Intelligent training method and system for audio noise reduction model

Abstract

The present invention discloses an intelligent training method and system for an audio noise reduction model, which relates to the technical field of audio noise reduction. The damage assessment module analyzes the spectrum distortion factor Psyz of each frame according to the spectrum state of the model training audio data and the real label audio data. n , which helps to more accurately measure the performance of the model under different spectral states. The model training module uses recurrent neural network technology to better capture the timing information of the audio signal and improve the adaptability of the noise reduction model to complex audio scenarios. The system fully considers the underfitting and overfitting states that may occur in the processing of training data. When processing the diversity of training data, the system not only focuses on the diversity of noise, but also considers the changes in audio signals in different speakers, environments, and pronunciation methods. This helps to improve the model's ability to learn different speech features and make it more fully adaptable to real-world audio noise reduction challenges.

Description

Technical Field

[0001] The present invention relates to the technical field of audio noise reduction, and in particular to an intelligent training method and system for an audio noise reduction model. Background Art

[0002] With the rapid development of modern science and technology, the widespread application of artificial intelligence technology has penetrated into many fields, among which the research and application of audio noise reduction models have become increasingly eye-catching. Audio noise reduction technology aims to extract clear and clean signals from noisy audio and provide high-quality audio input for various applications in the field of audio processing.

[0003] Specifically, the training process of the audio noise reduction model involves a lot of data processing and model adjustment, in which recurrent neural network technology is used to build the model to better capture the timing information in the audio signal. However, this system does not only focus on model building, but also emphasizes full and intelligent processing of data throughout the training process to improve the training effect and generalization ability of the model.

[0004] Although the system has made significant progress in the intelligent training of audio noise reduction models, there are still some key issues that are ignored in the current training process. One of them is to fully consider the underfitting or overfitting state when processing training data; when it is unclear whether it is in an underfitting or overfitting state, it may affect the performance, generalization ability and training results of the model. In the field of audio noise reduction, due to the diversity of actual noise scenes, the imbalance and specificity of training data may lead to insufficient performance of the model on certain noise types.

[0005] In addition, although the system has introduced data enhancement technology to deal with the complex audio characteristics in real-world scenarios, it may still have certain limitations in terms of the diversity of training data. The changes in audio signals in different speakers and environments, as well as the diversity of different tones and pronunciations, should also be considered more comprehensively. Therefore, the system still has room for further optimization in terms of the intelligence of model training to better adapt to the audio noise reduction challenges in the real world. Summary of the invention

[0006] 1. Technical issues to be resolved

[0007] In view of the deficiencies in the prior art, the present invention provides an intelligent training method and system for an audio noise reduction model, which solves the problems in the above-mentioned background technology.

[0008] (II) Technical solution

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions: an intelligent training system for an audio noise reduction model, comprising a data acquisition module, a preprocessing module, a damage assessment module, a model training module and a training evaluation module;

[0010] The data acquisition module is used to collect noise audio data and real label audio data in advance, and divide it into n groups of frames to obtain the spectrum states in n groups of model training audio data frames and real label audio data frames respectively, and collect relevant preset hyperparameter data information and relevant diversified data information before and after model training;

[0011] The preprocessing module is used to selectively add random noise using data enhancement technology, and superimpose different noises onto the audio data of the real label to generate a complete training set; and before model training, the sampling rate and bit depth of the input layer of the noisy audio data are the same as the sampling rate and bit depth of the target output layer;

[0012] The damage assessment module is used to analyze and obtain the corresponding spectrum distortion factor Psyz in each frame according to the spectrum status in n groups of model training audio data frames and real label audio data frames. n , and use statistics to find the average value to obtain the total spectrum distortion coefficient Psxs;

[0013] The model training module is used to establish an audio noise reduction model using recurrent neural network technology, perform feature extraction on the relevant preset hyperparameter data information and the relevant diversified data information to obtain the number of iterations Ddcs, the number of neurons in each layer Sjsz, the noise intensity change frequency Qdpl, the timbre difference value Ysyz and the tone difference value Ydyz, and obtain the preset influence factor Yyyz by associating the number of iterations Ddcs with the number of neurons in each layer Sjsz, and associating the preset influence factor Yyyz with the timbre difference value Ysyz to obtain the training richness coefficient Xfxs, and after dimensionless processing, comprehensively evaluate the training to obtain the comprehensive evaluation index Zpzs;

[0014] The training evaluation module is used to pre-set an evaluation threshold V, and compare and analyze the evaluation threshold V with the comprehensive evaluation index Zpzs to evaluate the fitting state of the current model training.

[0015] Preferably, the data acquisition module includes a first acquisition unit, a second acquisition unit and a third acquisition unit;

[0016] The first acquisition unit is used to divide the noise audio data and the real label audio data into n groups of frames, and use a spectrum analyzer to collect the corresponding spectrum Bpz in the n groups of model training audio data frames. i, and the corresponding spectrum Lpz in n sets of true label audio data frames i ;

[0017] The second acquisition unit is used to collect relevant preset hyperparameter data information before model training, including the number of layers Csz of the model, the number of neurons in each layer Sjsz, the number of iterations Ddcs, the learning rate Xxz, the batch value Plz and the convergence speed;

[0018] The third acquisition unit is used to collect relevant diversified data information after model training, including noise intensity change frequency Qdpl, training data volume Xslz, bit depth Wsd, sampling rate Cyl and different speech features, and the different speech features include pitch difference value Ydyz, speaking speed difference value Yslz and timbre difference value Ysyz.

