Intercom audio adaptive noise reduction method and system based on artificial intelligence

Abstract

The present invention provides an artificial intelligence-based walkie-talkie audio adaptive noise reduction method and system, specifically, the pre-constructed AI noise reduction basic network and the initial adaptive noise reduction network are disassembled into multiple network units, and a unique mapping relationship is established. By obtaining the first sample walkie-talkie audio sequence, each network unit of the initial adaptive noise reduction network is subjected to step-by-step network parameter learning to generate a temporary adaptive noise reduction network. Furthermore, the AI ​​noise reduction basic network and the temporary adaptive noise reduction network are subjected to interactive combined network parameter learning to generate a combined network unit including multiple first connected network units and second connected network units. Finally, by obtaining the second sample walkie-talkie audio sequence, the combined network unit is subjected to network knowledge parameter transfer learning until the network convergence requirements are met, and a target adaptive noise reduction network is generated, thereby achieving efficient and accurate noise reduction of the walkie-talkie audio and improving the communication quality.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an artificial intelligence-based walkie-talkie audio adaptive noise reduction method and system. Background Art

[0002] In the field of communication technology, walkie-talkies, as an important instant communication tool, play a key role in various scenarios such as public safety, industrial production and military communications. However, the communication process of walkie-talkies is often disturbed by environmental noise, resulting in a decrease in audio quality, which in turn affects the clarity and reliability of communication. Therefore, the development of effective walkie-talkie audio noise reduction technology has always been an important research direction in the field of communication technology.

[0003] Traditional methods for intercom audio noise reduction mainly rely on signal processing techniques, such as spectrum subtraction and Wiener filtering. Although these methods can suppress noise to a certain extent, the noise reduction effect is often limited in complex and changeable noise environments, and may introduce audio distortion. In recent years, with the rapid development of artificial intelligence technology, especially the widespread application of deep learning in the field of audio processing, new solutions have been provided for intercom audio noise reduction.

[0004] Deep learning technology can achieve effective audio noise reduction by learning the mapping relationship between a large number of noisy audio and clean audio. However, there are still many challenges in directly applying deep learning technology to walkie-talkie audio noise reduction. On the one hand, walkie-talkie audio has unique voice characteristics and noise patterns, which require targeted noise reduction network design; on the other hand, due to the real-time requirements of walkie-talkie communication, the noise reduction algorithm needs to have low computational complexity and efficient reasoning capabilities. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides an artificial intelligence-based walkie-talkie audio adaptive noise reduction method, the method comprising:

[0006] The pre-built AI denoising basic network and the initial adaptive denoising network are respectively disassembled into a plurality of network units; a unique mapping relationship is respectively established between each network unit in the pre-built AI denoising basic network and the initial adaptive denoising network;

[0007] Obtain a first sample walkie-talkie audio sequence, traverse each network unit of the initial adaptive noise reduction network, and based on the first sample walkie-talkie audio sequence, the pre-built AI noise reduction basic network, and the target network unit corresponding to the traversed network unit, perform step-by-step network parameter learning on the traversed network unit until the step-by-step network parameter learning of all network units of the initial adaptive noise reduction network is completed, and a temporary adaptive noise reduction network is generated;

[0008] Interactively combining network parameters of the pre-constructed AI denoising basic network and the temporary adaptive denoising network to generate a plurality of combined network units; the combined network unit includes a plurality of first connected network units determined from the pre-constructed AI denoising basic network, and a plurality of second connected network units determined from the temporary adaptive denoising network and corresponding to the plurality of first connected network units;

[0009] Obtain a second sample walkie-talkie audio sequence, traverse each of the multiple combined network units, and based on the second sample walkie-talkie audio sequence and the multiple first connected network units in the traversed combined network units, perform network knowledge parameter transfer learning on the multiple second connected network units in the combined network unit until each of the combined network units meets the network convergence requirements, thereby generating a target adaptive noise reduction network that completes network parameter learning.

[0010] In a possible implementation of the first aspect, obtaining a first sample walkie-talkie audio sequence, traversing each network unit of the initial adaptive noise reduction network, and performing step-by-step network parameter learning on the traversed network unit based on the first sample walkie-talkie audio sequence, the pre-built AI noise reduction basic network, and the target network unit corresponding to the traversed network unit until the step-by-step network parameter learning of all network units of the initial adaptive noise reduction network is completed, and a temporary adaptive noise reduction network is generated, including:

[0011] Obtain a first sample walkie-talkie audio sequence, load the first sample walkie-talkie audio sequence into the pre-built AI noise reduction basic network and the target network unit corresponding to the traversed network unit, use the target network unit to perform noise reduction processing on the first sample walkie-talkie audio sequence, and output the first noise reduction processing data generated by the target network unit;

[0012] Traverse each network unit of the initial adaptive noise reduction network, lock the network parameter information of other network units, and based on the first sample walkie-talkie audio sequence and the first noise reduction processing data of the target network unit, perform step-by-step network parameter learning on the network parameter information of the traversed network units until the step-by-step network parameter learning of all network units of the initial adaptive noise reduction network is completed to generate a temporary adaptive noise reduction network.

[0013] In a possible implementation of the first aspect, based on the first sample intercom audio sequence and the first noise reduction processed data of the target network unit, performing step-by-step network parameter learning on the network parameter information of the traversed network unit includes:

[0014] Loading the first sample intercom audio sequence, the first noise reduction processing data of the target network unit, and the set first mapping array into the traversed network unit, and outputting the second noise reduction processing data generated by the traversed network unit;

[0015] Performing a correspondence operation on the second noise reduction processed data of the traversed network unit and the set second mapping array, and outputting a first operation matching result;

[0016] Determining a first training error between the first operation matching result and the first noise reduction processed data of the target network unit;

[0017] If the first training error meets the convergence requirement, the step-by-step network parameter learning of the traversed network units is terminated.

[0018] In a possible implementation of the first aspect, obtaining a second sample walkie-talkie audio sequence, traversing each combined network unit in the multiple combined network units, and performing network knowledge parameter transfer learning on the multiple second connected network units in the combined network unit based on the second sample walkie-talkie audio sequence and the multiple first connected network units in the traversed combined network units, includes:

[0019] Obtain a second sample intercom audio sequence, load the second sample intercom audio sequence into a first network unit among a plurality of first connected network units in the traversed combined network unit, and generate a first network knowledge parameter transfer learning result generated by each network unit in the plurality of first connected network units;

[0020] Based on the second sample walkie-talkie audio sequence and the first network knowledge parameter transfer learning result generated by each network unit in the multiple first connected network units, network knowledge parameter transfer learning is performed on each network unit in the multiple second connected network units.

[0021] In a possible implementation of the first aspect, the performing network knowledge parameter transfer learning on each of the plurality of second connected network units based on the second sample intercom audio sequence and the first network knowledge parameter transfer learning result generated by each of the plurality of first connected network units includes:

[0022] Loading the second sample intercom audio sequence into the first network unit of the plurality of second connected network units in the traversed combined network unit, and loading the first mapping array and the first network knowledge parameter transfer learning result of each network unit in the plurality of first connected network units into each network unit in the plurality of second connected network units respectively, to generate the second network knowledge parameter transfer learning result generated by each network unit in the plurality of second connected network units;

[0023] Performing a corresponding relationship operation on the second network knowledge parameter transfer learning result of the last network unit in the plurality of second connected network units and the second mapping array, and outputting a second operation matching result;

[0024] Determine a second training error between a first network knowledge parameter transfer learning result of a last network unit in the plurality of first connected network units and a second operation matching result;

[0025] If the second training error meets the convergence requirement, the network knowledge parameter transfer learning of the traversed combined network unit is terminated.

