Speech error correction method and system based on edge computing

Abstract

The application provides a speech correction method and system based on edge computing, which comprises the following steps: obtaining speech data and environment data, aligning the speech data and the environment data, and forming speech data samples with environment labels; constructing a relationship graph based on a graph neural network according to the speech data samples with environment labels; constructing a speech correction model that fuses graph information, inputting the relationship graph into the speech correction model for training to obtain a target model, wherein the target model is deployed to an edge device to realize real-time inference by using edge computing; when the edge device receives new speech data and environment data, a corresponding relationship graph is constructed and input into the target model to generate corrected text output, wherein the target model dynamically adjusts the correction strategy according to the graph features to judge the current environment state, and specifically, the accuracy and robustness of edge speech correction are effectively enhanced through deep cooperation of speech and environment data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of speech error correction, and particularly relates to a speech error correction method and system based on edge computing. BACKGROUND

[0002] Under the background of rapid development of industrial automation and Internet of Things, speech interaction technology is widely used in industrial workshop scenarios due to its convenience and efficiency. Operators control equipment, record inspection information, etc. through voice commands, which significantly improves work efficiency.

[0003] It should be noted that the traditional speech recognition technology has high recognition accuracy, but in the application scenario of the industrial workshop, the effective consideration of environmental factors is often lacking, that is, when in a noisy industrial workshop, the sound of machine roar and the noise of equipment operation will seriously interfere with the speech signal, resulting in a sharp increase in speech recognition error rate. SUMMARY

[0004] Therefore, in the embodiments of the present application, a speech error correction method and system based on edge computing are provided, which aims to combine environmental data to accurately implement speech error correction in a complex environment, so as to meet the strict requirements of reliability and accuracy of speech interaction in industrial scenarios.

[0005] The first aspect of the embodiments of the present application provides a speech error correction method based on edge computing, which is applied to an industrial workshop scenario, and an edge device is installed in the industrial workshop. The method comprises:

[0006] Obtaining speech data and environmental data, and aligning the speech data and the environmental data to form speech data samples with environmental labels;

[0007] According to the speech data samples with environmental labels, a relationship graph based on a graph neural network is constructed;

[0008] A speech error correction model that fuses graph information is constructed, the relationship graph is input into the speech error correction model for training to obtain a target model, wherein the target model is deployed to the edge device, and edge computing is used to realize real-time inference;

[0009] When the edge device receives new speech data and environmental data, a corresponding relationship graph is constructed and input into the target model to generate corrected text output, wherein the target model dynamically adjusts the error correction strategy according to the graph features to determine the current environmental state.

[0010] Further, the step of obtaining speech data and environmental data, and aligning the speech data and the environmental data to form speech data samples with environmental labels comprises:

[0011] Acquire voice data, use beamforming technology to capture human voice directionally, and use short-time Fourier transform to convert time-domain voice signal into frequency-domain mel spectrum graph;

[0012] According to the frequency-domain mel spectrum graph, extract acoustic features;

[0013] Acquire environmental data including at least temperature, humidity, environmental noise decibel value, equipment speed, and load, and normalize the environmental data;

[0014] Use timestamp alignment algorithm to match the acoustic features under the voice frame and the normalized environmental data to form voice data samples with environmental labels.

[0015] Further, in the step of constructing a relationship graph based on a graph neural network according to the voice data samples with environmental labels, the relationship graph is composed of nodes and edges, the nodes include voice nodes and environmental nodes, and the acoustic features under the voice frame are taken as node attributes, and each voice frame corresponds to a voice node;

[0016] Each environmental parameter is taken as a node separately, and the acquisition timestamp of the corresponding environmental parameter is taken as an attribute;

[0017] Each voice node establishes a non-directed edge with all environmental nodes at the same time, and the weight is dynamically adjusted according to the influence degree of environmental factors on voice;

[0018] Between environmental nodes, edges are established between preset environmental parameters, and the edge weight is calculated by Pearson correlation coefficient to represent the correlation between parameters;

[0019] In addition, the relationship graph is graph structure encoded, and the graph structure is encoded into a low-dimensional vector representation using a graph embedding algorithm to preserve the topological relationship between voice and environmental factors.