[0019] Preferably, the preprocessing module includes an enhancement unit and a unification unit;

[0020] The enhancement unit is used to combine the fitting state. When the model training is in an underfitting state, according to the data enhancement technology, noise is randomly added, the pitch and timbre are changed, and different noises are superimposed on the audio data of the real label by using the synthetic data method to generate a complete training set, and the number of training cycles is increased;

[0021] The unification unit is used to observe whether the sampling rate Cyl and the bit depth Wsd in the input layer are consistent with the sampling rate Cyl and the bit depth Wsd in the target output layer during training. The specific contents are as follows:

[0022] If they are consistent, it is now in normal state;

[0023] If they are inconsistent, there is a risk of information loss and model instability, as well as the model’s inability to correctly learn and reduce the noise of the audio signal. At this time, uniformity is achieved through interpolation, downsampling, and upsampling.

[0024] Preferably, the damage assessment module includes a frame-by-frame analysis unit and an overall analysis unit;

[0025] The frame-by-frame analysis unit is used to train the corresponding spectrum Bpz in the audio data frame through n groups of models i , and the corresponding spectrum Lpz in n sets of true label audio data frames i , calculate and analyze to obtain the first spectrum distortion factor Psyz1, the second spectrum distortion factor Psyz2, ..., the nth spectrum distortion factor Psyz n , the specific formula is as follows:

[0026]

[0027] ...

[0028]

[0029] In the formula, n represents the number of frames.

[0030] Preferably, the overall analysis unit is used to analyze n groups of spectral distortion factors Psyz n Summarize and average to obtain the total spectrum distortion coefficient Psxs, which is obtained by the following formula:

[0031]

[0032] A distortion threshold K is preset, and the total spectrum distortion coefficient Psxs is compared and analyzed with the distortion threshold K to obtain a distortion level. The specific distortion level is as follows;

[0033] If the total spectrum distortion coefficient Psxs is greater than the distortion threshold K, a first distortion level is obtained, indicating that the distortion state of the overall model training audio data frame is in an unqualified state;

[0034] If the total spectrum distortion coefficient Psxs is equal to the distortion threshold K, a second distortion level is obtained, indicating that the distortion state of the overall model training audio data frame is in an unqualified state;

[0035] If the total spectrum distortion coefficient Psxs is less than the distortion threshold K, a third distortion level is obtained, indicating that the distortion state of the overall model training audio data frame is in a qualified state.

[0036] Preferably, the model training module includes a debugging analysis unit, a diverse analysis unit and a comprehensive analysis unit;

[0037] The debugging and analysis unit is used to manually adjust the hyperparameters before model training, and associate the number of iterations Ddcs with the number of neurons in each layer Sjsz according to the relevant preset hyperparameter data information, and obtain the preset influence factor Yyyz after dimensionless processing. The preset influence factor Yyyz is obtained by the following formula:

[0038]

[0039] Where Plz represents the batch value, Xxz represents the learning rate, Csz represents the number of layers, and p, x, d, c, and s represent the preset proportional coefficients of the batch value Plz, the learning rate Xxz, the number of iterations Ddcs, the number of layers Csz, and the number of neurons in each layer Sjsz, respectively.

[0040] Preferably, the diversity analysis unit is used to obtain the training richness coefficient Xfxs according to the relevant diversified data information, combined with the preset influencing factor Yyyz, and after dimensionless processing. The training richness coefficient Xfxs is obtained by the following formula:

[0041]

[0042] In the formula, Yslz represents the speaking speed difference value, Ysyz represents the timbre difference value, Ydyz represents the pitch difference value, Qdpl represents the noise intensity change frequency, Xslz represents the training data amount, Cyl represents the sampling rate, Wsd represents the bit depth, w1 represents the preset proportional coefficient of the sum of the speaking speed difference value Yslz, the timbre difference value Ysyz and the pitch difference value Ydyz, w2, w3, w4, w5 and w6 represent the preset proportional coefficients of the noise intensity change frequency Qdpl, the training data amount Xslz, the sampling rate Cyl, the bit depth Wsd and the preset influencing factor Yyyz respectively, and C represents the first correction constant.

[0043] Preferably, the comprehensive analysis unit is used to combine the relevant preset hyperparameter data information and the relevant diversified data information, and after dimensionless processing, analyze and obtain a comprehensive evaluation index Zpzs, and the comprehensive evaluation index Zpzs is obtained by the following formula:

[0044]

[0045] Wherein, Zyqd represents the noise intensity, α and β represent the preset proportional coefficients, B represents the second correction constant, and ln2 represents the logarithmic operation with base 2.

[0046] Preferably, the training evaluation module is used to compare and analyze the comprehensive evaluation index Zpzs with the evaluation threshold V to evaluate the fitting state of the current model training, and its specific contents are as follows:

[0047] If the comprehensive evaluation index Zpzs ≥ the evaluation threshold V, it indicates that the current model training is in an overfitting state;

[0048] If the comprehensive evaluation index Zpzs is less than the evaluation threshold V, it indicates that the current model training is in an underfitting state.