[0026] In a possible implementation of the first aspect, the pre-built AI denoising basic network and the initial adaptive denoising network are respectively decomposed into a plurality of network units, including:

[0027] Based on the network topology relationship architecture, at least every two adjacent network function units of the pre-built AI denoising basic network are respectively disassembled into one network unit to generate multiple network units corresponding to the pre-built AI denoising basic network, and each network function unit of the initial adaptive denoising network is respectively disassembled into one network unit to generate multiple network units corresponding to the initial adaptive denoising network;

[0028] The method further comprises:

[0029] Based on the network topology relationship architecture, the pre-built AI denoising basic network and the unique mapping relationship between each network unit in the initial adaptive denoising network are constructed respectively.

[0030] In a possible implementation of the first aspect, the method further includes:

[0031] Obtain network parameter information of the target adaptive denoising network that has completed network parameter learning;

[0032] The network parameter information of the target adaptive denoising network is subjected to nonlinear coding processing, and the network parameter information of the target adaptive denoising network is nonlinearly mapped from an original numerical domain to another numerical range, so as to generate a target adaptive denoising network after configuration coding.

[0033] In a possible implementation manner of the first aspect, the step of performing nonlinear encoding processing on the network parameter information of the target adaptive denoising network, nonlinearly mapping the network parameter information of the target adaptive denoising network from an original numerical domain to another numerical range, and generating a target adaptive denoising network after configuration encoding includes:

[0034] Analyzing the network parameters of the target adaptive denoising network that has completed network parameter learning to obtain analysis results of the network parameters, the analysis results including the value distribution range of the network parameters, the data type of the network parameters, the correlation between the network parameters, and the role of the network parameters in the target adaptive denoising network;

[0035] Based on the analysis result, a corresponding target nonlinear coding function is matched from a predefined nonlinear coding function, and each network parameter in the target adaptive denoising network is encoded according to the matched target nonlinear coding function and a predefined numerical range, and after the correlation between the network parameters is jointly encoded during the encoding process, a regularization term is introduced to limit the encoded numerical range;

[0036] After verifying the encoded network parameters, configure the encoded target adaptive denoising network.

[0037] In a possible implementation of the first aspect, the method further includes:

[0038] The target adaptive noise reduction network with completed network parameter learning is used to perform noise reduction on any input walkie-talkie audio data.

[0039] On the other hand, an embodiment of the present invention also provides an artificial intelligence-based walkie-talkie audio adaptive noise reduction system, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0040] Based on the above aspects, the embodiment of the present application realizes efficient and accurate intercom audio noise reduction processing. First, the pre-constructed AI noise reduction basic network and the initial adaptive noise reduction network are disassembled into multiple network units, and a unique mapping relationship between them is established, which provides a basis for subsequent step-by-step learning and interactive combination learning. Then, using the first example intercom audio sequence, each network unit of the initial adaptive noise reduction network is subjected to step-by-step network parameter learning to generate a temporary adaptive noise reduction network. This step significantly improves the initial adaptability and noise reduction effect of the noise reduction network. Furthermore, by interactively combining network parameter learning of the pre-constructed AI noise reduction basic network and the temporary adaptive noise reduction network, a combined network unit including multiple first connection network units and second connection network units is generated. This interactive combination learning strategy not only integrates the noise reduction knowledge and experience of the basic network, but also enables the temporary adaptive noise reduction network to further improve the noise reduction performance while maintaining its unique noise reduction capability. Finally, using the second sample walkie-talkie audio sequence, network knowledge parameter transfer learning is performed on each of the multiple combined network units until each combined network unit meets the network convergence requirements, generating a target adaptive noise reduction network that has completed network parameter learning, ensuring that the target adaptive noise reduction network has excellent noise reduction effect and stability in the complex and changeable walkie-talkie audio environment. Therefore, through the strategies of step-by-step learning, interactive combination learning, and network knowledge parameter transfer learning, the adaptive noise reduction capability of walkie-talkie audio is significantly improved, providing a clearer and higher-quality audio experience for walkie-talkie communication. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The present invention is a flowchart of an implementation of an artificial intelligence-based walkie-talkie audio adaptive noise reduction method provided in an embodiment of the present invention.

[0042] Figure 2 It is a schematic diagram of the hardware architecture of an artificial intelligence-based intercom audio adaptive noise reduction system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 The present invention is a flowchart of an artificial intelligence-based intercom audio adaptive noise reduction method provided by an embodiment of the present invention. The artificial intelligence-based intercom audio adaptive noise reduction method is introduced in detail below.

[0044] Step S110, decomposing the pre-built AI denoising basic network and the initial adaptive denoising network into a plurality of network units, respectively. A unique mapping relationship is established between each network unit in the pre-built AI denoising basic network and the initial adaptive denoising network.

[0045] In the server, a complex AI noise reduction basic network and an initial adaptive noise reduction network are pre-built. In order to better carry out subsequent parameter learning and optimization, the server disassembles the AI ​​noise reduction basic network and the initial adaptive noise reduction network into multiple network units according to the network topology architecture.

[0046] Taking the AI ​​denoising basic network as an example, assume that the AI ​​denoising basic network consists of multiple convolutional layers, pooling layers, and fully connected layers. The server disassembles every two adjacent convolutional layers and the pooling layers between them into a network unit based on the network topology. For example, there are two consecutive convolutional layers with convolution kernel sizes of 3x3 and 5x5, respectively, and a 2x2 maximum pooling layer. These three parts are combined into a network unit. In this way, the AI ​​denoising basic network is disassembled into multiple network units with specific functions and structures.

[0047] For the initial adaptive noise reduction network, the server decomposes each independent network functional unit, such as a single convolutional layer, a fully connected layer, etc., into a network unit.

[0048] After the disassembly is completed, the server establishes a unique mapping relationship for each network unit in the AI ​​noise reduction basic network and the initial adaptive noise reduction network. For example, a network unit in the AI ​​noise reduction basic network is responsible for preliminary feature extraction of the audio signal, while in the initial adaptive noise reduction network, the corresponding network unit is responsible for preliminary processing and conversion of similar features. This unique mapping relationship helps to make accurate comparisons and optimizations in the subsequent parameter learning process.

[0049] Step S120, obtain the first sample walkie-talkie audio sequence, traverse each network unit of the initial adaptive noise reduction network, and based on the first sample walkie-talkie audio sequence, the pre-built AI noise reduction basic network and the target network unit corresponding to the traversed network unit, perform step-by-step network parameter learning on the traversed network unit until the step-by-step network parameter learning of all network units of the initial adaptive noise reduction network is completed, and a temporary adaptive noise reduction network is generated.

[0050] The server obtains a series of first-sample walkie-talkie audio sequences, which contain various noise conditions and speech features. Then, the server begins to traverse each network unit of the initial adaptive noise reduction network and conducts step-by-step network parameter learning.

[0051] Assuming that the first network unit of the initial adaptive noise reduction network is a convolutional layer, the server first loads the first sample walkie-talkie audio sequence obtained into the corresponding target network unit in the pre-built AI noise reduction basic network (assuming it is also a convolutional layer). This target network unit performs noise reduction on the audio sequence and outputs the first noise reduction processed data.

[0052] Then, the server locks the network parameter information of other network units in the initial adaptive noise reduction network except the currently traversed convolutional layer, and only performs network parameter learning for the current convolutional layer. Based on the first example walkie-talkie audio sequence, the first noise reduction processing data generated by the target network unit, and the set first mapping array (e.g., a set of specific weight coefficients), the currently traversed convolutional layer processes the input audio sequence and outputs the second noise reduction processing data.

[0053] The server performs a correspondence operation on the second noise reduction data output by the convolution layer and the set second mapping array, such as calculating a mean square error between the two. Through this operation, the server obtains a first operation matching result.