[0020] Further, the voice correction model fused with graph information includes a graph feature extraction layer and an attention fusion layer, wherein the graph feature extraction layer uses GNN to extract features from the relationship graph, outputs high-order semantic features of each node, and the attention fusion layer fuses the acoustic features of the voice frame and the environmental graph features of the corresponding node through a multi-head attention mechanism to generate a voice representation containing environmental perception. In addition, an adaptive weight function is designed to dynamically adjust the weight of the noise reduction module according to the feature value of the environmental node.

[0021] Further, in the step of inputting the relationship graph into the voice correction model for training to obtain a target model, the cross-entropy loss of voice correction and the Laplacian regularization loss of graph structure are combined to ensure that the model corrects voice errors while maintaining the semantic consistency of the graph structure, and the loss function expression is:

[0022] ;

[0023] in, For cross-entropy loss in speech error correction, For graph structure loss, This is the balance coefficient.

[0024] Furthermore, the step of dynamically adjusting the error correction strategy based on graph features to determine the current environmental state includes:

[0025] When the target model determines that the current environment is a preset noise environment based on graph features, the noise reduction module will be triggered. Specifically, a noise reduction algorithm based on deep learning is used to process the noisy speech features, and then the noise-reduced speech features are input into the target model to generate corrected text output.

[0026] When the target model determines that the current environmental state is the preset operating state of the device based on the graph features, for voice commands involving device operation, the device operating parameters in the environmental data are used to perform semantic verification on the commands.

[0027] The risk of instruction execution is assessed through a pre-defined rule base and machine learning model.

[0028] A second aspect of this invention provides a speech error correction system based on edge computing power, used to implement the speech error correction method based on edge computing power provided in the first aspect of this invention, the system comprising:

[0029] The acquisition module is used to acquire voice data and environmental data, and to align the voice data and environmental data to form a voice data sample with environmental labels;

[0030] The module is used to construct a relational graph based on a graph neural network from voice data samples with environmental labels.

[0031] The training module is used to construct a speech error correction model that integrates graph information. The relationship graph is input into the speech error correction model for training to obtain a target model. The target model is deployed to the edge device to realize real-time inference using edge computing power.

[0032] The input module is used to construct a corresponding relationship graph and input it into the target model when the edge device receives new voice data and environmental data, and generate corrected text output. The target model dynamically adjusts the error correction strategy based on the graph features to determine the current environmental state.

[0033] A third aspect of the present invention provides a computer-readable storage medium, comprising:

[0034] The readable storage medium stores one or more programs that, when executed by a processor, implement the edge computing-based speech error correction method as described in the first aspect.

[0035] A fourth aspect of the present invention provides an electronic device, the electronic device including a memory and a processor, wherein:

[0036] The memory is used to store computer programs;

[0037] When the processor executes the computer program stored in the memory, it implements the speech error correction method based on edge computing power as described in the first aspect.

[0038] This invention provides a speech error correction method and system based on edge computing power. The method involves acquiring speech data and environmental data, aligning them to form speech data samples with environmental labels, constructing a relational graph based on a graph neural network, building a speech error correction model that integrates graph information, and training the model with the relational graph to obtain a target model. This target model is then deployed to an edge device, utilizing edge computing power for real-time inference. When the edge device receives new speech and environmental data, it constructs a corresponding relational graph and inputs it into the target model to generate corrected text output. The target model dynamically adjusts its error correction strategy based on graph features to determine the current environmental state. Specifically, through deep collaboration between speech and environmental data, the accuracy and robustness of edge-side speech error correction are effectively enhanced. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the implementation of a speech error correction method based on edge computing power, provided in Embodiment 1 of the present invention.

