A data recognition method, system and application based on a neural network model

Abstract

The application discloses a data recognition method and system based on a neural network model and application, relates to the field of audio processing, and solves the problem that the existing data recognition method is not good at recognizing and processing audio data, and comprises the following steps: S1, acquiring audio to be recognized, and intercepting a sound period of the target audio in the audio to be recognized to obtain target audio period collection data; S2, acquiring a plurality of groups of historical target audio, marking the audio sources of the historical target audio by setting a denoising index to obtain audio source marking data, and creating an audio neural network model according to the audio source marking data; and S3, recognizing and removing audio noise in each target audio interception period according to the audio neural network model, so that the audio denoising efficiency is improved.

Description

Technical Field

[0001] This invention belongs to the field of audio processing and relates to artificial feedback reinforcement learning technology, specifically a data recognition method, system, and application based on a neural network model. Background Technology

[0002] Existing data recognition methods have the following specific drawbacks when processing audio data:

[0003] Existing data recognition methods cannot extract the sound segments containing the target audio from the audio to be identified and perform targeted noise reduction on the extracted sound segments, resulting in low audio noise reduction efficiency.

[0004] Existing data recognition methods cannot create audio neural network models by labeling the source of historical target audio, nor can they identify and remove audio noise for each segment of target audio based on the audio neural network model, resulting in a low degree of automation in the audio denoising process.

[0005] To this end, we propose a data recognition method, system, and application based on a neural network model. Summary of the Invention

[0006] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a data recognition method, system and application based on a neural network model. This invention aims to improve the targeting and accuracy of the audio denoising process.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a data recognition method based on a neural network model, comprising the following specific steps:

[0008] Step S1: Obtain the audio to be identified, and extract the sound period containing the target audio from the audio to obtain the target audio period data;

[0009] Step S2: Obtain multiple sets of historical target audio, mark the audio sources of the historical target audio by setting multiple sets of noise reduction indicators, obtain audio source marking data, and create an audio neural network model based on the audio source marking data;

[0010] Step S3: Based on the audio neural network model, identify and remove audio noise for each target audio segment.

[0011] Furthermore, step S1 also includes the following specific steps:

[0012] Step S11: During the real-time audio monitoring of the audio recognition location through the audio acquisition device, the time point when the audio device starts working is marked as the audio start time point of the time period, the time point corresponding to the current moment is marked as the audio end time point, and the environmental recordings generated at the audio recognition location at the audio start time point and the audio end time point are marked as the audio to be recognized.

[0013] Step S12: Segment the audio to be identified into several time-continuous audio short frames, and select a sample audio short frame from the acquired audio short frames.

[0014] Step S13: Perform target audio energy analysis on the short-time frames of the sample audio, and obtain the target audio energy ratio corresponding to the short-time frames of the sample audio based on the analysis results;

[0015] Step S14: Obtain several historical audio short frames that have been identified with target audio, obtain the target audio energy ratio corresponding to each historical audio short frame, and compare the values ​​of the multiple target audio energy ratios obtained, and mark the target audio energy ratio with the smallest value as the target audio reference energy ratio.

[0016] Step S15: Obtain the target audio energy ratio for each audio short frame, and compare the obtained target audio energy ratio with the target audio reference energy ratio. Based on the comparison result, divide the audio short frames into valid audio short frames and invalid audio short frames.

[0017] Step S16: In the audio to be identified, the sound time period corresponding to the effective audio short frame is extracted to obtain multiple target audio extraction time periods;

[0018] Step S17: Define the acquired target audio segment as the target audio segment acquisition data.

[0019] Furthermore, step S13 also includes the following specific steps:

[0020] The sample audio short frame is divided into several audio signal sampling points, and the divided audio signal sampling points are named X1 signal sampling point to Xa signal sampling point in the order of the division time.

[0021] The intensity values ​​of the target audio signals corresponding to the sampling points X1 to Xa of the signal are obtained respectively.

[0022] At the X1 signal sampling point, the intensity values ​​of any ambient audio signal other than the target audio signal are obtained, and the intensity values ​​of multiple ambient audio signals are compared. The ambient audio signal intensity value with the largest value is marked as the X1 ambient signal intensity value.

[0023] Repeat the process of acquiring the environmental signal intensity value of X1, and acquire the environmental signal intensity values ​​corresponding to the sampling points of X2 signal to Xa signal respectively, to obtain the environmental signal intensity values ​​of X2 to Xa.