[0054] Next, the server determines the first training error between the first operation matching result and the first noise reduction processing data of the target network unit. If the first training error does not meet the convergence requirements (for example, the error is greater than a preset threshold), the server uses an optimization algorithm (such as a stochastic gradient descent algorithm) to adjust the network parameters of the current convolution layer (such as the weight and bias of the convolution kernel), and then performs the above processing and operation again until the first training error meets the convergence requirements.

[0055] In this way, the server traverses each network unit of the initial adaptive noise reduction network in turn, and performs similar step-by-step network parameter learning on each network unit until the learning of all network units of the initial adaptive noise reduction network is completed, and finally a temporary adaptive noise reduction network is generated.

[0056] Step S130, interactively combining network parameters of the pre-constructed AI denoising basic network and the temporary adaptive denoising network to generate a plurality of combined network units. The combined network unit includes a plurality of first connected network units determined from the pre-constructed AI denoising basic network, and a plurality of second connected network units determined from the temporary adaptive denoising network and corresponding to the plurality of first connected network units.

[0057] In this embodiment, after the server completes the step-by-step network parameter learning of the initial adaptive noise reduction network and generates a temporary adaptive noise reduction network, it starts to interactively combine network parameter learning of the pre-built AI noise reduction basic network and the temporary adaptive noise reduction network to generate multiple combined network units.

[0058] Assume that there are three network units A1, A2, and A3 in the pre-built AI noise reduction basic network, and three corresponding network units B1, B2, and B3 in the temporary adaptive noise reduction network. The server combines A1 and B1 into one combined network unit, A2 and B2 into another combined network unit, and A3 and B3 into the third combined network unit.

[0059] In each combined network unit, the network unit of the A series (such as A1) is defined as the first connected network unit, and the network unit of the B series (such as B1) is defined as the second connected network unit. In this way, the server determines the first connected network unit and the second connected network unit in multiple combined network units.

[0060] Step S140, obtain a second sample walkie-talkie audio sequence, traverse each of the multiple combined network units, and based on the second sample walkie-talkie audio sequence and the multiple first connected network units in the traversed combined network units, perform network knowledge parameter transfer learning on the multiple second connected network units in the combined network unit until each of the combined network units meets the network convergence requirements, thereby generating a target adaptive noise reduction network that completes network parameter learning.

[0061] The server obtains a new set of second sample walkie-talkie audio sequences, and then begins to traverse each of the multiple combined network units generated previously, and performs network knowledge parameter transfer learning on the second connected network units therein.

[0062] Taking one of the combined network units as an example, the server first loads the second sample walkie-talkie audio sequence into the first network unit (assuming it is A1) among the multiple first connected network units in the combined network unit. A1 processes the input audio sequence and generates a first network knowledge parameter transfer learning result.

[0063] Based on the first network knowledge parameter transfer learning result and the input second sample walkie-talkie audio sequence, the server performs network knowledge parameter transfer learning on each network unit (assuming it is B1) in the plurality of second connected network units.

[0064] Specifically, the server loads the second sample intercom audio sequence into B1, and loads the first mapping array and the first network knowledge parameter transfer learning result generated by A1 into B1 respectively. B1 processes these inputs and generates a second network knowledge parameter transfer learning result.

[0065] The server performs a corresponding relationship operation on the second network knowledge parameter transfer learning result generated by B1 and the second mapping array, such as calculating the cosine similarity between them. Through this operation, the server obtains the second operation matching result.

[0066] Then, the server determines the second training error between the first network knowledge parameter transfer learning result and the second operation matching result of the last network unit (assuming it is A3) among the multiple first connected network units. If this error meets the convergence requirement (for example, less than a preset threshold), the server considers that the combined network unit has met the network convergence requirement and ends its network knowledge parameter transfer learning.

[0067] In this way, the server traverses each combined network unit in turn, and performs similar network knowledge parameter transfer learning on the second connected network unit therein until all combined network units meet the network convergence requirements, and finally generates a target adaptive noise reduction network that completes network parameter learning.

[0068] Throughout the process, the server continuously processes large amounts of data, performs complex calculations and parameter adjustments to ensure that the generated target adaptive noise reduction network can effectively perform high-quality noise reduction on the intercom audio.

[0069] Based on the above steps, the embodiment of the present application realizes efficient and accurate intercom audio noise reduction processing. First, the pre-constructed AI noise reduction basic network and the initial adaptive noise reduction network are disassembled into multiple network units, and a unique mapping relationship between them is established, which provides a basis for subsequent step-by-step learning and interactive combination learning. Then, using the first example intercom audio sequence, each network unit of the initial adaptive noise reduction network is subjected to step-by-step network parameter learning to generate a temporary adaptive noise reduction network. This step significantly improves the initial adaptability and noise reduction effect of the noise reduction network. Furthermore, by interactively combining network parameter learning of the pre-constructed AI noise reduction basic network and the temporary adaptive noise reduction network, a combined network unit including multiple first connected network units and second connected network units is generated. This interactive combination learning strategy not only integrates the noise reduction knowledge and experience of the basic network, but also enables the temporary adaptive noise reduction network to further improve the noise reduction performance while maintaining its unique noise reduction capability. Finally, using the second sample walkie-talkie audio sequence, network knowledge parameter transfer learning is performed on each of the multiple combined network units until each combined network unit meets the network convergence requirements, generating a target adaptive noise reduction network that has completed network parameter learning, ensuring that the target adaptive noise reduction network has excellent noise reduction effect and stability in the complex and changeable walkie-talkie audio environment. Therefore, through the strategies of step-by-step learning, interactive combination learning, and network knowledge parameter transfer learning, the adaptive noise reduction capability of walkie-talkie audio is significantly improved, providing a clearer and higher-quality audio experience for walkie-talkie communication.

[0070] In a possible implementation, step S120 includes:

[0071] Step S121, obtain a first sample walkie-talkie audio sequence, load the first sample walkie-talkie audio sequence into the pre-built AI noise reduction basic network and the target network unit corresponding to the traversed network unit, use the target network unit to perform noise reduction processing on the first sample walkie-talkie audio sequence, and output the first noise reduction processing data generated by the target network unit.

[0072] Step S122, traverse each network unit of the initial adaptive noise reduction network, lock the network parameter information of other network units, and based on the first sample walkie-talkie audio sequence and the first noise reduction processing data of the target network unit, perform step-by-step network parameter learning on the network parameter information of the traversed network units until the step-by-step network parameter learning of all network units of the initial adaptive noise reduction network is completed to generate a temporary adaptive noise reduction network.

[0073] This embodiment can obtain a large number of first-example intercom audio sequences from the storage system. These first-example intercom audio sequences cover various complex noise conditions, including environmental noise, electromagnetic interference noise, human voice interference, etc., and also contain different voice features, such as different speaking speeds, intonations, accents, etc.

[0074] Assume that the first network unit in the initial adaptive noise reduction network is a convolutional layer with a specific structure. The server loads the first sample walkie-talkie audio sequence obtained into the corresponding target network unit in the pre-built AI noise reduction basic network (also assumed to be a convolutional layer, but may be different in structure and parameters from the convolutional layer in the initial adaptive noise reduction network). After receiving the audio sequence, this target network unit uses its internal convolution kernel and calculation logic to perform noise reduction on the audio.

[0075] After a series of calculations and processing, the target network unit outputs the first noise reduction processing data. This set of data reflects the initial noise reduction effect and feature extraction results of the target network unit on the input audio sequence.

[0076] Next, the server starts to traverse each network unit of the initial adaptive noise reduction network. Taking the second network unit as an example, assume that it is a fully connected layer. The server locks the network parameter information of other network units except the currently traversed fully connected layer to ensure that the parameters of other network units remain unchanged when learning the parameters of the current fully connected layer.

[0077] Based on the first sample walkie-talkie audio sequence obtained previously and the first noise reduction processed data output by the target network unit, the server performs step-by-step network parameter learning on the network parameter information of the currently traversed fully connected layer. The server inputs these data and a set specific first mapping array (e.g., a set of carefully designed weight coefficients) into the fully connected layer.

