A precise voice interaction system for complex scenarios

Abstract

The application relates to the technical field of voice interaction, and discloses a precise voice interaction system for complex scenes, which comprises the following modules: an acoustic feature extraction module, which extracts an acoustic feature vector from a voice signal and generates an acoustic feature representation for suppressing dialect differences; an adaptive semantic understanding module, which generates an initial semantic vector based on the acoustic feature representation and performs semantic enhancement, combines a dynamically adjusted semantic weighting fusion weight to perform weighted fusion on a historical dialogue semantic vector, and generates a user intention representation; a semantic decoding module, which decodes the user intention representation to generate a precise intention instruction; and a context management module, which generates a dialogue state vector according to the user intention representation, calculates the correlation degree between the dialogue state vector and the historical dialogue semantic vector, generates a dialogue context representation, and dynamically adjusts the semantic weighting fusion weight. The application can improve voice recognition accuracy, guarantee context coherence, and better adapt to complex and changeable real scene requirements.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of voice interaction, and particularly relates to a precise voice interaction system for complex scenes. BACKGROUND

[0002] With the development of science and technology, voice interaction systems are widely used in many fields. However, in complex scenes, the existing voice interaction systems have many problems. For example, in a noisy environment, the voice recognition accuracy is greatly reduced, background noise cannot be effectively filtered, and command recognition errors occur. When facing different accents, dialects and professional terms, the system has poor adaptability and cannot accurately understand user intentions. In the process of multi-round dialogue, the context coherence is insufficient, and semantic understanding deviation easily occurs. Moreover, in specific industry applications, there is a lack of deep integration of professional knowledge, which cannot meet the complex business requirements, greatly limiting the application range and user experience of the voice interaction system, and making it difficult to adapt to complex and variable real scene requirements.

[0003] Therefore, there is an urgent need to provide a technical solution to solve the above problems. SUMMARY

[0004] To solve the above technical problems, the present application provides a precise voice interaction system for complex scenes.

[0005] In a first aspect, the present application provides a precise voice interaction system for complex scenes, and the technical scheme of the system is as follows:

[0006] An acoustic feature extraction module is configured to extract an acoustic feature vector from a target voice signal and generate an acoustic feature representation that suppresses dialect differences through an adversarial training network containing a dialect classifier;

[0007] An adaptive semantic understanding module is configured to generate an initial semantic vector based on the acoustic feature representation, perform semantic enhancement on the initial semantic vector by matching a professional term semantic association matrix, and perform weighted fusion on historical dialogue semantic vectors by combining a dynamically adjusted semantic weighted fusion weight, to generate a user intention representation;

[0008] A semantic decoding module is configured to perform mapping decoding on the user intention representation based on a preset instruction template, to generate a precise intention instruction corresponding to the semantic content of the target voice signal;

[0009] A context management module is configured to generate a dialogue state vector according to the user intention representation, calculate the correlation degree between the dialogue state vector and the historical dialogue semantic vector through an attention mechanism, generate a dynamically updated dialogue context representation, and feed the dialogue context representation back to the adaptive semantic understanding module to dynamically adjust the semantic weighted fusion weight.

[0010] The beneficial effects of the accurate voice interaction system for complex scenarios proposed in this invention are as follows:

[0011] The system of this invention can solve many problems of existing voice interaction systems in noisy environments, dialects, professional terminology, and multi-turn dialogue scenarios. It improves the accuracy of voice recognition, enhances the system's adaptability, ensures contextual coherence, and deeply integrates professional domain knowledge, thereby expanding the scope of application, optimizing the user experience, and thus better adapting to the requirements of complex and ever-changing real-world scenarios.

[0012] Based on the above solution, the precise voice interaction system for complex scenarios of the present invention can be further improved as follows.

[0013] In one alternative approach, the acoustic feature extraction module is specifically used for:

[0014] The target speech signal is input into a deep neural network to extract the acoustic feature vector containing speech content and dialect characteristics;

[0015] Construct the adversarial training network comprising a generator and a discriminator; the generator is composed of the deep neural network and is used to generate acoustic feature representations that suppress dialect differences; the discriminator is a dialect classifier used to distinguish feature representations of different dialect categories.

[0016] Based on the optimization objective, the parameters of the generator and the discriminator are optimized through adversarial training, and the acoustic feature vector is input into the optimized generator to obtain the acoustic feature representation.