[0024] Calculate the ratio of the X1 target signal strength value to the X1 ambient signal strength value to obtain the X1 target audio intensity ratio. Calculate the ratio of the X2 target signal strength value to the X2 ambient signal strength value to obtain the X2 target audio intensity ratio. And so on, calculate the ratio of the Xa target signal strength value to the Xa ambient signal strength value to obtain the Xa target audio intensity ratio.

[0025] The target audio intensity ratios from X1 to Xa are averaged to obtain the target audio energy ratios corresponding to short frames of the sample audio.

[0026] Furthermore, step S15 also includes the following specific steps:

[0027] If the target audio energy ratio is greater than or equal to the target audio reference energy ratio, then the corresponding audio short frame is divided into a valid audio short frame.

[0028] If the target audio energy ratio is less than the target audio reference energy ratio, the corresponding audio short frame will be classified as an invalid audio short frame.

[0029] Furthermore, step S2 also includes the following specific steps:

[0030] Step S21: During the process of audio denoising of the target audio segment, several different types of audio denoising indicators are set, and the different types of audio denoising indicators are marked as Z1 denoising indicator to Zb denoising indicator respectively.

[0031] Step S22: Analyze the Z1 denoising index to the Zb denoising index based on the historical target audio to obtain the Z1 denoising index baseline range to the Zb denoising index baseline range.

[0032] Step S23: Obtain several historical audio truncation periods in the audio recognition field. In the historical audio truncation periods, the audio sources whose Z1 denoising index to Zb denoising index are all within the Z1 denoising index baseline range to the Zb denoising index baseline range are marked as first-type audio sources. The audio sources whose denoising index is not within the corresponding denoising index range are marked as second-type audio sources, thus obtaining audio source marking data.

[0033] Further, step S24: Divide the audio source labeling data into a denoised frequency training set and a denoised frequency test set according to the audio training-test ratio;

[0034] Step S25: Create an audio recognition model using an existing artificial intelligence platform, and train the audio recognition model using a denoised frequency training set until the audio recognition model is trained once for each historical audio segment in the denoised frequency training set.

[0035] Step S26: Test the audio recognition model using the denoised frequency test set and obtain the recognition accuracy. When the recognition accuracy is greater than or equal to the target recognition accuracy, the audio recognition model training is complete and an audio neural network model is obtained. When the recognition accuracy is less than the target recognition accuracy, continue to train the audio recognition model using the denoised frequency training set until the recognition accuracy is greater than or equal to the target recognition accuracy.

[0036] Furthermore, step S22 also includes the following specific steps:

[0037] Multiple sets of historical target audio that have been denoised are acquired. Among the acquired historical target audio, a sample target audio is selected. The Z1 denoising index value corresponding to each time point in the sample target audio is acquired, resulting in multiple Z1 denoising index values. The values ​​of the multiple Z1 denoising index values ​​are compared, and the range of values ​​formed by the largest and smallest Z1 denoising index values ​​is obtained to obtain the Z1 denoising index value range corresponding to the sample target audio.

[0038] Repeat the process of obtaining the Z1 denoising index value range corresponding to the sample target audio, obtain the Z1 denoising index value range corresponding to each historical target audio, obtain multiple Z1 denoising index value ranges, and perform interval union on the obtained multiple Z1 denoising index value ranges to obtain the Z1 denoising index baseline range.

[0039] Repeat the process of obtaining the Z1 denoising index benchmark interval to obtain the Z2 denoising index benchmark interval to the Zb denoising index benchmark interval.

[0040] Furthermore, step S3 also includes the following specific steps:

[0041] Obtain the audio neural network model, acquire target audio time segment data, and obtain multiple target audio segmentation time segments based on the target audio time segment data;

[0042] An audio neural network model is used to identify the second type of sound source in the target audio segment, and the second type of sound source in each target audio segment is removed as audio noise.

[0043] A data recognition system based on a neural network model includes:

[0044] Audio acquisition module: Acquires the audio to be identified and extracts the sound period containing the target audio from the audio to obtain the target audio period acquisition data;

[0045] Model creation module: Acquire multiple sets of historical target audio, mark the audio sources of the historical target audio by setting multiple sets of noise reduction indicators, obtain audio source marking data, and create an audio neural network model based on the audio source marking data;

[0046] Audio noise reduction module: Based on the audio neural network model, it identifies and removes audio noise from each target audio segment.