[0078] After receiving these inputs, the fully connected layer performs a series of matrix multiplication and addition operations and outputs the second denoising data. This set of data is the result calculated by the fully connected layer based on the input information and the current network parameters.

[0079] Then, the server performs a correspondence operation on the second noise reduction data output by the fully connected layer and the set second mapping array. For example, the mean square error between the two is calculated. Through complex mathematical operations, the server obtains the first operation matching result.

[0080] The server then determines a first training error between the first operation matching result and the first noise reduction processed data of the target network unit. If the error is greater than a preset threshold, it indicates that the network parameters of the current fully connected layer need to be further optimized.

[0081] The server then uses optimization algorithms such as stochastic gradient descent to adjust the network parameters of the fully connected layer, such as adjusting the connection weights and bias values ​​between neurons. After the adjustment is completed, the relevant data is input into the fully connected layer again, and the above processing and calculation process is repeated until the first training error meets the convergence requirements, that is, the error is less than or equal to the pre-set threshold.

[0082] In this way, the server traverses each network unit in the initial adaptive noise reduction network in turn, including various types of convolutional layers, pooling layers, fully connected layers, etc. Similar step-by-step network parameter learning and optimization adjustments are performed on each network unit to continuously improve the performance and noise reduction effect of the network unit.

[0083] After completing the above learning and optimization process for the last network unit of the initial adaptive noise reduction network, the server ends the step-by-step network parameter learning for all network units of the entire initial adaptive noise reduction network and generates a temporary adaptive noise reduction network.

[0084] In this process, the server needs to process massive amounts of data and perform complex mathematical operations. For each network unit, the server must precisely control the input of data, the adjustment of parameters, and the evaluation of results to ensure that the resulting temporary adaptive noise reduction network can better handle the noise problem in the intercom audio.

[0085] For example, for an audio sequence containing multiple channels and sampling points, the server needs to convert it into a matrix form suitable for network unit processing and consider the time domain and frequency domain characteristics of the audio during processing. At the same time, the server also needs to monitor the usage of computing resources in real time to ensure that the parameter learning task is completed efficiently with limited hardware resources.

[0086] Suppose that when processing a specific audio clip, the server finds that the output of the current network unit is far from the expected noise reduction effect. The server will increase the amplitude of parameter adjustment, increase the number of training iterations, or try to change the optimization algorithm to find a better network parameter configuration.

[0087] For example, when processing a series of audio sequences with similar noise characteristics, the server may find that the parameter adjustments of certain network units have certain regularities. The server will use these regularities to perform more targeted parameter initialization and adjustment strategies for subsequent similar network units in advance, thereby accelerating the speed of parameter learning and improving the learning effect.

[0088] Therefore, this embodiment continuously explores and optimizes throughout the entire process to generate a temporary adaptive noise reduction network with excellent performance, laying a foundation for subsequent further optimization and application.

[0089] In a possible implementation, step S122 includes:

[0090] Step S1221: load the first sample intercom audio sequence, the first noise reduction processing data of the target network unit, and the set first mapping array into the traversed network unit, and output the second noise reduction processing data generated by the traversed network unit.

[0091] Step S1222: performing a correspondence operation on the second noise reduction processed data of the traversed network unit and the set second mapping array, and outputting a first operation matching result.

[0092] Step S1223: determine a first training error between the first operation matching result and the first noise reduction processed data of the target network unit.

[0093] Step S1224: if the first training error meets the convergence requirement, then the step-by-step network parameter learning of the traversed network unit is terminated.

[0094] In this embodiment, the server has obtained the first sample walkie-talkie audio sequence and obtained the first noise reduction processing data of the target network unit. Next, the server will perform step-by-step network parameter learning on the traversed network units.

[0095] Assuming that the currently traversed network unit is a pooling layer with specific structure and parameters, the server first loads the first sample intercom audio sequence, the first noise reduction processed data of the target network unit and the set first mapping array into the pooling layer.

[0096] The first example walkie-talkie audio sequence is a complex audio data containing a variety of noise and voice features, such as wind, machine roar and speaker's voice, and the audio sampling frequency is 44.1kHz and the duration is 5 seconds. The first noise reduction processing data of the target network unit is the preliminary noise reduction result obtained after the pre-processing, which may contain the attenuation or enhancement information of certain frequency components in the audio. The set first mapping array can be a set of specific weight values ​​for weighted processing of the input data.

[0097] When these data are loaded into the pooling layer, the pooling layer processes them according to its internal calculation logic and parameters. After the pooling operation, such as using the maximum pooling or average pooling method, the input data is downsampled and the second noise reduction processing data generated by the traversed network unit is output.

[0098] It is assumed that the second noise reduction processing data is a feature matrix with reduced dimension, in which each element represents the processing result of the input audio in a specific area and feature dimension.

[0099] Next, the server performs a correspondence operation on the second noise reduction data output by the pooling layer and the set second mapping array. The set second mapping array can be another set of predefined weights or coefficients. For example, the performed correspondence operation can be to calculate the inner product or correlation coefficient between the two.

[0100] By performing this operation, the server obtains a first operation matching result. Assuming that this result is a value, it reflects a certain correlation or matching degree between the second noise reduction processed data and the second mapping array.

[0101] Then, the server determines a first training error between the first operation matching result and the first noise reduction processed data of the target network unit, which may involve calculating a difference, a mean square error or other indicators measuring the difference between the two.

[0102] Assuming that the first training error is calculated using a mean square error method, the server subtracts the first operation matching result and the first noise reduction processing data of the target network unit element by element, then squares and sums them, and then divides them by the total amount of data to obtain a mean square error value.

[0103] If the first training error is less than or equal to a preset convergence threshold, such as 0.01, the server considers that the parameters of the currently traversed network unit have achieved a good learning effect and meet the convergence requirements, and ends the step-by-step network parameter learning for this network unit.

[0104] However, if the first training error is greater than the convergence threshold, the server needs to further adjust the parameters of the network unit. For example, the server can use a stochastic gradient descent algorithm to update the parameters of the pooling layer, such as the size or step size of the pooling window.

[0105] Then, the server again loads the first sample walkie-talkie audio sequence, the first noise reduction processed data of the target network unit and the first mapping array into the adjusted network unit, and repeats the above processing and calculation process until the first training error meets the convergence requirement.

[0106] In the actual processing process, the server may need to perform such operations on a large number of audio data segments and continuously optimize the parameters of the network unit to improve the effect and accuracy of the noise reduction processing.

[0107] For example, for a particularly noisy audio, the server may need to adjust the parameters of the network unit multiple times to make the training error reach the convergence requirement. In the process of adjusting the parameters, the server will carefully analyze the calculation results and error change trends each time in order to optimize the parameters more specifically.

[0108] For another example, the server may process multiple first-instance walkie-talkie audio sequences from different sources simultaneously, each with unique noise characteristics and speech patterns. For different audio sequences, the initial response and training error of the network unit may be different. The server needs to take these differences into account and find a universal optimal parameter configuration that can adapt to multiple situations.

[0109] For example, when calculating the training error, the server may find that the error in certain frequency ranges is large, which may indicate that the network unit is insufficient in processing the audio features of these frequencies. The server will adjust the parameters of the network unit accordingly to enable it to have better processing capabilities for audio of specific frequencies.

[0110] Therefore, by continuously loading data, executing operations, calculating errors and adjusting parameters, the performance of the traversed network units is gradually optimized until a satisfactory convergence effect is achieved, thus laying the foundation for generating a high-quality temporary adaptive denoising network.

[0111] In a possible implementation, step S140 includes:

[0112] Step S141, obtain a second sample walkie-talkie audio sequence, load the second sample walkie-talkie audio sequence into the first network unit among multiple first connected network units in the traversed combined network unit, and generate a first network knowledge parameter transfer learning result generated by each network unit in the multiple first connected network units.