[0049] An intelligent training method for an audio noise reduction model comprises the following steps:

[0050] Step 1: First, collect noisy audio data and real labeled audio data, divide them into n groups of frames, and collect relevant preset hyperparameter data information and relevant diversified data information before and after model training;

[0051] Step 2: Then, add random noise selectively, and set the sampling rate and bit depth of the input layer of the noisy audio data to the same as the sampling rate and bit depth of the target output layer before model training;

[0052] Step 3: In addition, for n groups of frames, collect the spectrum states of n groups of model training audio data frames and real label audio data frames, and analyze and obtain the corresponding spectrum distortion factor Psyz in each frame. n , and use statistics to find the average value to obtain the total spectrum distortion coefficient Psxs;

[0053] Step 4: Secondly, using recurrent neural network technology, an audio noise reduction model is established, and the relevant preset hyperparameter data information and the relevant diversified data information are feature extracted, and the preset influence factor Yyyz, training richness coefficient Xfxs and comprehensive evaluation index Zpzs are obtained through machine learning calculation;

[0054] Step 5: Finally, the evaluation threshold V is compared with the comprehensive evaluation index Zpzs to evaluate the fitting status of the current model training.

[0055] (III) Beneficial effects

[0056] The present invention provides an intelligent training method and system for an audio noise reduction model, which has the following beneficial effects:

[0057] (1) The data acquisition module collects audio data containing noise and real labels in advance and divides them into n groups of frames. This framing method helps the model better capture the instantaneous characteristics of the signal, enabling the model to learn and process audio data more accurately, providing high-quality input for model training. The loss estimation module analyzes the spectral distortion factor Psyz of each frame based on the spectral state of the model training audio data and the real label audio data. n, which helps to more accurately measure the performance of the model under different spectral states. The model training module adopts recurrent neural network technology to better capture the timing information of audio signals and improve the adaptability of the noise reduction model to complex audio scenes. In the model training module, by associating preset hyperparameters and diversified data information, taking into account multiple parameters, and calculating the comprehensive evaluation index Zpzs, an intelligent evaluation of the richness of training is formed, which provides a favorable reference for model adjustment, and provides an intuitive and feasible method for judging the fitting state of the current model training through threshold comparison. In short, the system fully considers the underfitting and overfitting states that may occur in the processing of training data. When processing the diversity of training data, the system not only pays attention to the diversity of noise, but also considers the changes in audio signals in different speakers, environments and pronunciation methods, which helps to improve the model's learning ability for different speech features and make it more fully adaptable to the audio noise reduction challenges in the real world. The comprehensive evaluation index Zpzs is used to comprehensively evaluate the effect of model training. Such an evaluation mechanism helps to timely discover and deal with the underfitting or overfitting state of the model, thereby comprehensively improving the training effect and generalization ability of the audio noise reduction model.

[0058] (2) By detecting underfitting or overfitting, we can understand the performance of the model on the training data: If the model underfits, it may mean that its complexity is not enough to capture the true characteristics of the data, and the complexity of the model needs to be increased; while overfitting may indicate that the model is too complex and overfits the training data, and the complexity of the model needs to be reduced. Therefore, determining underfitting and overfitting can help further improve the generalization performance of the model. By adjusting the complexity of the model or using methods such as regularization techniques, the model can be better adapted to new and unseen data, rather than just performing well on the training data. At the same time, determining underfitting and overfitting helps understand how the model learns from the data, has a clearer understanding of the model's behavior, and can better explain the model's performance for specific problems. Adjusting for underfitting or overfitting helps optimize the model and improve its efficiency and performance. This helps reduce the waste of resources, such as computing resources, time, and labor costs.

[0059] (3) The enhancement unit improves the quality of the training set through data enhancement, the adaptive training cycle number adjustment improves the model fitting ability, and the unified unit maintains the consistency of input layer parameters and ensures the consistency of information, which together provide comprehensive and beneficial support for the training process of the audio denoising model.

[0060] (4) The combination of frame-by-frame analysis and overall analysis enables the system to fully understand the distortion of the audio data, while the distortion level assessment provides intuitive information on the degree of distortion, providing a practical reference indicator for the effectiveness and optimization of model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of a flow chart of an intelligent training system for an audio noise reduction model of the present invention;

[0062] Figure 2 A schematic diagram of the steps of an intelligent training method for an audio noise reduction model of the present invention. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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 ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] Example 1

[0065] See also Figure 1 , the present invention provides an intelligent training system for an audio noise reduction model, including a data acquisition module, a preprocessing module, a damage assessment module, a model training module and a training evaluation module;

[0066] The data acquisition module is used to collect noise audio data and real label audio data in advance, and divide it into n groups of frames to obtain the spectrum states in n groups of model training audio data frames and real label audio data frames respectively, and collect relevant preset hyperparameter data information and relevant diversified data information before and after model training;