[0047] Furthermore, data recognition methods can be applied in audio denoising.

[0048] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0049] 1. This invention improves the efficiency of audio denoising by extracting the sound segments containing the target audio in the audio to be identified and then performing targeted denoising on the extracted sound segments.

[0050] 2. This invention creates an audio neural network model by marking the source of historical target audio, and identifies and removes audio noise for each target audio segment based on the audio neural network model, thereby improving the automation of the audio denoising process. Attached Figure Description

[0051] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0052] Figure 1 This is an overall system block diagram of the present invention;

[0053] Figure 2 This is a diagram illustrating the implementation steps of the present invention. Detailed Implementation

[0054] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0055] Firstly, please refer to Figure 1 The audio neural network model in this invention involves artificial feedback reinforcement learning. This invention provides a technical solution: a data recognition method based on a neural network model, comprising the following specific steps:

[0056] Step S1: Obtain the audio to be identified, and extract the sound period containing the target audio from the audio to obtain the target audio period data;

[0057] Step S1 further includes the following specific steps:

[0058] Step S11: During the real-time audio monitoring of the audio recognition location through the audio acquisition device, the time point when the audio device starts working is marked as the audio start time point of the time period, the time point corresponding to the current moment is marked as the audio end time point, and the environmental recordings generated at the audio recognition location at the audio start time point and the audio end time point are marked as the audio to be recognized.

[0059] Step S12: Segment the audio to be identified into several time-continuous audio short frames, and select a sample audio short frame from the acquired audio short frames.

[0060] Step S13: Perform target audio energy analysis on the short-time frames of the sample audio, and obtain the target audio energy ratio corresponding to the short-time frames of the sample audio based on the analysis results;

[0061] Step S13 further includes the following specific steps:

[0062] The sample audio short frame is divided into several audio signal sampling points, and the divided audio signal sampling points are named X1 signal sampling point to Xa signal sampling point in the order of the division time.

[0063] The intensity values ​​of the target audio signals corresponding to the sampling points X1 to Xa of the signal are obtained respectively.

[0064] At the X1 signal sampling point, the intensity values ​​of any ambient audio signal other than the target audio signal are obtained, and the intensity values ​​of multiple ambient audio signals are compared. The ambient audio signal intensity value with the largest value is marked as the X1 ambient signal intensity value.

[0065] Repeat the process of acquiring the environmental signal intensity value of X1, and acquire the environmental signal intensity values ​​corresponding to the sampling points of X2 signal to Xa signal respectively, to obtain the environmental signal intensity values ​​of X2 to Xa.

[0066] Calculate the ratio of the X1 target signal strength value to the X1 ambient signal strength value to obtain the X1 target audio intensity ratio. Calculate the ratio of the X2 target signal strength value to the X2 ambient signal strength value to obtain the X2 target audio intensity ratio. And so on, calculate the ratio of the Xa target signal strength value to the Xa ambient signal strength value to obtain the Xa target audio intensity ratio.

[0067] The average of the target audio intensity ratios from X1 to Xa is calculated to obtain the target audio energy ratio corresponding to the short-time frames of the sample audio.

[0068] Step S14: Obtain several historical audio short frames that have been identified with target audio, obtain the target audio energy ratio corresponding to each historical audio short frame, and compare the values ​​of the multiple target audio energy ratios obtained, and mark the target audio energy ratio with the smallest value as the target audio reference energy ratio.

[0069] Step S15: Obtain the target audio energy ratio for each audio short frame, and compare the obtained target audio energy ratio with the target audio reference energy ratio. Based on the comparison result, divide the audio short frames into valid audio short frames and invalid audio short frames.

[0070] Step S15 further includes the following specific steps:

[0071] If the target audio energy ratio is greater than or equal to the target audio reference energy ratio, then the corresponding audio short frame is divided into a valid audio short frame.

[0072] If the target audio energy ratio is less than the target audio reference energy ratio, the corresponding audio short frame will be classified as an invalid audio short frame.