[0017] In one alternative approach, the adaptive semantic understanding module is specifically used for:

[0018] The acoustic feature representation is input into the semantic encoder to generate the initial semantic vector, the expression of which is:

[0019] v0 = Encoder(H inv )

[0020] In the formula, v0 is the initial semantic vector, H inv For the acoustic feature representation, Encoder(·) is the encoder function;

[0021] By calculating the similarity between the initial semantic vector and the semantic association matrix of the technical terms, matching term vectors are selected for semantic enhancement to obtain the enhanced semantic vector, the expression of which is:

[0022]

[0023] In the formula, α irepresents the semantic matching degree of the initial semantic vector and the i-th professional term in the professional term semantic association matrix, v represents the enhanced semantic vector, represents the i-th professional term vector in the professional term semantic association matrix, N represents the number of professional term vectors in the professional term semantic association matrix; represents the transpose matrix of the initial semantic vector; represents the j-th professional term vector in the professional term semantic association matrix;

[0024] According to the dynamically adjusted semantic weighted fusion weight, the enhanced semantic vector and the historical dialogue semantic vector are weighted and fused to generate the user intent representation, and the expression is:

[0025]

[0026] In the formula, β t represents the t-th element of the dynamically adjusted semantic weighted fusion weight β, represents the key vector of the t-th element in the historical dialogue semantic vector, Q current represents the query vector of the enhanced semantic vector, represents the t-th element of the historical dialogue semantic vector, d represents the dimension of the semantic vector, v intent represents the user intent representation.

[0027] In an optional manner, the semantic decoding module is specifically configured to:

[0028] The preset instruction template is constructed to contain a plurality of instruction template vectors, and each instruction template vector corresponds to an executable operation;

[0029] The user intent representation is calculated to be similar to each instruction template vector respectively, and the confidence score of each instruction template vector is calculated, and the expression is:

[0030]

[0031] In the formula, s j represents the confidence score of the j-th instruction template vector, T represents the preset instruction template, T (j) represents the j-th instruction template vector, T (m) represents the m-th instruction template vector, M represents the number of instruction template vectors;

[0032] Based on the instruction template vector with the highest confidence score, and combined with the semantic details of the user intent representation, the accurate intent instruction is generated, and the expression is:

[0033]

[0034] In the formula, TemplateFill represents a template filling function, used to inject semantic details of v intent into placeholders of the instruction template vector, and CMD represents the precise intention instruction.

[0035] In an optional manner, the context management module is specifically used for:

[0036] characterizing the user intention v intent inputting the state encoder to generate the dialogue state vector;

[0037] calculating the attention correlation degree of the dialogue state vector and each round vector in the historical dialogue semantic vector, which is expressed as:

[0038]

[0039] In the formula, γ t represents the attention correlation degree of the t-th round vector in the historical dialogue semantic vector, represents the transpose matrix of the dialogue state vector V current ;

[0040] weighting and summing the historical dialogue state vectors according to the correlation degree weight to generate the dynamically updated dialogue context representation, which is expressed as:

[0041]

[0042] In the formula, c ctx represents the dialogue context representation;

[0043] feeding back the dialogue context representation to the adaptive semantic understanding module to dynamically adjust the semantic weighted fusion weight, which is expressed as:

[0044] β=Sigmoid(W β ·c ctx +b β )

[0045] In the formula, W β represents a weight mapping matrix, and b β represents a bias vector.

[0046] In an optional manner, the method further comprises:

[0047] The preprocessing module is used for receiving an original speech signal and performing collaborative processing on the original speech signal through a noise suppression model and a speech enhancement model to generate the target speech signal.

[0048] In an optional mode, the professional term semantic association matrix is constructed by a professional term atlas, and a multi-level semantic relationship between terms is mined by a graph neural network.

[0049] In an optional mode, the optimization target is that the generator minimizes a speech content classification loss while maximizing a dialect classification error of the discriminator, and the discriminator minimizes a dialect classification loss.

[0050] In a second aspect, the application provides a precise voice interaction method for a complex scene, and the technical solution of the method is as follows.