[0113] Step S142: Based on the second sample intercom audio sequence and the first network knowledge parameter transfer learning result generated by each network unit in the plurality of first connected network units, network knowledge parameter transfer learning is performed on each network unit in the plurality of second connected network units.

[0114] This embodiment has completed the previous series of operations and now enters the stage of performing network knowledge parameter transfer learning on the second connection network unit in the combined network unit based on the second sample walkie-talkie audio sequence.

[0115] The server obtains a new set of second sample walkie-talkie audio sequences from the storage system. These audio sequences have more diverse and complex noise conditions and different voice characteristics, such as signal attenuation caused by long-distance transmission and voice overlap caused by multiple people speaking at the same time.

[0116] Assuming that in the combined network unit currently being processed, the plurality of first connected network units include a convolution layer and a pooling layer. The server first loads the second sample walkie-talkie audio sequence into the first network unit, i.e., the convolution layer, among the plurality of first connected network units in the combined network unit.

[0117] After receiving the second sample walkie-talkie audio sequence, the convolution layer performs convolution operation with the input audio sequence through its internal convolution kernel. The convolution kernel slides on the audio sequence to extract and integrate audio features at different positions. After a series of complex calculations, the convolution layer generates the corresponding feature map, which is part of the learning result of the first network knowledge parameter transfer.

[0118] Next, the audio sequence and the output of the convolutional layer are passed to the pooling layer. The pooling layer downsamples the output feature map of the convolutional layer, for example, by selecting the maximum value or average value in each region, further compressing the features and reducing the amount of data. The output of the pooling layer also becomes part of the learning result of the first network knowledge parameter transfer.

[0119] With the first network knowledge parameter transfer learning results generated by each of the multiple first connected network units, the server performs network knowledge parameter transfer learning on each of the multiple second connected network units based on the second sample walkie-talkie audio sequence and these results.

[0120] Assuming that the plurality of second connection network units include a fully connected layer and an activation function layer, the server first inputs the second sample walkie-talkie audio sequence and the first network knowledge parameter transfer learning result generated by the first connection network unit into the fully connected layer.

[0121] Each neuron in the fully connected layer is connected to all the outputs of the previous layer, integrating the inputs through a weighted sum. In this process, the fully connected layer uses the knowledge obtained from the first connected network unit and the information of the second example walkie-talkie audio sequence to adjust its weights and biases to better capture the patterns and features in the audio.

[0122] Subsequently, the output of the fully connected layer is passed to the activation function layer, which performs a nonlinear transformation on the output of the fully connected layer, introducing nonlinear factors so that the network can learn more complex patterns.

[0123] During this process, the server continuously monitors and adjusts the parameters of each network unit to ensure the accuracy and effectiveness of learning. For example, the server may find that some neurons in the fully connected layer do not respond strongly enough to certain types of audio features, so it adjusts the corresponding weights to enable it to capture important information more sensitively.

[0124] For a second sample walkie-talkie audio sequence containing an emergency call and background noise, the server continuously optimizes the parameters of the second connection network unit during processing, so that the network can better distinguish the emergency call voice and background noise, thereby improving the noise reduction effect and voice clarity.

[0125] The server may simultaneously process the second sample walkie-talkie audio sequences obtained in multiple different scenarios, such as construction sites, sea rescue, and deep mountain exploration. The audio in each scenario has its own unique characteristics and challenges. For the audio in the construction site scenario, there may be strong mechanical noise; the audio in the sea rescue scenario may be interfered by the waves and wind; the audio in the deep mountain exploration scenario may have a weak signal and be accompanied by the sound of wild animals.

[0126] The server needs to flexibly adjust the parameters of the second connection network unit according to the audio characteristics of different scenarios, so that the network can adapt to various complex situations. For example, for construction site scene audio dominated by mechanical noise, the server may strengthen the parameter adjustment related to high-frequency noise suppression in the fully connected layer; for maritime rescue audio with greater interference from waves and wind, the server may focus on optimizing the parameters of the activation function layer to enhance the processing ability of low-frequency fluctuations.

[0127] In addition, the server also needs to consider the correlation and differences between different audio sequences. If it is found that multiple audio sequences have similar noise patterns, the server can use these commonalities to accelerate the parameter learning process; if an audio sequence has unique noise characteristics, the server needs to analyze and process it separately to ensure that the network can effectively learn the knowledge and parameters to deal with this special situation.

[0128] In summary, by processing the second sample walkie-talkie audio sequence, making full use of the first network knowledge parameter transfer learning results generated by multiple first connected network units, performing detailed and in-depth network knowledge parameter transfer learning on multiple second connected network units, and continuously improving the performance and adaptability of the entire combined network unit, so as to achieve a better walkie-talkie audio noise reduction effect.

[0129] In a possible implementation, step S142 includes:

[0130] Step S1421, load the second sample walkie-talkie audio sequence into the first network unit among the multiple second connected network units in the traversed combined network unit, and load the first mapping array and the first network knowledge parameter transfer learning result of each network unit in the multiple first connected network units into each network unit in the multiple second connected network units respectively, to generate the second network knowledge parameter transfer learning result generated by each network unit in the multiple second connected network units.

[0131] Step S1422, performing a corresponding relationship operation on the second network knowledge parameter transfer learning result of the last network unit in the plurality of second connected network units and the second mapping array, and outputting a second operation matching result.

[0132] Step S1423, determining a second training error between the first network knowledge parameter transfer learning result of the last network unit in the plurality of first connected network units and the second operation matching result.

[0133] Step S1424: if the second training error meets the convergence requirement, the network knowledge parameter transfer learning of the traversed combined network unit is terminated.

[0134] In this embodiment, during the process of network knowledge parameter transfer learning, the server has obtained the second sample walkie-talkie audio sequence and the first network knowledge parameter transfer learning result generated by each network unit in the plurality of first connected network units.

[0135] The server first loads the second sample walkie-talkie audio sequence into the first network unit of the multiple second connected network units in the traversed combined network unit. Assume that this first network unit is a fully connected layer with a specific structure. At the same time, the server transfers the first mapping array (which can be a set of pre-set weight coefficients) and the first network knowledge parameter transfer learning results of each network unit in the multiple first connected network units, and loads them into this fully connected layer respectively.

[0136] After receiving this rich input information, the fully connected layer begins to perform complex calculations and processing. It adjusts the connection weights between neurons based on the input data and the loaded weight coefficients, and outputs the processing results through the activation function. This processing result is part of the learning result of the second network knowledge parameter transfer generated by the fully connected layer.

[0137] Assume that the next second connection network unit is a ReLU activation function layer. The server passes the output of the fully connected layer and the corresponding first mapping array and the first network knowledge parameter transfer learning result to this activation function layer. The activation function layer performs a nonlinear transformation on the input, further enhances the network's expression ability, and outputs a new result, which is also part of the second network knowledge parameter transfer learning result.

[0138] In this way, the server processes multiple second connection network units in sequence, and each unit generates a corresponding second network knowledge parameter transfer learning result based on the input and loaded information.

[0139] When the last network unit among the plurality of second connected network units is processed, the server obtains the second network knowledge parameter transfer learning result generated by it. Then, the server performs a corresponding relationship operation on the second network knowledge parameter transfer learning result of the last network unit and the second mapping array. Assuming that the second mapping array is a set of specific standard values ​​or reference coefficients, the corresponding relationship operation performed can be to calculate the Euclidean distance, cosine similarity, etc. between the two. Through such an operation, the server outputs the second operation matching result.

[0140] Next, the server determines the second training error between the first network knowledge parameter transfer learning result of the last network unit in the plurality of first connected network units and the second operation matching result. For example, the server can calculate the sum of the squares of the differences between the two and divide it by the number of samples to obtain the mean square error.