[0067] The preprocessing module is used to selectively add random noise using data enhancement technology, and superimpose different noises onto the audio data of the real label to generate a complete training set; and before model training, the sampling rate and bit depth of the input layer of the noisy audio data are the same as the sampling rate and bit depth of the target output layer, thereby further improving the stability and generalization ability of the model;

[0068] The damage assessment module is used to analyze and obtain the corresponding spectrum distortion factor Psyz in each frame according to the spectrum status in n groups of model training audio data frames and real label audio data frames. n , and use statistics to find the average value to obtain the total spectrum distortion coefficient Psxs;

[0069] The model training module is used to establish an audio noise reduction model using recurrent neural network technology, perform feature extraction on the relevant preset hyperparameter data information and the relevant diversified data information to obtain the number of iterations Ddcs, the number of neurons in each layer Sjsz, the noise intensity change frequency Qdpl, the timbre difference value Ysyz and the tone difference value Ydyz, and obtain the preset influence factor Yyyz by associating the number of iterations Ddcs with the number of neurons in each layer Sjsz, and associating the preset influence factor Yyyz with the timbre difference value Ysyz to obtain the training richness coefficient Xfxs, and after dimensionless processing, comprehensively evaluate the training to obtain the comprehensive evaluation index Zpzs;

[0070] The training evaluation module is used to pre-set an evaluation threshold V, and compare and analyze the evaluation threshold V with the comprehensive evaluation index Zpzs to evaluate the fitting state of the current model training.

[0071] During the operation of this system, firstly, through the data acquisition module, the system can pre-collect audio data containing noise and true labels and divide them into frames, taking into full account the diversity of actual noise scenes. This provides the model with rich training data to better adapt to different noise types and intensities. The introduction of the damage assessment module enables the system to conduct a detailed analysis of the spectrum state during the training process, especially the spectrum distortion factor Psyz. n The statistical mean calculation of the model helps to understand the distortion in each frame and provides strong support for the adjustment of the model. The model training module uses recurrent neural network technology, makes full use of time series information, extracts relevant preset hyperparameters and diversified data information, and comprehensively considers multi-dimensional factors to obtain a more comprehensive comprehensive evaluation index Zpzs of the model training effect, and compares it with the threshold to provide a reliable evaluation of the current model training fitting status, which helps to timely discover and solve the underfitting or overfitting problems that may exist in the training process, and further improve the performance of the model in practical applications.

[0072] Example 2

[0073] Please refer to Figure 1 , specifically: the data acquisition module includes a first acquisition unit, a second acquisition unit and a third acquisition unit;

[0074] The first acquisition unit is used to divide the noise audio data and the real label audio data into n groups of frames, and use a spectrum analyzer to collect the corresponding spectrum Bpz in the n groups of model training audio data frames. i , and the corresponding spectrum Lpz in n sets of true label audio data frames i ;

[0075] The second acquisition unit is used to collect relevant preset hyperparameter data information before model training, including the number of layers Csz of the model, the number of neurons in each layer Sjsz, the number of iterations Ddcs, the learning rate Xxz, the batch value Plz and the convergence speed;

[0076] The third acquisition unit is used to collect relevant diversified data information after model training, including noise intensity change frequency Qdpl, training data volume Xslz, bit depth Wsd, sampling rate Cyl and different speech features, and the different speech features include pitch difference value Ydyz, speaking speed difference value Yslz and timbre difference value Ysyz.

[0077] In this embodiment, the first acquisition unit captures the frequency spectrum of n groups of frames, which helps the system to fully understand the frequency domain characteristics of the audio signal and provides rich spectral data for model training; the second acquisition unit collects relevant preset hyperparameter data information before model training, which provides the system with pre-training configuration parameters, helps to optimize the model structure and training process, and further improves the stability of the system and adjusts the convergence speed; the third acquisition unit obtains relevant diversified data information after model training, which helps the system to more comprehensively consider the complexity of actual audio scenes and make the model more adaptable and generalized; based on the above collected data information, the system can perform a comprehensive performance evaluation; through the analysis of spectral information, hyperparameters and diversified data, the performance of the model in different scenarios can be more accurately evaluated, providing strong support for further optimization and adjustment.

[0078] Example 3

[0079] Please refer to Figure 1 ,Specifically: the preprocessing module includes an enhancement unit and a unification unit;

[0080] The enhancement unit is used to combine the fitting state. When the model training is in an underfitting state, according to the data enhancement technology, randomly add noise, change the pitch and timbre, and use synthetic data to superimpose different noises on the audio data of the real label to generate a complete training set, while increasing the number of training cycles, or using a regularization method;

[0081] The unification unit is used to observe whether the sampling rate Cyl and the bit depth Wsd in the input layer are consistent with the sampling rate Cyl and the bit depth Wsd in the target output layer during training. The specific contents are as follows:

[0082] If they are consistent, it is now in normal state;

[0083] If they are inconsistent, there is a risk of information loss and model instability, as well as the model’s inability to correctly learn and reduce the noise of the audio signal. At this time, uniformity is achieved through interpolation, downsampling, and upsampling.