[0051] An acoustic feature vector is extracted from a target voice signal, and an acoustic feature representation suppressing dialect differences is generated by an adversarial training network containing a dialect classifier;

[0052] An initial semantic vector is generated based on the acoustic feature representation, the initial semantic vector is semantically enhanced by matching a professional term semantic association matrix, and a historical dialogue semantic vector is weighted and fused by combining a dynamically adjusted semantic weighting fusion weight, to generate a user intention representation;

[0053] The user intention representation is mapped and decoded based on a preset instruction template, to generate a precise intention instruction corresponding to the semantic content of the target voice signal;

[0054] A dialogue state vector is generated according to the user intention representation, an association degree between the dialogue state vector and the historical dialogue semantic vector is calculated by an attention mechanism, a dynamically updated dialogue context representation is generated, and the semantic weighting fusion weight is dynamically adjusted.

[0055] The precise voice interaction method for a complex scene of the application has the following beneficial effects:

[0056] The method of the application can solve many problems of existing voice interaction systems in noisy environments, accents, dialects, professional terms and multi-round dialogue scenes, improve voice recognition accuracy, enhance system adaptability, ensure context coherence, deeply integrate professional knowledge, thereby expand the application range, optimize user experience, and better adapt to complex and variable real scene requirements.

[0057] The above description is only a summary of the technical solutions of the application, in order to more clearly understand the technical means of the application, the application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the application more obvious and easy to understand, the following specific embodiments of the application are described. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to make the technical solutions of the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description only show some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative effort based on these drawings.

[0059] The accompanying drawings are only used to show the embodiments, and are not considered as limiting the present application. Moreover, the same reference signs are used to represent the same components throughout the drawings. In the drawings:

[0060] Figure 1 A structural schematic diagram of an embodiment of a complex scene-oriented accurate voice interaction system of the present application;

[0061] Figure 2 A flowchart of an embodiment of a complex scene-oriented accurate voice interaction method of the present application. DETAILED DESCRIPTION

[0062] Exemplary embodiments of the present application will be described in detail with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein.

[0063] Figure 1 A structural schematic diagram of an embodiment of a complex scene-oriented accurate voice interaction system provided by the present application is shown. As shown in the figure, Figure 1 The system comprises:

[0064] An acoustic feature extraction module 110 is configured to extract an acoustic feature vector from a target voice signal and generate an acoustic feature representation suppressing dialect difference through an adversarial training network comprising a dialect classifier.

[0065] The target voice signal refers to the original voice waveform data input by a user through a microphone or other audio acquisition device, or the voice waveform data after noise reduction processing. The acoustic feature vector refers to a high-dimensional representation extracted from the target voice signal, including but not limited to voice content and dialect characteristics. The adversarial training network refers to a joint training architecture comprising a generator network (feature extractor) and a discriminator network (dialect classifier), which suppresses the dialect difference in the acoustic feature through an adversarial game. The acoustic feature representation refers to the dialect invariance feature output after processing by the adversarial training network, which has the same dimension as the acoustic feature vector but suppresses the dialect correlation information.

[0066] The adaptive semantic understanding module 120 is configured to generate an initial semantic vector based on the acoustic feature representation, perform semantic enhancement on the initial semantic vector by matching a professional term semantic correlation matrix, and perform weighted fusion on a historical dialogue semantic vector by combining a dynamically adjusted semantic weighted fusion weight, to generate a user intent representation.

[0067] The initial semantic vector is a d-dimensional vector generated by a semantic encoder performing nonlinear transformation on the acoustic feature representation, and initially represents a semantic intent in the voice signal. The professional term semantic correlation matrix is constructed from a professional term graph and mined by a graph neural network to obtain a multi-level semantic relationship between terms. The row vector of the professional term semantic correlation matrix represents a d-dimensional semantic vector of a professional term, and the matrix element reflects the co-occurrence probability or semantic similarity between terms. The dynamically adjusted semantic weighted fusion weight represents the contribution proportion of the historical dialogue semantic vector to the current intent understanding, and the weight value is calculated in real time by an attention mechanism. The historical dialogue semantic vector is a sequence of semantic vectors stored in the previous k rounds of dialogue, and each round of dialogue corresponds to a d-dimensional vector. The user intent representation is a d-dimensional vector after fusing the field enhanced semantic and historical context.

[0068] The semantic decoding module 130 is configured to perform mapping decoding on the user intent representation based on a preset instruction template to generate a precise intent instruction corresponding to the semantic content of the target voice signal.