[0141] If the second training error is less than or equal to a preset convergence threshold (for example, 0.005), the server considers that the network knowledge parameter transfer learning of the traversed combined network unit has achieved good results and meets the convergence requirements, thereby ending the network knowledge parameter transfer learning of the combined network unit.

[0142] In the actual processing process, the server may face various complex situations. For example, for a second sample walkie-talkie audio sequence recorded in a noisy factory environment, it contains the huge noise of machine operation and the shouts of workers. When the server loads the data to the second connection network unit, it needs to accurately control the input of the data and the adjustment of the parameters to ensure that valuable voice information can be effectively extracted from the noise.

[0143] For another example, for an audio sequence containing multiple frequency band features, when the server processes multiple second-connected network units, it needs to pay attention to the transmission and processing effects of information in different frequency bands in the network. If it is found that the information of certain frequency bands is lost or distorted during the transmission process, the server will adjust the parameters of the corresponding network unit in time to ensure the integrity and accuracy of the information.

[0144] For example, the server may simultaneously process multiple second-sample walkie-talkie audio sequences with similar but slightly different characteristics, such as audio from different factory floors but with similar workflows. In this case, the server will try to find the commonalities and differences between these audios in order to more efficiently adjust the network parameters and improve the network's generalization and adaptability.

[0145] If the server finds that the error is large and does not meet the convergence requirements when calculating the second training error, it will take appropriate measures to optimize it. This may include further adjusting the weights of the network units, increasing the amount of training data, trying different activation functions, or adjusting the structure of the network.

[0146] Therefore, in the entire process of network knowledge parameter transfer learning, it is necessary to finely operate and manage data, and continuously evaluate and optimize network parameters to ensure that the target adaptive noise reduction network finally generated can effectively handle various complex walkie-talkie audio situations and achieve high-quality noise reduction effects.

[0147] In a possible implementation, step S110 includes:

[0148] Step S111: Based on the network topology relationship architecture, at least every two adjacent network function units of the pre-built AI denoising basic network are respectively disassembled into one network unit to generate multiple network units corresponding to the pre-built AI denoising basic network, and each network function unit of the initial adaptive denoising network is respectively disassembled into one network unit to generate multiple network units corresponding to the initial adaptive denoising network.

[0149] The method also includes: based on the network topology relationship architecture, respectively constructing the pre-constructed AI noise reduction basic network and the unique mapping relationship between each network unit in the initial adaptive noise reduction network.

[0150] In the processing process of the above embodiment, the first thing faced is the pre-built AI denoising basic network and the initial adaptive denoising network.

[0151] For the pre-built AI denoising basic network, it may be a deep neural network composed of multiple complex layers and modules. The server disassembles it based on the network topology relationship architecture. Assume that this AI denoising basic network contains a series of convolutional layers, pooling layers, and activation function layers connected in sequence.

[0152] The server first observes the connection and interaction between adjacent network function units. For example, there are two consecutive convolutional layers. The convolution kernel size of the first convolutional layer is 3x3, the step size is 1, and a certain number of feature maps are output; the convolution kernel size of the second convolutional layer is 5x5, and the step size is 2. In between, there is also a 2x2 maximum pooling layer for downsampling. The server will disassemble these three adjacent network function units, namely the two convolutional layers and the middle pooling layer, into one network unit. In this way, the server gradually disassembles the entire AI denoising basic network into multiple network units with specific functions and structures according to the data flow and functional dependencies between the layers in the network. Finally, multiple network units corresponding to the pre-built AI denoising basic network are generated.

[0153] For the initial adaptive noise reduction network, its structure may be relatively simple or have different characteristics. The server also relies on the network topology, but in this case, due to its structural characteristics, the server separates each independent network function unit of the initial adaptive noise reduction network, such as a single fully connected layer, a specific activation function layer, or a simple convolutional layer, into a network unit. In this way, multiple network units corresponding to the initial adaptive noise reduction network are generated.

[0154] After completing the disassembly of the two networks, the server must also build a unique mapping relationship between the pre-built AI noise reduction basic network and the initial adaptive noise reduction network based on the network topology relationship architecture.

[0155] For example, in the pre-built AI noise reduction basic network, a disassembled network unit is responsible for preliminary feature extraction and compression of the audio signal, and its output feature information needs to be passed to the subsequent specific network unit for further processing and optimization. The server will clarify this transfer relationship and establish a unique mapping from the network unit responsible for preliminary feature extraction and compression to the subsequent processing network unit.

[0156] For example, in the initial adaptive noise reduction network, a disassembled network unit is responsible for performing a certain transformation operation on the input audio data, and its result needs to be accurately passed to another network unit responsible for integration and decision-making. The server will clearly define the relationship between data transmission and functional connection to ensure that the output of each network unit can be passed to the corresponding next network unit according to the predetermined path and method.

[0157] This provides a clear structure and path for subsequent network parameter learning, optimization and integration.

[0158] In actual operation, the server may process an extremely complex AI noise reduction basic network, which includes convolutional layers of various types and sizes, recurrent neural network layers, and attention mechanism modules. For such a network, the server needs to have a deep understanding of the functions and effects of each layer, as well as the relationship between them, in order to accurately disassemble and map them.

[0159] For example, there is a network part with multiple branch structures, each branch is composed of different numbers and types of layers. The server will carefully analyze the input and output of each branch, as well as their interaction with other branches, and then disassemble these branches into appropriate network units according to function and data flow, and establish precise mapping relationships to ensure that the flow and processing of information in the network is accurate and efficient.

[0160] For another example, the initial adaptive noise reduction network may contain some network function units optimized for specific noise types or audio scenarios. When the server disassembles and establishes mapping relationships, it will fully consider the particularity and pertinence of these units to ensure that they can play their due role in the subsequent learning and optimization process, and that the collaboration with other network units is coordinated and effective.

[0161] In other words, by decomposing the network topology relationship architecture and constructing a unique mapping relationship, a solid foundation is laid for achieving efficient and accurate network learning and optimization.

[0162] In a possible implementation, the method further includes:

[0163] Step S150, obtaining network parameter information of the target adaptive denoising network that has completed network parameter learning.

[0164] Step S160, performing nonlinear coding processing on the network parameter information of the target adaptive denoising network, nonlinearly mapping the network parameter information of the target adaptive denoising network from the original numerical domain to another numerical range, and generating a target adaptive denoising network after configuration coding.

[0165] In a possible implementation, step S160 includes:

[0166] Step S161, analyzing the network parameters of the target adaptive noise reduction network that has completed network parameter learning to obtain analysis results of the network parameters, wherein the analysis results include the numerical distribution range of the network parameters, the data type of the network parameters, the correlation between the network parameters, and the role of the network parameters in the target adaptive noise reduction network.

[0167] Step S162, based on the analysis result, a corresponding target nonlinear coding function is matched from a predefined nonlinear coding function, and each network parameter in the target adaptive denoising network is encoded according to the matched target nonlinear coding function and a predefined numerical range. After the correlation between the network parameters is jointly encoded during the encoding process, a regularization term is introduced to limit the encoded numerical range.

[0168] Step S163, after verifying the encoded network parameters, configure the encoded target adaptive noise reduction network.

[0169] In this embodiment, after completing the network parameter learning of the target adaptive noise reduction network, the server will obtain all network parameter information of the network. These parameters cover many detailed values ​​such as the weights, biases, and connection strengths between neurons of each layer in the network.

[0170] Next, the server starts to perform nonlinear encoding processing on these network parameter information. First, the server performs an in-depth analysis of the network parameters of the target adaptive denoising network that has completed network parameter learning.

[0171] For example, the server found that the values ​​of some network parameters have a wide range, from very small values ​​to large values; some parameters are of floating point data type, while others may be integers. At the same time, the server also detected strong correlations between some parameters, such as the weights of some convolutional layers showing a trend of synchronous changes, while other parameters are mainly responsible for suppressing high-frequency noise in the target adaptive noise reduction network, and their role and effect have specific patterns.