[0040] Figure 2 This is a structural block diagram of a speech error correction system based on edge computing power provided in Embodiment 2 of the present invention;

[0041] Figure 3 This is a structural block diagram of an electronic device provided in Embodiment 3 of the present invention. Detailed Implementation

[0042] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0043] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0045] Example 1

[0046] Embodiment 1 of this invention provides a voice error correction method based on edge computing power, applied in an industrial workshop scenario. Edge devices are installed in the industrial workshop; these edge devices refer to intelligent hardware devices located at the network edge layer, close to the data acquisition source or user terminal. They possess certain localized computing, storage, and communication capabilities, and can process data directly at the "edge" close to the data source, without needing to upload all data to the cloud. This reduces latency, saves bandwidth, and maintains functional availability in offline or weak network environments. Examples of edge devices include intelligent sensors (such as microphone arrays and environmental sensors), intelligent cameras (integrated with voice acquisition functions), etc. Please refer to [link to relevant documentation]. Figure 1 This is a flowchart of the implementation of a speech error correction method based on edge computing power, specifically including steps S01 to S04.

[0047] Step S01: Acquire voice data and environmental data, and align the voice data and environmental data to form a voice data sample with environmental tags.

[0048] Specifically, the process involves acquiring speech data, using beamforming technology to directionally capture human voices, and then using short-time Fourier transform to convert the time-domain speech signal into a frequency-domain Mel-spectrum. The formula is as follows:

[0049] ;

[0050] in, It is a time-domain speech signal. Here, m is the frame index, k is the frequency index, and T is the frame length;

[0051] Based on the frequency domain Mel spectrogram, acoustic features are extracted. These acoustic features can be MFCC (Melbourne Frequency Cepstral Coefficients), PLP (Perceptual Linear Prediction Coefficients), etc. Specifically, low-frequency features are extracted by taking the logarithm of the Mel spectrogram and then performing a discrete cosine transform, as shown in the formula:

[0052] ;

[0053] Where S(k) is the Mel spectrum, K is the number of Mel filters, and n is the MFCC coefficient index;

[0054] Acquire environmental data including at least temperature, humidity, ambient noise levels (decibels), equipment speed, and load, and normalize the environmental data. It is understood that equipment speed and load are equipment operating parameters. The normalization formula is as follows:

[0055] X represents the original data. min and X max These are the minimum and maximum values ​​of the data, respectively.

[0056] A timestamp alignment algorithm is used to match the acoustic features of the speech frame with the normalized environmental data to form speech data samples with environmental labels. Specifically, a time window Δt (e.g., 50ms) is defined, and the environmental data E(t) is aligned with the speech frame V(t) according to the timestamp t, which can be expressed as:

[0057] .

[0058] Step S02: Construct a relational graph based on a graph neural network based on voice data samples with environmental labels.

[0059] It should be noted that the relationship graph consists of nodes and edges. The nodes include speech nodes and environment nodes. The acoustic features of the speech frame are used as node attributes, and each speech frame corresponds to one speech node. Taking MFCC features as an example, the MFCC features are used as speech node v. voice The attribute, with dimension D voice , can be represented as:

[0060] ;

[0061] Each environmental parameter is treated as a separate node, and its acquisition timestamp is assigned as an attribute. For example, the normalized environmental parameter is used as the environmental node v. env Attributes, such as noise decibel value d Bnorm Temperature T norm Etc. can be represented as:

[0062] ;

[0063] Each speech node establishes an undirected edge with all environmental nodes at the same time, and the weights are dynamically adjusted according to the degree of influence of environmental factors on speech, such as the noise weight w. noise It can be represented as: It is understandable that the higher the noise level in decibels, the greater the weight of the corresponding edge.

[0064] Between environmental nodes, edge connections are established for preset environmental parameters, such as temperature and humidity, and equipment speed and current in the load. The edge weights are calculated using the Pearson correlation coefficient to characterize the correlation between parameters.

[0065] In addition, graph structure encoding is performed on the relationship graph. A graph embedding algorithm is used to encode the graph structure into a low-dimensional vector representation, preserving the topological relationships between speech and environmental factors. Specifically, GraphSAGE aggregation is used for graph structure encoding, and the current node representation is updated by aggregating neighbor node features. The formula is as follows:

[0066] ;

[0067] in, Let N(v) be the feature vector of node v at layer l, and N(v) be the set of neighboring nodes. Let σ be the learnable weight matrix, and σ be the activation function.