[0084] In this embodiment, the enhancement unit helps to generate a richer and more diverse training set, and improves the model's ability to learn various noises and audio features; in the enhancement unit, for the case where the model training is in an underfitting state, the number of training cycles is increased, allowing the model to learn the training data more times. This adaptive adjustment of the number of training cycles helps to alleviate the underfitting problem and further improve the degree of fit of the model to the training data. The unified unit observes whether the sampling rate and bit depth in the input layer are consistent with those in the target output layer during training. If not, there may be a risk of information loss and model instability. The uniformity of the input layer parameters is achieved through interpolation, downsampling, and upsampling, which helps to ensure the consistency of the input and output layer parameters of the model during training and improve the stability and generalization ability of the model. The detection and adjustment of the input layer parameters by the unified unit helps to maintain the consistency of information. The maintenance of information consistency can reduce information loss in the model learning process and reduce the model's misunderstanding of the input signal, thereby improving the noise reduction effect.

[0085] Example 4

[0086] Please refer to Figure 1 ,Specifically: the damage assessment module includes a frame-by-frame analysis unit and an overall analysis unit;

[0087] The frame-by-frame analysis unit is used to train the corresponding spectrum Bpz in the audio data frame through n groups of models i , and the corresponding spectrum Lpz in n sets of true label audio data frames i , calculate and analyze to obtain the first spectrum distortion factor Psyz1, the second spectrum distortion factor Psyz2, ..., the nth spectrum distortion factor Psyz n , the specific formula is as follows:

[0088]

[0089] ...

[0090]

[0091] In the formula, n represents the number of frames.

[0092] The overall analysis unit is used to analyze n groups of spectral distortion factors Psyz n Summarize and average to obtain the total spectrum distortion coefficient Psxs, which is obtained by the following formula:

[0093]

[0094] A distortion threshold K is preset, and the total spectrum distortion coefficient Psxs is compared and analyzed with the distortion threshold K to obtain a distortion level. The specific distortion level is as follows;

[0095] If the total spectrum distortion coefficient Psxs is greater than the distortion threshold K, a first distortion level is obtained, indicating that the distortion state of the overall model training audio data frame is in an unqualified state;

[0096] If the total spectrum distortion coefficient Psxs is equal to the distortion threshold K, a second distortion level is obtained, indicating that the distortion state of the overall model training audio data frame is in an unqualified state;

[0097] If the total spectrum distortion coefficient Psxs is less than the distortion threshold K, a third distortion level is obtained, indicating that the distortion state of the overall model training audio data frame is in a qualified state.

[0098] In this embodiment, the frame-by-frame analysis unit calculates the corresponding spectral distortion factors Psyz1, Psyz2, ..., Psyzn in each frame to perform a detailed analysis of the spectral distortion of the overall model training audio data frame, which helps the system to deeply understand the spectral differences of each frame and further improve the fineness analysis of the audio signal distortion; the overall analysis unit summarizes n groups of spectral distortion factors Psyzn and calculates the average to obtain the total spectral distortion coefficient Psxs, which helps to comprehensively consider the distortion of the overall model training audio data frame and provides a global distortion evaluation index. The distortion threshold K is set in advance, and the total spectral distortion coefficient Psxs is compared with the distortion threshold K to obtain the corresponding distortion level, which helps the system to quickly understand the overall distortion level of the model training and provide guidance for subsequent adjustment and optimization. Through the distortion level information, the system can timely understand the distortion state of the overall model training audio data frame and take corresponding adjustment measures according to the distortion level. This real-time feedback helps to optimize the training process and improve the noise reduction performance of the model.

[0099] Example 5

[0100] Please refer to Figure 1 ,Specifically: the model training module includes a debugging analysis unit, a diverse analysis unit and a comprehensive analysis unit;

[0101] The debugging and analysis unit is used to manually adjust the hyperparameters before model training, and associate the number of iterations Ddcs with the number of neurons in each layer Sjsz according to the relevant preset hyperparameter data information, and obtain the preset influence factor Yyyz after dimensionless processing. The preset influence factor Yyyz is obtained by the following formula:

[0102]

[0103] Wherein, Plz represents the batch value, Xxz represents the learning rate, Csz represents the number of layers, p, x, d, c and s represent the preset proportional coefficients of the batch value Plz, the learning rate Xxz, the number of iterations Ddcs, the number of layers Csz and the number of neurons in each layer Sjsz, respectively, where 0.02≤p≤0.12, 0.11≤x≤0.22, 0.15≤d≤0.25, 0.10≤c≤0.25, 0.05≤s≤0.16, and 0.50≤p+x+d+c+s≤1.0.

[0104] The diversity analysis unit is used to obtain the training richness coefficient Xfxs according to the relevant diversified data information, combined with the preset influencing factor Yyyz, and after dimensionless processing. The training richness coefficient Xfxs is obtained by the following formula:

[0105]

[0106] Wherein, Yslz represents the speech rate difference value, Ysyz represents the timbre difference value, Ydyz represents the pitch difference value, Qdpl represents the noise intensity change frequency, Xslz represents the amount of training data, Cyl represents the sampling rate, Wsd represents the bit depth, w1 represents the preset proportional coefficient of the sum of the speech rate difference value Yslz, the timbre difference value Ysyz and the pitch difference value Ydyz, w2, w3, w4, w5 and w6 represent the noise intensity change frequency Qdpl, respectively. pl, the preset proportional coefficient of the training data volume Xslz, the sampling rate Cyl, the bit depth Wsd and the preset influencing factor Yyyz, wherein 0.06≤w1≤0.21, 0.10≤w2≤0.17, 0.05≤w3≤0.15, 0.08≤w4≤0.22, 0.02≤w5≤0.15, 0.03≤w6≤0.10, and 0.40≤w1+w2+w3+w4+w5+w6≤1.0, C is represented as the first correction constant.