[0172] Based on these analysis results, the server matches the corresponding target nonlinear coding function from a series of predefined nonlinear coding functions. For example, for the parameters responsible for suppressing high-frequency noise, an exponential nonlinear coding function is matched; for the parameter groups with strong correlation, a coding function that can consider the joint correlation is selected.

[0173] Then, the server encodes each network parameter in the target adaptive noise reduction network according to the matching target nonlinear coding function and the predefined value range. During the encoding process, the server will perform joint encoding for the correlated network parameters, taking into account their mutual relationship to achieve more effective compression and optimization.

[0174] At the same time, in order to avoid excessive fluctuations or anomalies in the encoded values, the server introduces regularization terms to limit the range of the encoded values. For example, by setting upper and lower limits, it ensures that the encoded parameter values ​​do not exceed a reasonable range, thereby maintaining the stability and reliability of the network.

[0175] After encoding, the server will strictly verify the parameters of the encoded network. The server will use a set of test data to input into the encoded target adaptive noise reduction network to observe whether the output results meet expectations.

[0176] Suppose the server uses a series of walkie-talkie audio data containing different noise types and intensities as a test set. By comparing the network's processing effect on these test data before and after encoding, the server evaluates whether encoding has a positive impact on noise reduction performance.

[0177] If the verification results show that the encoded target adaptive denoising network can effectively process the test data and meets or exceeds expectations in terms of noise reduction effect and computational efficiency, the server will configure the encoded target adaptive denoising network to prepare it for actual application.

[0178] For example, when processing a piece of walkie-talkie audio from a construction site, the encoded network was able to accurately identify and reduce the intense noise generated by large machinery while clearly preserving the conversations between workers.

[0179] For example, for a piece of walkie-talkie audio recorded under extreme weather conditions, including the sounds of strong winds and heavy rain, the encoded network can effectively remove these environmental noises and highlight the important voice information.

[0180] During the whole process, the server may encounter various situations. For example, some encoding methods may cause the network to handle noise of a certain frequency poorly. The server will promptly discover and readjust the encoding function or parameters to optimize network performance.

[0181] For example, during the verification phase, if the encoded network is found to have errors or performance degradation when processing certain complex multi-source sound scenes, the server will conduct an in-depth analysis of the cause, which may be due to the regularization term being set too strictly or too loosely, or the encoding function being inappropriately selected. The server will then make targeted improvements and optimizations accordingly until the encoded target adaptive noise reduction network performs well in various test scenarios.

[0182] Therefore, it can be ensured that the target adaptive noise reduction network can serve the noise reduction task of the intercom audio in a more efficient and stable manner after the nonlinear coding processing.

[0183] In a possible implementation manner, this embodiment can finally use the target adaptive noise reduction network that has completed network parameter learning to perform noise reduction processing on any input intercom audio data.

[0184] In this embodiment, the target adaptive noise reduction network has been constructed and parameter learning has been completed, and nonlinear coding processing and verification configuration have been performed on it. Next, the server will apply this optimized target adaptive noise reduction network to actual intercom audio data noise reduction processing.

[0185] The server receives any input walkie-talkie audio data from various scenarios. These data may come from different environments and scenarios such as construction sites, fire rescue sites, ships sailing on the sea, and deep mountain expedition teams.

[0186] Suppose that the first thing received is a piece of walkie-talkie audio data from a construction site. This audio is full of the roar of various large machinery, the shouts of workers, and other environmental background noise. The server inputs this audio data into the target adaptive noise reduction network.

[0187] The various network units in the target adaptive noise reduction network begin to work together. The convolution layer in the network first extracts features from the input audio data to capture various patterns and frequency components in the audio. Then, the pooling layer downsamples the extracted features to reduce the amount of data while retaining key information.

[0188] During the processing, the network will identify and suppress noises of different frequencies and intensities based on the learned parameters and patterns. For example, for low-frequency and high-intensity noises such as mechanical roar, the network will adjust the corresponding weights and parameters based on previous learning experience to effectively attenuate it.

[0189] For voice signals such as workers' shouts that contain important information, the network will try its best to retain their original features and frequency components to ensure the integrity and clarity of the information.

[0190] Suppose the server then receives a walkie-talkie audio clip from a fire rescue scene. It contains the whirring sound of a burning fire, the sound of rescue equipment running, and the tense conversations between firefighters. The target adaptive noise reduction network will quickly analyze these complex sound features.

[0191] For high-frequency and sharp sounds of fire burning, the network adjusts the parameters of the filter to appropriately weaken them. At the same time, for key instructions and communication voices between firefighters, the network will enhance and highlight them to better convey important information.

[0192] When processing intercom audio from a ship sailing at sea, you may encounter the sound of waves, wind, and communication between crew members. The target adaptive noise reduction network can accurately distinguish these sounds and focus on processing environmental noise such as waves and wind to improve the quality of crew communication voice.

[0193] For the intercom audio of a deep mountain expedition team, which may contain birdsong, stream sound, and team members' shouts, the target adaptive noise reduction network can reduce the interference of natural environmental sounds while retaining the team members' shouts.

[0194] When processing the audio data of these intercoms in different scenarios, the server will monitor the noise reduction effect in real time. If it is found that the noise reduction processing of some parts is not ideal, such as the noise of a certain frequency is not fully suppressed, or important voice information is distorted, the server will fine-tune the parameters of the network based on the feedback information to further optimize the noise reduction performance.

[0195] In the process of processing a large amount of intercom audio data, the server will also accumulate various statistical information and experience. For example, the server will find that certain types of noise have similar characteristics in different scenarios, and can optimize the network parameter settings in advance to improve processing efficiency.

[0196] At the same time, the server will continue to improve and perfect the performance of the target adaptive noise reduction network based on user feedback and actual application needs, so that it can better adapt to various complex and changeable walkie-talkie audio environments and provide users with clearer and more accurate audio communication services.

[0197] Therefore, the target adaptive noise reduction network that has completed network parameter learning can be fully utilized to efficiently and accurately reduce the noise of any input walkie-talkie audio data from different scenarios and environments, continuously improving service quality and user experience.

[0198] Figure 2 The hardware structure of the intercom audio adaptive noise reduction system 100 based on artificial intelligence for implementing the intercom audio adaptive noise reduction method based on artificial intelligence provided by an embodiment of the present invention is shown as follows: Figure 2As shown, the artificial intelligence-based intercom audio adaptive noise reduction system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0199] In one possible design, the intercom audio adaptive noise reduction system 100 based on artificial intelligence can be a single server or a server group. The server group can be centralized or distributed (for example, the intercom audio adaptive noise reduction system 100 based on artificial intelligence can be a distributed system). In some embodiments, the intercom audio adaptive noise reduction system 100 based on artificial intelligence can be local or remote. For example, the intercom audio adaptive noise reduction system 100 based on artificial intelligence can access information and / or data stored in the machine-readable storage medium 120 via a network. For another example, the intercom audio adaptive noise reduction system 100 based on artificial intelligence can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the intercom audio adaptive noise reduction system 100 based on artificial intelligence can be implemented on an intercom audio adaptive noise reduction system based on artificial intelligence. By way of example only, the intercom audio adaptive noise reduction system based on artificial intelligence can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any combination thereof.

[0200] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the artificial intelligence-based intercom audio adaptive noise reduction system 100 uses to execute or use to complete the exemplary method described in the present invention.

[0201] In the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the artificial intelligence-based walkie-talkie audio adaptive noise reduction method of the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected via the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0202] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned artificial intelligence-based intercom audio adaptive noise reduction system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0203] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned artificial intelligence-based walkie-talkie audio adaptive noise reduction method is implemented.

