learning according to claim 9, characterized in that: Step S54 includes the following steps: Perform frequency band decomposition on the parameter correction data to obtain a frequency band feature set; perform parameter mapping on the frequency band feature set to generate a frequency band mapping matrix; Perform compensation calculation on the frequency band mapping matrix to obtain a compensation coefficient set; Perform boundary constraint processing on the compensation coefficient set to generate a constraint parameter set; perform gradient calculation on the constraint parameter set to generate a gradient feature set; Optimizing parameters of the gradient feature set is calculated to obtain an optimized parameter set; and adjusting parameters of the candidate equalizer parameter set is performed based on the optimized parameter set to generate final equalizer parameters.

Images