[0073] Step S16: In the audio to be identified, the sound time period corresponding to the effective audio short frame is extracted to obtain multiple target audio extraction time periods;

[0074] Step S17: Define the acquired target audio segment as the target audio segment acquisition data;

[0075] The beneficial effects of step S1 above are as follows:

[0076] By using real-time audio monitoring and dynamic time-segmentation marking mechanisms, the integrity and timeliness of audio acquisition are ensured, providing raw data support for subsequent analysis. The short-time frame segmentation strategy refines the audio into local analysis units, which not only preserves temporal continuity but also facilitates transient feature capture, effectively balancing computational efficiency and feature resolution.

[0077] The system innovatively introduces the target audio energy ratio index, which quantifies the intensity relationship between the target signal and the ambient noise, and constructs a feature parameter with strong anti-interference ability. Compared with the traditional energy threshold method, it significantly improves the target audio recognition rate in complex environments. Based on historical data, an adaptive benchmark energy ratio is generated, which enables the system to have environmental adaptability and dynamically distinguish between effective signals and background noise, avoiding the misjudgment problem of fixed thresholds in changing scenarios.

[0078] By using energy comparison to accurately extract effective time periods, this approach removes silent or noise-dominated periods while preserving the complete temporal distribution of the target audio. This provides high-quality input data for subsequent audio recognition tasks and reduces the consumption of system resources by invalid data. Through the organic combination of signal processing and data filtering, this solution constructs a complete technical chain from raw audio acquisition to target time period extraction, significantly improving target detection accuracy and system operating efficiency in complex audio scenarios.

[0079] Step S2: Obtain multiple sets of historical target audio, mark the audio sources of the historical target audio by setting multiple sets of noise reduction indicators, obtain audio source marking data, and create an audio neural network model based on the audio source marking data;

[0080] Step S2 further includes the following specific steps:

[0081] Step S21: During the process of audio denoising of the target audio segment, several different types of audio denoising indicators are set, and the different types of audio denoising indicators are marked as Z1 denoising indicator to Zb denoising indicator respectively.

[0082] Step S22: Analyze the Z1 denoising index to the Zb denoising index based on the historical target audio to obtain the Z1 denoising index baseline range to the Zb denoising index baseline range.

[0083] Step S22 further includes the following specific steps:

[0084] Multiple sets of historical target audio that have been denoised are acquired. Among the acquired historical target audio, a sample target audio is selected. The Z1 denoising index value corresponding to each time point in the sample target audio is acquired, resulting in multiple Z1 denoising index values. The values ​​of the multiple Z1 denoising index values ​​are compared, and the range of values ​​formed by the largest and smallest Z1 denoising index values ​​is obtained to obtain the Z1 denoising index value range corresponding to the sample target audio.

[0085] Repeat the process of obtaining the Z1 denoising index value range corresponding to the sample target audio, obtain the Z1 denoising index value range corresponding to each historical target audio, obtain multiple Z1 denoising index value ranges, and perform interval union on the obtained multiple Z1 denoising index value ranges to obtain the Z1 denoising index baseline range.

[0086] Repeat the process of obtaining the Z1 denoising index benchmark interval to obtain the Z2 denoising index benchmark interval to the Zb denoising index benchmark interval respectively.

[0087] Step S23: Obtain several historical audio truncation periods in the audio recognition field. In the historical audio truncation periods, the audio sources whose Z1 denoising index to Zb denoising index are all within the Z1 denoising index baseline range to the Zb denoising index baseline range are marked as first-type audio sources. The audio sources whose denoising index is not within the corresponding denoising index range are marked as second-type audio sources, thus obtaining audio source marking data.

[0088] Step S24: Divide the audio source labeling data into a denoised frequency training set and a denoised frequency test set according to the audio training-test ratio;

[0089] Step S25: Create an audio recognition model using an existing artificial intelligence platform, and train the audio recognition model using a denoised frequency training set until the audio recognition model is trained once for each historical audio segment in the denoised frequency training set.

[0090] Step S26: Test the audio recognition model using the denoised frequency test set and obtain the recognition accuracy. When the recognition accuracy is greater than or equal to the target recognition accuracy, the audio recognition model training is complete and an audio neural network model is obtained. When the recognition accuracy is less than the target recognition accuracy, continue to train the audio recognition model using the denoised frequency training set until the recognition accuracy is greater than or equal to the target recognition accuracy.