[0107] In this embodiment, the debugging and analysis unit provides an opportunity to manually adjust the hyperparameters before model training. By associating the number of iterations Ddcs and the number of neurons in each layer Sjsz, and performing dimensionless processing, a preset influencing factor Yyyz is obtained. This factor comprehensively considers a variety of data information, thereby providing an effective debugging guide and helping to quickly adjust the model structure. The diversity analysis unit obtains the training richness coefficient Xfxs based on the relevant diversified data information and the preset influencing factor Yyyz, and comprehensively evaluates the training richness by considering the diversity factors such as the speech speed difference value Yslz, the timbre difference value Ysyz, the pitch difference value Ydyz, the noise intensity change frequency Qdpl, the training data volume Xslz, the sampling rate Cyl, and the bit depth Wsd, which helps the system to more comprehensively understand the diversity of training data and improve the model's learning ability for different audio features.

[0108] Example 6

[0109] Please refer to Figure 1 Specifically, the comprehensive analysis unit is used to combine the relevant preset hyperparameter data information and the relevant diversified data information, and after dimensionless processing, analyze and obtain the comprehensive evaluation index Zpzs, and the comprehensive evaluation index Zpzs is obtained by the following formula:

[0110]

[0111] Wherein, Zyqd represents the noise intensity, α and β represent preset proportional coefficients, wherein 0.10≤α≤0.50, 0.11≤β≤0.50, and 0.30≤α+β≤1.0, B represents the second correction constant, and ln2 represents the logarithmic operation with base 2.

[0112] The training evaluation module is used to compare and analyze the comprehensive evaluation index Zpzs with the evaluation threshold V to evaluate the fitting state of the current model training. The specific contents are as follows:

[0113] If the comprehensive evaluation index Zpzs ≥ the evaluation threshold V, it indicates that the current model training is in an overfitting state;

[0114] If the comprehensive evaluation index Zpzs is less than the evaluation threshold V, it indicates that the current model training is in an underfitting state.

[0115] In this embodiment, the unit combines relevant preset hyperparameter data information and relevant diversified data information, and after dimensionless processing, analyzes and obtains the comprehensive evaluation index Zpzs. This index comprehensively evaluates the overall state of model training by considering multiple factors such as noise intensity Zyqd. Among them, the introduction of the preset proportional coefficient and the second correction constant B makes the index more comprehensive and accurate, which helps the system to understand the complexity of model training more deeply. The training evaluation module compares and analyzes the comprehensive evaluation index Zpzs with the pre-set evaluation threshold V to evaluate the fitting state of the current model training, which provides a clear guide for the system to help users better understand the fitting performance of model training, thereby adjusting the model structure and hyperparameters and improving the training effect of the audio noise reduction model.

[0116] Example 7

[0117] Please refer to Figure 1 and Figure 2 Specifically: an intelligent training method for an audio noise reduction model comprises the following steps:

[0118] Step 1: First, collect noisy audio data and real labeled audio data, divide them into n groups of frames, and collect relevant preset hyperparameter data information and relevant diversified data information before and after model training;

[0119] Step 2: Then, add random noise selectively, and set the sampling rate and bit depth of the input layer of the noisy audio data to the same as the sampling rate and bit depth of the target output layer before model training;

[0120] Step 3: In addition, for n groups of frames, collect the spectrum states of n groups of model training audio data frames and real label audio data frames, and analyze and obtain the corresponding spectrum distortion factor Psyz in each frame. n , and use statistics to find the average value to obtain the total spectrum distortion coefficient Psxs;

[0121] Step 4: Secondly, using recurrent neural network technology, an audio noise reduction model is established, and the relevant preset hyperparameter data information and the relevant diversified data information are feature extracted, and the preset influence factor Yyyz, training richness coefficient Xfxs and comprehensive evaluation index Zpzs are obtained through machine learning calculation;

[0122] Step 5: Finally, the evaluation threshold V is compared with the comprehensive evaluation index Zpzs to evaluate the fitting status of the current model training.