[0068] Step S03: Construct a speech error correction model that integrates graph information, input the relationship graph into the speech error correction model for training, and obtain the target model.

[0069] The target model is deployed to the edge device, utilizing edge computing power to achieve real-time inference. Specifically, the speech error correction model that integrates graph information includes a graph feature extraction layer and an attention fusion layer. The graph feature extraction layer uses a GNN (such as GCN or GAT) to extract features from the relation graph, outputting high-order semantic features for each node. It should be noted that the propagation rule of a graph convolutional network (GCN) can be expressed as:

[0070] ;

[0071] in, To add a self-loop adjacency matrix, for The degree matrix of the diagonal nodes, Let l be the feature matrix of the nodes in the l-th layer. This is the weight matrix.

[0072] The attention fusion layer uses a multi-head attention mechanism to fuse the acoustic features of the speech frame with the environmental graph features of the corresponding node, generating a speech representation that includes environmental awareness. The multi-head attention mechanism is used to calculate the weighted sum of the speech feature Q and the environmental graph features K and V, expressed as:

[0073] ;

[0074] ;

[0075] ;

[0076] Concat() is the concatenation operation. This is the weight matrix. For the first i The calculation results of each attention head, , , Let Q, K, and V be the linear transformation matrices. Let K be the dimension.

[0077] Furthermore, an adaptive weighting function is designed to dynamically adjust the weights of the noise reduction module based on the feature values ​​of the environmental nodes. For example, when the decibel value exceeds a threshold, the weights of the noise reduction module are increased from 0.3 to 0.8 using a sigmoid function, thereby enhancing the noise suppression capability in the speech signal. In this embodiment of the invention, the dynamic adjustment of the noise reduction module weights based on the noise decibel value is expressed as follows:

[0078] ;

[0079] Where α is the weight of the noise reduction module, β is the hyperparameter for adjusting the slope, and θ is the threshold.

[0080] Before model training, the dataset can be augmented using data augmentation techniques (such as adding synthetic noise or simulating abnormal device conditions). During model training, the cross-entropy loss for speech error correction and the Laplacian regularization loss for graph structure are combined to ensure that the model maintains the semantic consistency of the graph structure while correcting speech errors. The loss function expression is as follows:

[0081] ;

[0082] ;

[0083] ;

[0084] in, For cross-entropy loss in speech error correction, For graph structure loss (such as smoothness loss of node embedding vectors). Here, N is the balance coefficient, and C is the number of samples and the number of classes. Let y be the predicted text and y be the real text. i, j ) is the edge, w ij For edge weights, h i , h j The node embedding vector.

[0085] In this embodiment of the invention, the AdamW optimizer is employed, combined with learning rate preheating and cosine annealing strategies, to prevent model overfitting and accelerate convergence. The AdamW optimizer is used to update θ, expressed as:

[0086] ;

[0087] Where η is the learning rate. , ϵ represents the first and second moment estimates, and ϵ is the smoothing term.

[0088] The cosine annealing strategy can be expressed as:

[0089] ;

[0090] η min η max For the minimum / maximum learning rate, T cur T max This represents the current / total number of training rounds.

[0091] Step S04: When the edge device receives new voice data and environmental data, it constructs the corresponding relationship graph and inputs it into the target model to generate the corrected text output.

[0092] The target model dynamically adjusts its error correction strategy based on graph features to determine the current environmental state. Specifically, when the target model determines that the current environmental state is a preset noise environment based on graph features, the noise reduction module is triggered. Specifically, a deep learning-based noise reduction algorithm (such as a convolutional neural network noise reduction model) is used to process the noisy speech features. The denoised speech features are then input into the target model to generate corrected text output. More specifically, noise is categorized into periodic noise (such as machine noise), random noise (such as workshop noise), and sudden noise (such as abnormal noises from equipment malfunctions). For different types, corresponding noise reduction filters are applied. For example, periodic noise can be addressed using an adaptive notch filter, with the following formula:

[0093] ;

[0094] r is the attenuation coefficient, and ω0 is the noise frequency;