[0091] The beneficial effects of step S2 are as follows:

[0092] By constructing multiple sets of noise reduction index benchmark intervals, a quantitative assessment of audio quality is achieved. Compared with the single index threshold method, this scheme can more comprehensively reflect the multi-dimensional characteristics of audio signals. The dynamic benchmark intervals generated based on historical data enable the system to have environmental adaptability, automatically distinguish between effective sound sources and noise interference, and avoid the misjudgment problem of fixed thresholds in complex scenarios.

[0093] By creating training and testing sets using audio source labeled data, a closed-loop optimization system was constructed, enabling the model to continuously optimize its feature extraction capabilities during iterative training. The audio neural network model, employing a deep learning architecture, can automatically learn the complex mapping relationship between audio source features and denoising metrics, achieving a better balance between noise suppression and target audio source preservation compared to traditional machine learning models. Finally, through multi-metric collaborative analysis and neural network modeling, the target audio source recognition rate in complex audio scenarios was significantly improved, while reducing the interference of environmental noise on the recognition results.

[0094] Step S3: Based on the audio neural network model, identify and remove audio noise for each target audio segment.

[0095] Step S3 further includes the following specific steps:

[0096] Obtain the audio neural network model, acquire target audio time segment data, and obtain multiple target audio segmentation time segments based on the target audio time segment data;

[0097] An audio neural network model is used to identify the second type of sound source in the target audio segment, and the second type of sound source in each target audio segment is removed as audio noise.

[0098] In this application, if a corresponding calculation formula appears, the above calculation formula is a dimensionless calculation. The weighting coefficient, proportional coefficient and other coefficients in the formula are set to quantify each parameter to obtain a result value. The size of the weighting coefficient and proportional coefficient is only required to not affect the proportional relationship between the parameter and the result value.

[0099] Secondly, please refer to Figure 2 Based on another concept of the same invention, a data recognition system based on a neural network model is proposed, including an audio acquisition module, a model creation module, an audio denoising module, and a server. The audio acquisition module, the model creation module, and the audio denoising module are respectively connected to the server, and the server controls the audio acquisition module, the model creation module, and the audio denoising module respectively.

[0100] The audio acquisition module acquires the audio to be identified and extracts the sound period containing the target audio from the audio to obtain the target audio period acquisition data;

[0101] Specifically as follows:

[0102] During the real-time audio monitoring of the audio recognition location using audio acquisition devices, the time point when the audio device starts working is marked as the audio start time point of the time period, the time point corresponding to the current moment is marked as the audio end time point, and the environmental recordings generated at the audio recognition location at the audio start time point and the audio end time point are marked as the audio to be recognized.

[0103] It should be noted here that:

[0104] In this application, the audio to be identified can be updated in real time according to the change of the current time value;

[0105] It should be noted here that:

[0106] In this application, the audio recognition locations mentioned herein are public places that comply with the rules for audio signal collection and do not involve personal privacy issues.

[0107] The audio to be identified is segmented into several short audio frames that are consecutive in time, and a sample audio frame is selected from the acquired short audio frames.

[0108] It should be noted here that:

[0109] In this application, the frame duration corresponding to the sample audio short frame involved herein is specifically 20ms.

[0110] Perform target audio energy analysis on short-time frames of sample audio, and obtain the target audio energy ratio corresponding to the short-time frames of sample audio based on the analysis results;

[0111] Specifically as follows:

[0112] The sample audio short frame is divided into several audio signal sampling points, and the divided audio signal sampling points are named X1 signal sampling point to Xa signal sampling point in the order of the division time.

[0113] It should be noted here that:

[0114] In this application, X refers to the symbol corresponding to the signal sampling point, and a refers to the quantity value corresponding to the signal sampling point.

[0115] The intensity values ​​of the target audio signals corresponding to the sampling points X1 to Xa of the signal are obtained respectively.

[0116] It should be noted here that:

[0117] In this application, the target audio signal referred to herein is specifically a human voice signal;

[0118] At the X1 signal sampling point, the intensity values ​​of any ambient audio signal other than the target audio signal are obtained, and the intensity values ​​of multiple ambient audio signals are compared. The ambient audio signal intensity value with the largest value is marked as the X1 ambient signal intensity value.

[0119] Repeat the process of acquiring the environmental signal intensity value of X1, and acquire the environmental signal intensity values ​​corresponding to the sampling points of X2 signal to Xa signal respectively, to obtain the environmental signal intensity values ​​of X2 to Xa.