[0204] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. An artificial intelligence-based walkie-talkie audio adaptive noise reduction method, characterized in that: The method comprises: The pre-built AI denoising basic network and the initial adaptive denoising network are respectively disassembled into a plurality of network units; a unique mapping relationship is respectively established between each network unit in the pre-built AI denoising basic network and the initial adaptive denoising network; Obtain a first sample walkie-talkie audio sequence, traverse each network unit of the initial adaptive noise reduction network, and based on the first sample walkie-talkie audio sequence, the pre-built AI noise reduction basic network, and the target network unit corresponding to the traversed network unit, perform step-by-step network parameter learning on the traversed network unit until the step-by-step network parameter learning of all network units of the initial adaptive noise reduction network is completed, and a temporary adaptive noise reduction network is generated; Interactively combining network parameters of the pre-constructed AI denoising basic network and the temporary adaptive denoising network to generate a plurality of combined network units; the combined network unit includes a plurality of first connected network units determined from the pre-constructed AI denoising basic network, and a plurality of second connected network units determined from the temporary adaptive denoising network and corresponding to the plurality of first connected network units; Obtain a second sample walkie-talkie audio sequence, traverse each of the multiple combined network units, and based on the second sample walkie-talkie audio sequence and the multiple first connected network units in the traversed combined network units, perform network knowledge parameter transfer learning on the multiple second connected network units in the combined network unit until each of the combined network units meets the network convergence requirements, thereby generating a target adaptive noise reduction network that completes network parameter learning; The step of obtaining a first sample intercom audio sequence, traversing each network unit of the initial adaptive noise reduction network, and performing step-by-step network parameter learning on the traversed network unit based on the first sample intercom audio sequence, the pre-built AI noise reduction basic network, and the target network unit corresponding to the traversed network unit, until the step-by-step network parameter learning of all network units of the initial adaptive noise reduction network is completed, and a temporary adaptive noise reduction network is generated, including: Obtain a first sample walkie-talkie audio sequence, load the first sample walkie-talkie audio sequence into the pre-built AI noise reduction basic network and the target network unit corresponding to the traversed network unit, use the target network unit to perform noise reduction processing on the first sample walkie-talkie audio sequence, and output the first noise reduction processing data generated by the target network unit; Traversing each network unit of the initial adaptive noise reduction network, locking the network parameter information of other network units, and performing step-by-step network parameter learning on the network parameter information of the traversed network units based on the first sample intercom audio sequence and the first noise reduction processing data of the target network unit, until the step-by-step network parameter learning of all network units of the initial adaptive noise reduction network is completed, and a temporary adaptive noise reduction network is generated; Based on the first sample intercom audio sequence and the first noise reduction processed data of the target network unit, performing step-by-step network parameter learning on the network parameter information of the traversed network unit, including: Loading the first sample intercom audio sequence, the first noise reduction processing data of the target network unit, and the set first mapping array into the traversed network unit, and outputting the second noise reduction processing data generated by the traversed network unit; Performing a correspondence operation on the second noise reduction processed data of the traversed network unit and the set second mapping array, and outputting a first operation matching result; Determining a first training error between the first operation matching result and the first noise reduction processed data of the target network unit; If the first training error meets the convergence requirement, the step-by-step network parameter learning of the traversed network units is terminated.

2. The method for adaptive noise reduction of intercom audio based on artificial intelligence according to claim 1 is characterized in that: The obtaining of the second sample intercom audio sequence, traversing each of the multiple combined network units, and performing network knowledge parameter transfer learning on the multiple second connected network units in the combined network unit based on the second sample intercom audio sequence and the multiple first connected network units in the traversed combined network units, includes: Obtain a second sample intercom audio sequence, load the second sample intercom audio sequence into a first network unit among a plurality of first connected network units in the traversed combined network unit, and generate a first network knowledge parameter transfer learning result generated by each network unit in the plurality of first connected network units; Based on the second sample walkie-talkie audio sequence and the first network knowledge parameter transfer learning result generated by each network unit in the multiple first connected network units, network knowledge parameter transfer learning is performed on each network unit in the multiple second connected network units.

3. The method for adaptive noise reduction of intercom audio based on artificial intelligence according to claim 2 is characterized in that: The performing network knowledge parameter transfer learning on each of the plurality of second connected network units based on the second sample intercom audio sequence and the first network knowledge parameter transfer learning result generated by each of the plurality of first connected network units comprises: Loading the second sample intercom audio sequence into the first network unit of the plurality of second connected network units in the traversed combined network unit, and loading the first mapping array and the first network knowledge parameter transfer learning result of each network unit in the plurality of first connected network units into each network unit in the plurality of second connected network units respectively, and generating the second network knowledge parameter transfer learning result generated by each network unit in the plurality of second connected network units; Performing a corresponding relationship operation on the second network knowledge parameter transfer learning result of the last network unit in the plurality of second connected network units and the second mapping array, and outputting a second operation matching result; Determine a second training error between a first network knowledge parameter transfer learning result of a last network unit in the plurality of first connected network units and a second operation matching result; If the second training error meets the convergence requirement, the network knowledge parameter transfer learning of the traversed combined network unit is terminated.

4. The method for adaptive noise reduction of intercom audio based on artificial intelligence according to claim 1 is characterized in that: The pre-built AI denoising basic network and the initial adaptive denoising network are respectively disassembled into multiple network units, including: Based on the network topology relationship architecture, at least every two adjacent network function units of the pre-built AI denoising basic network are respectively disassembled into one network unit to generate multiple network units corresponding to the pre-built AI denoising basic network, and each network function unit of the initial adaptive denoising network is respectively disassembled into one network unit to generate multiple network units corresponding to the initial adaptive denoising network; The method further comprises: Based on the network topology relationship architecture, the pre-built AI denoising basic network and the unique mapping relationship between each network unit in the initial adaptive denoising network are constructed respectively.

5. The method for adaptive noise reduction of intercom audio based on artificial intelligence according to claim 1 is characterized in that: The method further comprises: Obtain network parameter information of the target adaptive denoising network that has completed network parameter learning; The network parameter information of the target adaptive denoising network is subjected to nonlinear coding processing, and the network parameter information of the target adaptive denoising network is nonlinearly mapped from an original numerical domain to another numerical range, so as to generate a target adaptive denoising network after configuration coding.

6. The method for adaptive noise reduction of intercom audio based on artificial intelligence according to claim 5 is characterized in that: The step of performing nonlinear encoding processing on the network parameter information of the target adaptive denoising network, nonlinearly mapping the network parameter information of the target adaptive denoising network from an original numerical domain to another numerical range, and generating a target adaptive denoising network after configuration encoding comprises: Analyzing the network parameters of the target adaptive denoising network that has completed network parameter learning to obtain analysis results of the network parameters, the analysis results including the value distribution range of the network parameters, the data type of the network parameters, the correlation between the network parameters, and the role of the network parameters in the target adaptive denoising network; Based on the analysis result, a corresponding target nonlinear coding function is matched from a predefined nonlinear coding function, and each network parameter in the target adaptive denoising network is encoded according to the matched target nonlinear coding function and a predefined numerical range, and after the correlation between the network parameters is jointly encoded during the encoding process, a regularization term is introduced to limit the encoded numerical range; After verifying the encoded network parameters, configure the encoded target adaptive denoising network.

7. The method for adaptive noise reduction of intercom audio based on artificial intelligence according to claim 6 is characterized in that: The method further comprises: The target adaptive noise reduction network with completed network parameter learning is used to perform noise reduction on any input walkie-talkie audio data.

8. An artificial intelligence-based intercom audio adaptive noise reduction system, characterized in that: The artificial intelligence-based intercom audio adaptive noise reduction system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the artificial intelligence-based intercom audio adaptive noise reduction method described in any one of claims 1 to 7.

Images