[0095] Random noise can be denoised using deep learning models (such as DnCNN), which learn the mapping relationship between noise and clean speech through end-to-end training;

[0096] When the target model determines that the current environmental state is the preset operating state of the device based on graph features, for voice commands involving device operation, semantic verification of the commands is performed using device operating parameters (such as speed and load) in the environmental data. For example, the voice-recognized command is combined with the current device state to determine whether the command is reasonable. If the current device speed has reached its upper limit, and the voice command is "increase speed," the command is corrected or a warning is issued. In this embodiment of the invention, a safety domain model of the device operating state is constructed, mapping parameters such as speed and load to safety regions in a multi-dimensional space. For example, the device speed S and load L must satisfy a set of inequalities:

[0097] ;

[0098] When a voice command involves device operation, the command parameters are compared with the security domain. If they exceed the range, error correction is triggered.

[0099] The risk of instruction execution is assessed using a pre-defined rule base and machine learning model. It should be noted that a risk quantification model is constructed, calculating a risk value based on the overall severity of the instruction, the current state of the equipment, and the urgency of the operation, expressed as:

[0100] ;

[0101] ω1, ω2, ω3 are weighting coefficients;

[0102] Then, different handling methods are adopted according to the risk value, for example:

[0103] R≥0.8: Force correction command and issue warning (e.g., change "increase speed to limit" to "increase speed by 10%));

[0104] 0.5≤R<0.8: Provide multiple correction suggestions for operator confirmation;

[0105] R<0.5: Only record potential risks, do not interfere with instruction execution.

[0106] In summary, the proposed speech correction method based on edge computing power in this invention involves acquiring speech data and environmental data, aligning them to form speech data samples with environmental labels, constructing a relational graph based on a graph neural network, building a speech correction model that integrates graph information, and training the model with the relational graph to obtain a target model. This target model is then deployed to an edge device, utilizing edge computing power for real-time inference. When the edge device receives new speech and environmental data, it constructs the corresponding relational graph and inputs it into the target model to generate corrected text output. The target model dynamically adjusts its correction strategy based on graph features to determine the current environmental state. Specifically, through deep collaboration between speech and environmental data, the accuracy and robustness of edge-based speech correction are effectively enhanced.

[0107] Example 2

[0108] Embodiment 2 of the present invention provides a speech error correction system 200 based on edge computing power. Please refer to [link / reference]. Figure 2 Here is a structural block diagram of a speech error correction system based on edge computing power. The speech error correction system 200 based on edge computing power includes:

[0109] The acquisition module 21 is used to acquire voice data and environmental data, and to align the voice data and environmental data to form a voice data sample with environmental tags;

[0110] The construction module 22 is used to construct a relational graph based on a graph neural network based on speech data samples with environmental labels. The relational graph consists of nodes and edges. The nodes include speech nodes and environmental nodes. The acoustic features under the speech frame are used as node attributes. Each speech frame corresponds to one speech node.

[0111] Each environmental parameter is treated as a separate node, and the collection timestamp of the corresponding environmental parameter is assigned as an attribute.

[0112] Each speech node establishes an undirected edge with all environmental nodes at the same time, and the weights are dynamically adjusted according to the degree of influence of environmental factors on speech.

[0113] Between environmental nodes, edge connections are established based on preset environmental parameters, and edge weights are calculated using the Pearson correlation coefficient to characterize the correlation between parameters;

[0114] In addition, the relationship graph is encoded using graph structure encoding, and graph embedding algorithm is used to encode the graph structure into a low-dimensional vector representation, preserving the topological relationship between speech and environmental factors;

[0115] Training module 23 is used to construct a speech error correction model that integrates graph information. The relationship graph is input into the speech error correction model for training to obtain a target model. The target model is deployed to the edge device to realize real-time inference using edge computing power. The speech error correction model that integrates graph information includes a graph feature extraction layer and an attention fusion layer. The graph feature extraction layer uses GNN to extract features from the relationship graph and outputs high-order semantic features of each node. The attention fusion layer uses a multi-head attention mechanism to fuse the acoustic features of the speech frame with the environmental graph features of the corresponding node to generate a speech representation that includes environmental awareness. In addition, an adaptive weight function is designed to dynamically adjust the weight of the anti-noise module according to the feature values ​​of the environmental nodes.