[0123] Example: An audio processing factory introduced an intelligent training method and system for an audio noise reduction model. The following is an example of an audio processing factory:

[0124] Data acquisition: frame number n is 3; spectrum Bpz1 is 10.2; spectrum Bpz2 is 15.2; spectrum Bpz3 is 11.6; spectrum Lpz1 is 10.5; spectrum Lpz2 is 12.7; spectrum Lpz3 is 10.4;

[0125] The batch value Plz is 37; the learning rate Xxz is 67%; the number of iterations Ddcs is 10; the number of layers Csz is 28; the number of neurons in each layer Sjsz is 54;

[0126] Preset proportional coefficients: p is 0.10; x is 0.15; d is 0.20; c is 0.11; s is 0.08;

[0127] The speech rate difference value Yslz is 2.1; the timbre difference value Ysyz is 3.5; the pitch difference value Ydyz is 5.1; the noise intensity change frequency Qdpl is 84%; the training data volume Xslz is 102; the sampling rate Cyl is 81%; the bit depth Wsd is 24;

[0128] Preset proportional coefficients: w1 is 0.15; w2 is 0.13; w3 is 0.10; w4 is 0.14; w5 is 0.08; w6 is 0.06; the first correction constant C is 0.48;

[0129] The noise intensity Zyqd is 12.4;

[0130] Preset proportional coefficient: α is 0.21; β is 0.26; the second correction constant B is 0.06;

[0131] With the above data, the following calculations can be performed:

[0132] First spectral distortion factor

[0133] Second spectral distortion factor

[0134] The third spectral distortion factor

[0135] Total spectral distortion factor

[0136] Default impact factor

[0137] Training richness coefficient

[0138]

[0139] Comprehensive evaluation index

[0140] If the evaluation threshold V is 20, the comprehensive evaluation index Zpzs < evaluation threshold V, indicating that the current model training is in an underfitting state.

[0141] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent training system for an audio noise reduction model, characterized in that: It includes data acquisition module, preprocessing module, damage assessment module, model training module and training evaluation module; The data acquisition module is used to collect noise audio data and real label audio data in advance, and divide it into n groups of frames to obtain the spectrum states in n groups of model training audio data frames and real label audio data frames respectively, and collect relevant preset hyperparameter data information and relevant diversified data information before and after model training; The preprocessing module is used to selectively add random noise using data enhancement technology, and superimpose different noises onto the audio data of the real label to generate a complete training set; Before model training, the sampling rate and bit depth of the input layer of the noisy audio data are the same as the sampling rate and bit depth of the target output layer; The damage assessment module is used to analyze and obtain the corresponding spectrum distortion factor Psyz in each frame according to the spectrum status in n groups of model training audio data frames and real label audio data frames. n , and use statistics to find the average value to obtain the total spectrum distortion coefficient Psxs; The model training module is used to establish an audio noise reduction model using recurrent neural network technology, perform feature extraction on the relevant preset hyperparameter data information and the relevant diversified data information to obtain the number of iterations Ddcs, the number of neurons in each layer Sjsz, the noise intensity change frequency Qdpl, the timbre difference value Ysyz and the tone difference value Ydyz, and obtain the preset influence factor Yyyz by associating the number of iterations Ddcs with the number of neurons in each layer Sjsz, and associating the preset influence factor Yyyz with the timbre difference value Ysyz to obtain the training richness coefficient Xfxs, and after dimensionless processing, comprehensively evaluate the training to obtain the comprehensive evaluation index Zpzs; The training evaluation module is used to pre-set an evaluation threshold V, and compare and analyze the evaluation threshold V with the comprehensive evaluation index Zpzs to evaluate the fitting state of the current model training.

2. The intelligent training system for an audio noise reduction model according to claim 1, characterized in that: The data acquisition module includes a first acquisition unit, a second acquisition unit and a third acquisition unit; The first acquisition unit is used to divide the noise audio data and the real label audio data into n groups of frames, and use a spectrum analyzer to collect the corresponding spectrum Bpz in the n groups of model training audio data frames. i , and the corresponding spectrum Lpz in n sets of true label audio data frames i ; The second acquisition unit is used to collect relevant preset hyperparameter data information before model training, including the number of layers Csz of the model, the number of neurons in each layer Sjsz, the number of iterations Ddcs, the learning rate Xxz, the batch value Plz and the convergence speed; The third acquisition unit is used to collect relevant diversified data information after model training, including noise intensity change frequency Qdpl, training data volume Xslz, bit depth Wsd, sampling rate Cyl and different speech features, and the different speech features include pitch difference value Ydyz, speaking speed difference value Yslz and timbre difference value Ysyz.

3. The intelligent training system for an audio noise reduction model according to claim 2, characterized in that: The preprocessing module includes an enhancement unit and a unification unit; The enhancement unit is used to combine the fitting state. When the model training is in an underfitting state, according to the data enhancement technology, noise is randomly added, the pitch and timbre are changed, and different noises are superimposed on the audio data of the real label by using the synthetic data method to generate a complete training set, and the number of training cycles is increased; The unification unit is used to observe whether the sampling rate Cyl and the bit depth Wsd in the input layer are consistent with the sampling rate Cyl and the bit depth Wsd in the target output layer during training. The specific contents are as follows: If they are consistent, it is now in normal state; If they are inconsistent, there is a risk of information loss and model instability, as well as the model’s inability to correctly learn and reduce the noise of the audio signal. At this time, uniformity is achieved through interpolation, downsampling, and upsampling.

4. The intelligent training system for an audio noise reduction model according to claim 3, characterized in that: The damage assessment module includes a frame-by-frame analysis unit and an overall analysis unit; The frame-by-frame analysis unit is used to train the corresponding spectrum Bpz in the audio data frame through n groups of models i , and the corresponding spectrum Lpz in n sets of true label audio data frames i , calculate and analyze to obtain the first spectrum distortion factor Psyz1, the second spectrum distortion factor Psyz2, ..., the nth spectrum distortion factor Psyz n , the specific formula is as follows: ... In the formula, n represents the number of frames.