[0069] The preset instruction template is a template including a plurality of instruction template vectors, and each instruction template vector includes fixed text and a parameter placeholder. The precise intent instruction is a structured instruction output by the semantic decoding module 130, including an operation type, a target device, and a parameter value. The precise intent instruction is used to drive a downstream device to execute a user instruction, and ensures the accuracy of the operation in a complex scenario. For example, inputting "turn off all lights after the meeting ends", the precise intent instruction triggers a conference room light-off sequence.

[0070] The context management module 140 is configured to generate a dialogue state vector according to the user intent representation, calculate the correlation degree between the dialogue state vector and the historical dialogue semantic vector by an attention mechanism, generate a dynamically updated dialogue context representation, and feed the dialogue context representation back to the adaptive semantic understanding module to dynamically adjust the semantic weighted fusion weight.

[0071] The dialogue state vector is a low-dimensional vector obtained by compressively encoding the user intent representation, and represents the context state of the current dialogue. The dialogue state vector is used to calculate the correlation degree with the historical dialogue to generate a dynamic context representation. The dialogue context representation is a d-dimensional vector obtained by weighted fusion of the historical dialogue state vectors, and dynamically reflects the semantic context of multiple rounds of dialogue.

[0072] In an alternative way, the acoustic feature extraction module 110 is specifically configured to:

[0073] inputting the target speech signal into a deep neural network to extract the acoustic feature vector containing speech content and dialect characteristics.

[0074] wherein the target speech signal is represented as T represents the time length of the target speech signal, and F represents the frequency points per frame. The speech content includes phonemes and semantics, and the dialect characteristics include pronunciation habits and tone patterns. The acoustic feature vector is represented as

[0075] constructing the adversarial training network containing a generator and a discriminator; the generator is composed of the deep neural network and is configured to generate an acoustic feature representation suppressing dialect differences; and the discriminator is a dialect classifier configured to distinguish feature representations of different dialect categories.

[0076] wherein the generator includes 4 layers of residual CNN, and the output dimension d = 256; and the discriminator includes 2 layers of fully connected network, and the output dimension is the number of dialect categories.

[0077] based on an optimization objective, optimizing parameters of the generator and the discriminator through adversarial training, and inputting the acoustic feature vector into the optimized generator to obtain the acoustic feature representation.

[0078] wherein the optimization objective is that the generator minimizes the speech content classification loss while maximizing the dialect classification error of the discriminator, and the discriminator minimizes the dialect classification loss.

[0079] Specifically, the generator loss function is: is a mathematical expectation operator, representing a statistical average of the input speech signal X; C(·; θ C ) represents a speech content classifier, and the parameter is θ C ; logC(H; θ C ) represents the log-likelihood probability of speech content classification; 1-logD(H; θ D ) represents the log error of the dialect classifier in discriminating the generated features; and λ represents the weight coefficient of the adversarial loss. The discriminator loss function is: θ D represents the trainable parameters of the discriminator network.

[0080] In an alternative way, the adaptive semantic understanding module 120 is specifically configured to:

[0081] inputting the acoustic feature representation into a semantic encoder to generate the initial semantic vector, and the expression is:

[0082] v0=Encoder(H inv )

[0083] wherein v0 is the initial semantic vector, H inv is the acoustic feature representation, and Encoder(·) is an encoder function.

[0084] By calculating the similarity between the initial semantic vector and the professional term semantic association matrix, the matching term vector is screened for semantic enhancement, and an enhanced semantic vector is obtained, and its expression is:

[0085]

[0086] wherein α i represents the semantic matching degree of the initial semantic vector and the i-th professional term in the professional term semantic association matrix, and v1 represents the enhanced semantic vector, represents the i-th professional term vector in the professional term semantic association matrix, and N represents the number of professional term vectors in the professional term semantic association matrix. represents the transpose matrix of the initial semantic vector. represents the j-th professional term vector in the professional term semantic association matrix.

[0087] According to the dynamically adjusted semantic weighted fusion weight, the enhanced semantic vector and the historical dialogue semantic vector are weighted and fused to generate the user intent representation, and its expression is:

[0088]

[0089] wherein β t represents the t-th element of the dynamically adjusted semantic weighted fusion weight β, represents the key vector of the t-th element in the historical dialogue semantic vector, Q current represents the query vector of the enhanced semantic vector, represents the t-th element of the historical dialogue semantic vector, and d represents the dimension of the semantic vector, v intent represents the user intent representation.