[0116] Combining the cross-entropy loss for speech error correction with the Laplacian regularization loss for graph structure, the model ensures that it maintains the semantic consistency of the graph structure while correcting speech errors. The loss function expression is as follows:

[0117] ;

[0118] in, For cross-entropy loss in speech error correction, For graph structure loss, This is the balance coefficient;

[0119] The input module 24 is used to construct a corresponding relationship graph and input it into the target model when the edge device receives new voice data and environmental data, thereby generating corrected text output. The target model dynamically adjusts the error correction strategy based on the graph features to determine the current environmental state. Specifically, when the target model determines that the current environmental state is a preset noise environment based on the graph features, the noise reduction module is triggered. Specifically, a noise reduction algorithm based on deep learning is used to process the noisy voice features, and then the noise-reduced voice features are input into the target model to generate corrected text output.

[0120] When the target model determines that the current environmental state is the preset operating state of the device based on the graph features, for voice commands involving device operation, the device operating parameters in the environmental data are used to perform semantic verification on the commands.

[0121] The risk of instruction execution is assessed through a pre-defined rule base and machine learning model.

[0122] Furthermore, in some other embodiments of the present invention, the acquisition module 21 includes:

[0123] The transformation unit is used to acquire speech data, employs beamforming technology to capture human voice in a directional manner, and uses short-time Fourier transform to convert the time-domain speech signal into a frequency-domain Mel spectrogram.

[0124] The extraction unit is used to extract acoustic features based on the frequency domain Mel spectrum.

[0125] The acquisition unit is used to acquire environmental data including at least temperature, humidity, ambient noise decibels, equipment speed, and load, and to normalize the environmental data.

[0126] The matching unit is used to match the acoustic features of the speech frame with the normalized environmental data using a timestamp alignment algorithm to form a speech data sample with environmental labels.

[0127] Example 3

[0128] Embodiment 3 of the present invention proposes an electronic device, please refer to [link / reference]. Figure 3 This is a structural block diagram of an electronic device, including a memory 20, a processor 10, and a computer program 30 stored in the memory and executable on the processor. When the processor 10 executes the computer program 30, it implements the speech error correction method based on edge computing power as described above.

[0129] In some embodiments, the processor 10 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip, used to run program code stored in memory 20 or process data, such as executing access restriction programs.

[0130] The memory 20 includes at least one type of readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 20 can be an internal storage unit of an electronic device, such as the hard disk of the electronic device. In other embodiments, the memory 20 can also be an external storage device of the electronic device, such as a plug-in hard disk, SmartMediaCard (SMC), SecureDigital (SD) card, FlashCard, etc., equipped on the electronic device. Furthermore, the memory 20 can include both internal and external storage units of the electronic device. The memory 20 can be used not only to store application software and various types of data of the electronic device, but also to temporarily store data that has been output or will be output.

[0131] This invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned speech error correction method based on edge computing power.

[0132] Those skilled in the art will understand that the logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0133] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0134] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0135] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0136] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A speech error correction method based on edge computing power, characterized in that, Applied to industrial workshop scenarios where edge devices are installed, the method includes: Acquire voice data and environmental data, and align the voice data and environmental data to form a voice data sample with environmental labels; Based on voice data samples with environmental labels, a relational graph based on a graph neural network is constructed. A speech error correction model integrating graph information is constructed, and the relationship graph is input into the speech error correction model for training to obtain a target model. The target model is then deployed to the edge device to achieve real-time inference using edge computing power. When the edge device receives new voice data and environmental data, it constructs a corresponding relationship graph and inputs it into the target model to generate corrected text output. The target model dynamically adjusts the error correction strategy based on the graph features to determine the current environmental state.