[0120] It should be noted here that:

[0121] In this application, the environmental audio signals referred to herein include, but are not limited to, natural environmental sounds, man-made environmental sounds, and biological noise generated by organisms other than humans.

[0122] Calculate the ratio of the X1 target signal strength value to the X1 ambient signal strength value to obtain the X1 target audio intensity ratio. Calculate the ratio of the X2 target signal strength value to the X2 ambient signal strength value to obtain the X2 target audio intensity ratio. And so on, calculate the ratio of the Xa target signal strength value to the Xa ambient signal strength value to obtain the Xa target audio intensity ratio.

[0123] The average of the target audio intensity ratios from X1 to Xa is calculated to obtain the target audio energy ratio corresponding to the short-time frames of the sample audio.

[0124] Acquire several historical audio short frames that have already identified target audio, obtain the target audio energy ratio corresponding to each historical audio short frame, and compare the values ​​of the multiple target audio energy ratios obtained, and mark the target audio energy ratio with the smallest value as the target audio reference energy ratio;

[0125] Repeat the process of obtaining the target audio energy ratio corresponding to the sample audio short frame. Obtain the target audio energy ratio for each audio short frame, and compare the obtained target audio energy ratio with the target audio reference energy ratio. Based on the comparison result, divide the audio short frames into valid audio short frames and invalid audio short frames.

[0126] If the target audio energy ratio is greater than or equal to the target audio reference energy ratio, then the corresponding audio short frame is divided into a valid audio short frame.

[0127] If the target audio energy ratio is less than the target audio reference energy ratio, the corresponding audio short frame will be classified as an invalid audio short frame.

[0128] In the audio to be identified, the sound time period corresponding to the effective audio short frame is extracted to obtain multiple target audio extraction time periods;

[0129] The target audio segment is defined as the target audio segment data.

[0130] The model creation module acquires multiple sets of historical target audio, sets multiple sets of denoising indicators to mark the audio sources of the historical target audio, obtains audio source marking data, and creates an audio neural network model based on the audio source marking data;

[0131] Acquire multiple sets of historical target audio that have already undergone denoising, and create an audio neural network model based on the historical target audio.

[0132] It should be noted here that:

[0133] In this application, the historical target audio referred to herein is the target audio that has already undergone noise reduction.

[0134] Specifically as follows:

[0135] During the process of audio denoising of the target audio segment, several different types of audio denoising indicators are set, and these different types of audio denoising indicators are marked as Z1 denoising indicator to Zb denoising indicator respectively.

[0136] Based on the analysis of historical target audio, the Z1 denoising index to the Zb denoising index are obtained, and the baseline range of the Z1 denoising index to the Zb denoising index is obtained.

[0137] Specifically as follows:

[0138] It should be noted here that:

[0139] In this application, Z is the symbol corresponding to the denoising index, and a is the numerical value corresponding to the denoising index.

[0140] In this application, the Z1 denoising index can be the Mel frequency cepstral coefficient, the Z2 denoising index can be the spectral centroid, and the Z3 denoising index can be the spectral zero-crossing rate.

[0141] Select a sample target audio from the multiple historical target audios. Obtain the Z1 denoising index value corresponding to each time point in the sample target audio. Obtain multiple Z1 denoising index values. Compare the values ​​of the multiple Z1 denoising index values. Obtain the value range formed by the largest and smallest Z1 denoising index values ​​to obtain the Z1 denoising index value range corresponding to the sample target audio.

[0142] Repeat the process of obtaining the Z1 denoising index value range corresponding to the sample target audio, obtain the Z1 denoising index value range corresponding to each historical target audio, obtain multiple Z1 denoising index value ranges, and perform interval union on the obtained multiple Z1 denoising index value ranges to obtain the Z1 denoising index baseline range.

[0143] Repeat the process of obtaining the Z1 denoising index benchmark interval to obtain the Z2 denoising index benchmark interval to the Zb denoising index benchmark interval respectively.

[0144] Several historical audio segmentation periods are obtained in the audio recognition field. In the historical audio segmentation periods, the audio sources whose Z1 denoising index to Zb denoising index are all within the Z1 denoising index baseline range to the Zb denoising index baseline range are marked as first type audio sources. The audio sources whose denoising index is not within the corresponding denoising index range are marked as second type audio sources, thus obtaining audio source marking data.