[0090] In an optional manner, the semantic decoding module 130 is specifically configured to:

[0091] construct the preset instruction template including a plurality of instruction template vectors, each instruction template vector corresponding to an executable operation;

[0092] The user intention representation is matched with each instruction template vector respectively to calculate a confidence score of each instruction template vector, which is expressed as:

[0093]

[0094] wherein s j represents the confidence score of the jth instruction template vector, T represents the preset instruction template, and T (j) represents the jth instruction template vector, T (m) represents the mth instruction template vector, and M represents the number of instruction template vectors.

[0095] The accurate intention instruction is generated based on the instruction template vector with the highest confidence score and in combination with the semantic details of the user intention representation, which is expressed as:

[0096]

[0097] wherein TemplateFill represents a template filling function for injecting the semantic details of v intent into placeholders of the instruction template vector, and CMD represents the accurate intention instruction.

[0098] In an optional manner, the context management module 140 is specifically configured to:

[0099] input the user intention representation v intent into a state encoder to generate the dialogue state vector;

[0100] calculate an attention correlation degree of the dialogue state vector and each round vector in the historical dialogue semantic vector, which is expressed as:

[0101]

[0102] wherein γ t represents the attention correlation degree of the tth round vector in the historical dialogue semantic vector, represents the transpose matrix of the dialogue state vector V current ;

[0103] weight-sum the historical dialogue state vectors according to the correlation degree weights to generate the dynamically updated dialogue context representation, which is expressed as:

[0104]

[0105] wherein c ctx represents the dialogue context representation.

[0106] The dialogue context representation is fed back to the adaptive semantic understanding module to dynamically adjust the semantic weighted fusion weight, which is expressed as:

[0107] β = Sigmoid(W β · c ctx + b β )

[0108] wherein W β represents a weight mapping matrix, and b β represents a bias vector.

[0109] In an optional manner, the technical scheme further comprises:

[0110] The preprocessing module is configured to receive an original speech signal and perform collaborative processing on the original speech signal by a noise suppression model and a speech enhancement model to generate the target speech signal.

[0111] Specifically, a professional speech acquisition device is used to receive an original speech signal input by a user; a signal processing algorithm (such as a short-time energy calculation method, a spectrum analysis method, etc.) is used to extract representative features from the received original speech signal. The noise suppression model analyzes and judges the input speech features based on the knowledge learned in its training process (for example, a neural network model will analyze and judge the input speech features according to the internal weights, biases, etc.). The original speech signal is input into a pre-trained speech enhancement model, and the model optimizes the speech features according to the internal structure and parameters. The de-noised speech features obtained after the noise suppression model processing and the optimized speech features obtained after the speech enhancement model processing are combined according to a certain fusion strategy. It can be a simple weighted summation method, which sets different weights according to the importance of the processing results of the two models, or a more complex nonlinear fusion method, which fully combines the characteristics of the two models to form the optimized speech features. The corresponding signal synthesis technology (such as an inverse short-time Fourier transform method) is used to convert the optimized speech features back to a time-domain speech signal form, i.e., a target speech signal. In this process, it is necessary to ensure that the synthesized speech signal is natural and coherent in waveform without obvious distortion or abrupt connection points, and finally the target speech signal is generated, which has lower noise interference, higher speech quality, and better intelligibility and clarity compared with the original speech signal.

[0112] The technical scheme of the embodiment can solve many problems of the existing voice interaction system in noisy environments, accents, dialects, professional terms, and multi-round dialogue scenarios, improve the voice recognition rate, enhance the system adaptability, ensure the context coherence, deeply integrate professional knowledge, thereby expand the application range, optimize the user experience, and thus better adapt to complex and variable real-world scenarios.