2. The speech error correction method based on edge computing power according to claim 1, characterized in that, The step of acquiring voice data and environmental data, and aligning the voice data and environmental data to form a voice data sample with environmental labels includes: Acquire speech data, use beamforming technology to capture human voice in a directional manner, and use short-time Fourier transform to convert the time-domain speech signal into a frequency-domain Mel spectrogram; Based on the frequency domain Mel spectrum, acoustic features are extracted; Acquire environmental data including at least temperature, humidity, ambient noise decibels, equipment speed, and load, and normalize the environmental data. A timestamp alignment algorithm is used to match the acoustic features of the speech frame with the normalized environmental data to form a speech data sample with environmental labels.

3. The speech error correction method based on edge computing power according to claim 2, characterized in that, In the step of constructing a relational graph based on a graph neural network according to speech data samples with environmental labels, the relational graph consists of nodes and edges. The nodes include speech nodes and environmental nodes. The acoustic features under the speech frame are used as node attributes, and each speech frame corresponds to one speech node. Each environmental parameter is treated as a separate node, and the collection timestamp of the corresponding environmental parameter is assigned as an attribute. Each speech node establishes an undirected edge with all environmental nodes at the same time, and the weights are dynamically adjusted according to the degree of influence of environmental factors on speech. Between environmental nodes, edge connections are established based on preset environmental parameters, and edge weights are calculated using the Pearson correlation coefficient to characterize the correlation between parameters; In addition, the relationship graph is encoded using graph structure encoding, and a graph embedding algorithm is used to encode the graph structure into a low-dimensional vector representation, preserving the topological relationship between speech and environmental factors.

4. The speech error correction method based on edge computing power according to claim 3, characterized in that, The speech error correction model that integrates graph information includes a graph feature extraction layer and an attention fusion layer. The graph feature extraction layer uses a GNN to extract features from the relation graph and outputs high-order semantic features of each node. The attention fusion layer uses a multi-head attention mechanism to fuse the acoustic features of the speech frame with the environmental graph features of the corresponding node to generate a speech representation that includes environmental awareness. In addition, an adaptive weight function is designed to dynamically adjust the weight of the noise reduction module according to the feature values ​​of the environmental nodes.

5. The speech error correction method based on edge computing power according to claim 4, characterized in that, In the step of inputting the relationship graph into the speech correction model for training to obtain the target model, the cross-entropy loss of speech correction and the Laplacian regularization loss of the graph structure are combined to ensure that the model maintains the semantic consistency of the graph structure while correcting speech errors. The loss function expression is as follows: ； in, For cross-entropy loss in speech error correction, For graph structure loss, This is the balance coefficient.

6. The speech error correction method based on edge computing power according to claim 5, characterized in that, The steps of dynamically adjusting the error correction strategy based on graph features to determine the current environmental state include: When the target model determines that the current environment is a preset noise environment based on graph features, the noise reduction module will be triggered. Specifically, a noise reduction algorithm based on deep learning is used to process the noisy speech features, and then the noise-reduced speech features are input into the target model to generate corrected text output. When the target model determines that the current environmental state is the preset operating state of the device based on the graph features, for voice commands involving device operation, the device operating parameters in the environmental data are used to perform semantic verification on the commands. The risk of instruction execution is assessed through a pre-defined rule base and machine learning model.

7. A speech error correction system based on edge computing power, characterized in that, The system for implementing the speech error correction method based on edge computing power as described in any one of claims 1-6 includes: The acquisition module is used to acquire voice data and environmental data, and to align the voice data and environmental data to form a voice data sample with environmental labels; The module is used to construct a relational graph based on a graph neural network from voice data samples with environmental labels. The training module is used to construct a speech error correction model that integrates graph information. The relationship graph is input into the speech error correction model for training to obtain a target model. The target model is deployed to the edge device to realize real-time inference using edge computing power. The input module is used to construct a corresponding relationship graph and input it into the target model when the edge device receives new voice data and environmental data, and generate corrected text output. The target model dynamically adjusts the error correction strategy based on the graph features to determine the current environmental state.

8. A computer-readable storage medium, characterized in that, include: The readable storage medium stores one or more programs that, when executed by a processor, implement the speech error correction method based on edge computing power as described in any one of claims 1-6.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, wherein: The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the speech error correction method based on edge computing power as described in any one of claims 1-6.

Images