[0145] It should be noted here that:

[0146] The first type of audio source mentioned here refers to the audio source corresponding to the historical target audio, specifically a human voice source. The second type of audio source mentioned here refers to a non-human voice source.

[0147] The audio source tagging data is divided into a denoised frequency training set and a denoised frequency test set according to the audio training-test ratio.

[0148] It should be noted here that:

[0149] In this application, the audio training-to-test ratio is specifically set to 7:3, that is, the ratio of the number of medical denoised frequencies in the denoised frequency training set to the number of medical denoised frequencies in the denoised frequency test set is 7:3.

[0150] An audio recognition model is created using an existing artificial intelligence platform. The audio recognition model is trained using a denoised frequency training set until each historical audio segment in the denoised frequency training set is used to train the audio recognition model.

[0151] The audio recognition model is tested using a denoised frequency test set, and the recognition accuracy is obtained. When the recognition accuracy is greater than or equal to the target recognition accuracy, the audio recognition model is trained and an audio neural network model is obtained. When the recognition accuracy is less than the target recognition accuracy, the audio recognition model is trained again using a denoised frequency training set until the recognition accuracy is greater than or equal to the target recognition accuracy.

[0152] It should be noted here that:

[0153] In this application, the target recognition accuracy is specifically set to 95%.

[0154] The audio noise reduction module identifies and removes audio noise for each target audio segment based on an audio neural network model.

[0155] Specifically as follows:

[0156] Obtain the audio neural network model, acquire target audio time segment data, and obtain multiple target audio segmentation time segments based on the target audio time segment data;

[0157] An audio neural network model is used to identify the second type of sound source in the target audio segment, and the second type of sound source in each target audio segment is removed as audio noise.

[0158] Thirdly, the present invention provides the application of the data recognition method described in any of the aforementioned real-time modes in audio denoising.

[0159] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for data recognition based on a neural network model, characterized in that, The method comprises the following steps: Step S1: obtaining the audio to be identified, and intercepting the sound period in which the target audio exists in the audio to be identified to obtain target audio period collection data; Step S2: obtaining a plurality of groups of historical target audio, marking the audio source of the historical target audio by setting a denoising index to obtain audio source marking data, and creating an audio neural network model according to the audio source marking data; Step S3: identifying and removing the audio noise of each target audio interception period according to the audio neural network model; In step S1, the following steps are further included: Step S11: in the process of real-time audio monitoring of the audio identification site by the audio acquisition device, marking the time point at which the audio device starts to work as the period audio start time point, marking the time point corresponding to the current time as the audio end time point, and marking the environmental recording of the audio identification site at the audio start time point and the audio end time point as the audio to be identified; Step S12: dividing the audio to be identified into a plurality of time-continuous audio short time frames, and selecting a sample audio short time frame from the plurality of audio short time frames; Step S13: performing target audio energy analysis on the sample audio short time frame, and obtaining the target audio energy ratio corresponding to the sample audio short time frame according to the analysis result; Step S14: obtaining a plurality of historical audio short time frames in which the target audio has been identified, obtaining the target audio energy ratio corresponding to each historical audio short time frame, and comparing the values of the plurality of target audio energy ratios, and marking the target audio energy ratio with the smallest value as the target audio reference energy ratio; Step S15: obtaining the target audio energy ratio of each audio short time frame, comparing the values of the obtained target audio energy ratio and the target audio reference energy ratio, and dividing the audio short time frame into an effective audio short time frame and an invalid audio short time frame according to the comparison result; Step S16: in the audio to be identified, intercepting the sound period corresponding to the effective audio short time frame to obtain a plurality of target audio interception periods; Step S17: defining the obtained target audio interception period as the target audio period collection data; In step S13, the following steps are further included: The sample audio short time frame is divided into a plurality of audio signal sampling points, and the divided audio signal sampling points are sequentially named as X1 signal sampling point to Xa signal sampling point according to the time sequence of division; The intensity values of the target audio signals corresponding to X1 signal sampling point to Xa signal sampling point are obtained respectively to obtain X1 target signal intensity value to Xa target signal intensity value; In X1 signal sampling point, the intensity values of any environmental audio signals other than the target audio signal are obtained, and the values of the plurality of environmental audio signal intensity values are compared to obtain the maximum environmental audio signal intensity value X1 environmental signal intensity value; The environmental signal intensity values corresponding to X2 signal sampling point to Xa signal sampling point are obtained respectively to obtain X2 environmental signal intensity value to Xa environmental signal intensity value; Calculate the ratio of the X1 target signal strength value and the X1 environment signal strength value to obtain the X1 target audio intensity ratio. Calculate the average of the X1 target audio intensity ratio to the Xa target audio intensity ratio to obtain the target audio energy ratio corresponding to the sample audio short-time frame.