5. The intelligent training system for an audio noise reduction model according to claim 4, characterized in that: The overall analysis unit is used to analyze n groups of spectral distortion factors Psyz n Summarize and average to obtain the total spectrum distortion coefficient Psxs, which is obtained by the following formula: A distortion threshold K is preset, and the total spectrum distortion coefficient Psxs is compared and analyzed with the distortion threshold K to obtain a distortion level. The specific distortion level is as follows; If the total spectrum distortion coefficient Psxs is greater than the distortion threshold K, a first distortion level is obtained, indicating that the distortion state of the overall model training audio data frame is in an unqualified state; If the total spectrum distortion coefficient Psxs is equal to the distortion threshold K, a second distortion level is obtained, indicating that the distortion state of the overall model training audio data frame is in an unqualified state; If the total spectrum distortion coefficient Psxs is less than the distortion threshold K, a third distortion level is obtained, indicating that the distortion state of the overall model training audio data frame is in a qualified state.

6. The intelligent training system for an audio noise reduction model according to claim 3, characterized in that: The model training module includes a debugging analysis unit, a diverse analysis unit and a comprehensive analysis unit; The debugging and analysis unit is used to manually adjust the hyperparameters before model training, and associate the number of iterations Ddcs with the number of neurons in each layer Sjsz according to the relevant preset hyperparameter data information, and obtain the preset influence factor Yyyz after dimensionless processing. The preset influence factor Yyyz is obtained by the following formula: Where Plz represents the batch value, Xxz represents the learning rate, Csz represents the number of layers, and p, x, d, c, and s represent the preset proportional coefficients of the batch value Plz, the learning rate Xxz, the number of iterations Ddcs, the number of layers Csz, and the number of neurons in each layer Sjsz, respectively.

7. The intelligent training system for an audio noise reduction model according to claim 6, characterized in that: The diversity analysis unit is used to obtain the training richness coefficient Xfxs according to the relevant diversified data information, combined with the preset influencing factor Yyyz, and after dimensionless processing. The training richness coefficient Xfxs is obtained by the following formula: In the formula, Yslz represents the speaking speed difference value, Ysyz represents the timbre difference value, Ydyz represents the pitch difference value, Qdpl represents the noise intensity change frequency, Xslz represents the training data amount, Cyl represents the sampling rate, Wsd represents the bit depth, w1 represents the preset proportional coefficient of the sum of the speaking speed difference value Yslz, the timbre difference value Ysyz and the pitch difference value Ydyz, w2, w3, w4, w5 and w6 represent the preset proportional coefficients of the noise intensity change frequency Qdpl, the training data amount Xslz, the sampling rate Cyl, the bit depth Wsd and the preset influencing factor Yyyz respectively, and C represents the first correction constant.

8. The intelligent training system for an audio noise reduction model according to claim 6, characterized in that: The comprehensive analysis unit is used to combine the relevant preset hyperparameter data information and the relevant diversified data information, and after dimensionless processing, analyze and obtain a comprehensive evaluation index Zpzs, and the comprehensive evaluation index Zpzs is obtained by the following formula: Wherein, Zyqd represents the noise intensity, α and β represent the preset proportional coefficients, B represents the second correction constant, and ln2 represents the logarithmic operation with base 2.

9. The intelligent training system for an audio noise reduction model according to claim 1, characterized in that: The training evaluation module is used to compare and analyze the comprehensive evaluation index Zpzs with the evaluation threshold V to evaluate the fitting state of the current model training. The specific contents are as follows: If the comprehensive evaluation index Zpzs ≥ the evaluation threshold V, it indicates that the current model training is in an overfitting state; If the comprehensive evaluation index Zpzs is less than the evaluation threshold V, it indicates that the current model training is in an underfitting state.

10. An intelligent training method for an audio noise reduction model, comprising an intelligent training system for an audio noise reduction model according to any one of claims 1 to 9, characterized in that: The following steps are included: Step 1: First, collect noisy audio data and real labeled audio data, divide them into n groups of frames, and collect relevant preset hyperparameter data information and relevant diversified data information before and after model training; Step 2: Then, add random noise selectively, and set the sampling rate and bit depth of the input layer of the noisy audio data to the same as the sampling rate and bit depth of the target output layer before model training; Step 3: In addition, for n groups of frames, collect the spectrum states of n groups of model training audio data frames and real label audio data frames, and analyze and obtain the corresponding spectrum distortion factor Psyz in each frame. n , and use statistics to find the average value to obtain the total spectrum distortion coefficient Psxs; Step 4: Secondly, using recurrent neural network technology, an audio noise reduction model is established, and the relevant preset hyperparameter data information and the relevant diversified data information are feature extracted, and the preset influence factor Yyyz, training richness coefficient Xfxs and comprehensive evaluation index Zpzs are obtained through machine learning calculation; Step 5: Finally, the evaluation threshold V is compared with the comprehensive evaluation index Zpzs to evaluate the fitting status of the current model training.

Images