[0113] Figure 2 A flowchart of an embodiment of a complex scene-oriented accurate voice interaction method provided by the present application is shown. As shown, it includes the following steps: Figure 2

[0114] S1, extracting an acoustic feature vector from a target voice signal, and generating an acoustic feature representation suppressing dialect differences through an adversarial training network containing a dialect classifier;

[0115] S2, generating an initial semantic vector based on the acoustic feature representation, performing semantic enhancement on the initial semantic vector through a professional term semantic association matrix, and performing weighted fusion on a historical dialogue semantic vector in combination with a dynamically adjusted semantic weighted fusion weight, to generate a user intent representation;

[0116] S3, performing mapping decoding on the user intent representation based on a preset instruction template, to generate an accurate intent instruction corresponding to the semantic content of the target voice signal;

[0117] S4, generating a dialogue state vector according to the user intent representation, calculating the correlation degree between the dialogue state vector and the historical dialogue semantic vector through an attention mechanism, generating a dynamically updated dialogue context representation, and dynamically adjusting the semantic weighted fusion weight.

[0118] The technical solution of the embodiment can solve many problems of existing voice interaction systems in noisy environments, accents, dialects, professional terms, and multi-round dialogue scenes, improve voice recognition accuracy, enhance system adaptability, ensure context coherence, deeply integrate professional knowledge, thereby expanding the application range, optimizing user experience, and thus better adapting to complex and variable real-world scene requirements.

[0119] In addition, the system provided in the above embodiment is only exemplified by the division of the above functional modules when implementing its functions. In actual application, the above functions can be completed by different functional modules according to needs, that is, the system is divided into different functional modules according to actual conditions to complete all or part of the above described functions. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process is described in the method embodiment, which is not repeated here.

[0120] ​The above description is only the preferred embodiment of the present application and the explanation of the applied technical principles. It should be understood by those skilled in the art that the disclosed range of the present application is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the disclosed concept. For example, the technical solutions formed by the mutual replacement of the above features and the technical features disclosed in the present application (but not limited to) with similar functions.

[0121] It should be noted that the terms "first", "second" and the like in the specification and claims of the present application are used to distinguish similar objects, and represent the limitation of a specific order or sequence. The order of use of similar objects can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described.

[0122] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A precise voice interaction system for complex scenarios, characterized in that, The system includes: The acoustic feature extraction module is used to extract acoustic feature vectors from the target speech signal and generate acoustic feature representations that suppress dialect differences through an adversarial training network containing a dialect classifier. An adaptive semantic understanding module is used to generate an initial semantic vector based on the acoustic feature representation, enhance the semantics of the initial semantic vector by matching the semantic association matrix of professional terms, and perform weighted fusion of historical dialogue semantic vectors by combining dynamically adjusted semantic weighted fusion weights to generate a user intent representation. The semantic decoding module is used to perform mapping decoding on the user intent representation based on a preset instruction template to generate an accurate intent instruction corresponding to the semantic content of the target speech signal; The context management module is used to generate a dialogue state vector based on the user intent representation, calculate the correlation between the dialogue state vector and the historical dialogue semantic vector through an attention mechanism, generate a dynamically updated dialogue context representation, and feed the dialogue context representation back to the adaptive semantic understanding module to dynamically adjust the semantic weighted fusion weights. The adaptive semantic understanding module is specifically used for: The acoustic feature representation is input into the semantic encoder to generate the initial semantic vector, the expression of which is: In the formula, The initial semantic vector. This represents the acoustic features. The encoder function calculates the similarity between the initial semantic vector and the semantic association matrix of the technical terms, filters out matching term vectors for semantic enhancement, and obtains the enhanced semantic vector, the expression of which is: In the formula, , This represents the first semantic vector in the semantic association matrix of the terminology. The semantic matching degree of each technical term. This represents the enhanced semantic vector. The term in the semantic association matrix represents the first term. A vector of technical terms, This represents the number of terminology vectors in the semantic association matrix of the terminology; The transpose matrix representing the initial semantic vector; The term in the semantic association matrix represents the first term. A vector of technical terms; based on the dynamically adjusted semantic weighting fusion weights, the enhanced semantic vector and the historical dialogue semantic vector are weighted and fused to generate the user intent representation, the expression of which is: In the formula, , , This represents the dynamically adjusted semantic weighted fusion weights. The One element, The first element in the semantic vector of the historical dialogue represents the... A key vector of n elements, The query vector representing the enhanced semantic vector. The first element representing the semantic vector of the historical dialogue One element, The dimension of the semantic vector. This represents the user's intent. The semantic decoding module is specifically used for: Construct the preset instruction template containing multiple instruction template vectors, each instruction template vector corresponding to an executable operation; The similarity of the user intent representation with each instruction template vector is calculated, and the confidence score of each instruction template vector is calculated, expressed as follows: , Indicates the first Confidence score of each instruction template vector. This refers to the preset instruction template. Indicates the first One instruction template vector, Indicates the first One instruction template vector, Indicates the number of instruction template vectors; Based on the instruction template vector with the highest confidence score, and combined with the semantic details of the user intent representation, the precise intent instruction is generated, and its expression is: In the formula, , This represents a template filling function, used to fill... Placeholders for the semantic detail injection instruction template vector. This indicates the precise intent instruction; The context management module is specifically used for: representing the user intent. The input state encoder generates the dialogue state vector; The attention correlation between the spoken state vector and each round vector in the historical dialogue semantic vector is calculated, and its expression is as follows: ; This represents the attention relevance of the vector in the t-th round of the historical dialogue semantic vector. Represents the dialogue state vector The transpose of the matrix; The historical dialogue state vectors are weighted and summed according to their relevance weights to generate a dynamically updated dialogue context representation, the expression of which is: ; The dialogue context representation is represented; the dialogue context representation is fed back to the adaptive semantic understanding module to dynamically adjust the semantic weighted fusion weights, the expression of which is: In the formula, Represents the weight mapping matrix. This represents the bias vector.