2. The data recognition method based on a neural network model according to claim 1, characterized in that, The step S15 further includes the following specific steps: If the target audio energy ratio is greater than or equal to the target audio reference energy ratio, the corresponding audio short-time frame is divided into an effective audio short-time frame. If the target audio energy ratio is less than the target audio reference energy ratio, the corresponding audio short-time frame is divided into an invalid audio short-time frame. 3.The data recognition method based on a neural network model according to claim 1, characterized in that, The step S2 includes the following specific steps: Step S21: In the process of audio denoising on the target audio interception period, set Z1 denoising index to Zb denoising index; Step S22: Analyze the Z1 denoising index to the Zb denoising index according to the historical target audio to obtain the Z1 denoising index reference interval to the Zb denoising index reference interval; Step S23: Obtain a plurality of historical audio interception periods in the audio recognition field, and in the historical audio interception period, mark the sound source whose denoising index is in the corresponding denoising index reference interval as a first type of sound source, and mark the sound source whose denoising index is not in the corresponding denoising index reference interval as a second type of sound source, to obtain audio source marking data.

4. The data recognition method based on a neural network model according to claim 1, characterized in that, The step S2 further includes the following specific steps: Step S24: Divide the audio source marking data into a denoising audio training set and a denoising audio test set according to an audio training test ratio; Step S25: Create an audio recognition model through an existing artificial intelligence platform, train the audio recognition model using the denoising audio training set until the historical audio interception periods in the denoising audio training set are trained on the audio recognition model once; Step S26: Test the audio recognition model using the denoising audio test set, and obtain the recognition accuracy, when the recognition accuracy is greater than or equal to the target recognition accuracy, the audio recognition model training is completed, and an audio neural network model is obtained, when the recognition accuracy is less than the target recognition accuracy, continue to train the audio recognition model using the denoising audio training set until the recognition accuracy is greater than or equal to the target recognition accuracy.

5. The data recognition method based on a neural network model according to claim 3, characterized in that, The step S22 further includes the following specific steps: Obtain a plurality of groups of historical target audio that have completed denoising, select a sample target audio from the obtained plurality of historical target audio, obtain the Z1 denoising index value corresponding to each time point in the sample target audio, obtain a plurality of Z1 denoising index values, and compare the values of the obtained plurality of Z1 denoising index values, obtain the value range composed of the maximum Z1 denoising index value and the minimum Z1 denoising index value, and obtain the Z1 denoising index value interval corresponding to the sample target audio. respectively obtain a plurality of Z1 de-noising index value intervals, and obtain a Z1 de-noising index reference interval by interval union of the plurality of Z1 de-noising index value intervals; The process of obtaining the Z1 de-noising index reference interval is repeated to obtain a Z2 de-noising index reference interval to a Zb de-noising index reference interval.

6. The data recognition method based on a neural network model according to claim 1, characterized in that, The step S3 further includes the following steps: An audio neural network model is obtained, target audio time period collection data is obtained, and a plurality of target audio time period collection data is obtained. The second type of sound source in the target audio time period is identified using the audio neural network model, and the second type of sound source in each target audio time period is removed as audio noise.

7. A neural network model based data recognition system suitable for a neural network model based data recognition method according to any one of claims 1 to 6, characterized in that, The data recognition system comprises: An audio collection module: obtaining the audio to be identified, and cutting the sound period in which the target audio exists in the audio to be identified to obtain target audio time period collection data; A model creation module: obtaining a plurality of sets of historical target audio, marking the sound source of the historical target audio by setting a de-noising index to obtain audio sound source marking data, and creating an audio neural network model according to the audio sound source marking data; An audio de-noising module: identifying and removing the audio noise in each target audio time period according to the audio neural network model.

8. Application of the data recognition method of any one of claims 1-6 in audio de-noising.

Images