2. The precise voice interaction system for complex scenarios according to claim 1, characterized in that, The acoustic feature extraction module is specifically used for: The target speech signal is input into a deep neural network to extract the acoustic feature vector containing speech content and dialect characteristics; Construct the adversarial training network comprising a generator and a discriminator; the generator is composed of the deep neural network and is used to generate acoustic feature representations that suppress dialect differences; the discriminator is a dialect classifier used to distinguish feature representations of different dialect categories. Based on the optimization objective, the parameters of the generator and the discriminator are optimized through adversarial training, and the acoustic feature vector is input into the optimized generator to obtain the acoustic feature representation.

3. The precise voice interaction system for complex scenarios according to claim 1, characterized in that, Also includes: The preprocessing module is used to receive the original speech signal and perform collaborative processing on the original speech signal through a noise suppression model and a speech enhancement model to generate the target speech signal.

4. The precise voice interaction system for complex scenarios according to claim 1, characterized in that, The semantic association matrix of the terminology is constructed from a terminology graph and the multi-level semantic relationships between terms are mined through a graph neural network.

5. The precise voice interaction system for complex scenarios according to claim 2, characterized in that, The optimization objective is that the generator minimizes the speech content classification loss, while maximizing the dialect classification error of the discriminator, and the discriminator minimizes the dialect classification loss.

6. A precise voice interaction method for complex scenarios, employing the precise voice interaction system for complex scenarios as described in any one of claims 1 to 5, characterized in that, The method includes: Acoustic feature vectors are extracted from the target speech signal, and an adversarial training network containing a dialect classifier is used to generate acoustic feature representations that suppress dialect differences. An initial semantic vector is generated based on the acoustic feature representation. The initial semantic vector is then semantically enhanced by matching the semantic association matrix of professional terms. Finally, the historical dialogue semantic vector is weighted and fused with dynamically adjusted semantic weighting fusion weights to generate a user intent representation. The user intent representation is mapped and decoded based on a preset instruction template to generate an accurate intent instruction corresponding to the semantic content of the target speech signal; A dialogue state vector is generated based on the user intent representation. The correlation between the dialogue state vector and the historical dialogue semantic vector is calculated through an attention mechanism to generate a dynamically updated dialogue context representation, thereby dynamically adjusting the semantic weighted fusion weights.

7. The accurate voice interaction method for complex scenarios according to claim 6, characterized in that, The step of extracting acoustic feature vectors from the target speech signal and generating acoustic feature representations that suppress dialect differences through an adversarial training network including a dialect classifier further includes: The target speech signal is input into a deep neural network to extract the acoustic feature vector containing speech content and dialect characteristics; Construct the adversarial training network comprising a generator and a discriminator; the generator is composed of the deep neural network and is used to generate acoustic feature representations that suppress dialect differences; the discriminator is a dialect classifier used to distinguish feature representations of different dialect categories. Based on the optimization objective, the parameters of the generator and the discriminator are optimized through adversarial training, and the acoustic feature vector is input into the optimized generator to obtain the acoustic feature representation.

Images