Intelligent voice service management system and method driven by natural language understanding

Through the intelligent voice service management system driven by natural language understanding, combined with the integration results of data preprocessing and multiple large language models, the problems of low accuracy of speech recognition and poor interpretation ability in the existing technology are solved, and the ability to interpret complex psychological states is achieved. Efficient voice recognition and personalized drug rehabilitation support.

CN118194875BActive Publication Date: 2024-11-26HANGZHOU HUA TING TECH CO LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202410415917.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-08
Publication Date
2024-11-26
Estimated Expiration
2044-04-08

AI Technical Summary

Technical Problem

The existing voice service management system is not adaptable to environmental factors such as accent and noise, resulting in low accuracy in speech recognition and poor ability to interpret complex psychological states.

Method used

Adopting a natural language understanding-driven intelligent voice service management system, through the voice reception module, semantic understanding module, dialogue management module and speech synthesis module, combining data preprocessing, time-frequency feature extraction, iterative scaling noise prediction, intention recognition and slot filling technology to improve the accuracy of speech recognition and enhance the ability to interpret complex psychological states.

Benefits of technology

It improves the accuracy of speech recognition and the ability to interpret complex psychological states, provides personalized task guidance and feedback, and enhances the efficiency and quality of drug rehabilitation work in prisons.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118194875B_ABST
    Figure CN118194875B_ABST
Patent Text Reader

Abstract

The present application discloses an intelligent voice service management system and method driven by natural language understanding. The system includes: a voice receiving module for receiving the needs of drug addicts through voice; a semantic understanding module for converting the needs into text information to extract the semantic information of the text information; a dialogue management module for determining the next action according to the semantic information to obtain a reply text; a speech synthesis module for converting the reply text into speech to feed back to the drug addicts; and a dialogue log database for recording each round of dialogue for supervision and subsequent analysis. The present application scheme can improve recognition accuracy and enhance the ability to interpret psychological states.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of management systems, and in particular to a natural language understanding-driven voice service management system and method. Background Art

[0002] The voice service management system is a system that uses modern information technology, especially voice recognition, processing and feedback technology, to assist prison drug rehabilitation work. This system can improve the efficiency and quality of prison drug rehabilitation work, for example, by automatically recording and analyzing prisoners' voice data to monitor their mental state and drug rehabilitation progress, thereby providing more personalized and timely intervention measures.

[0003] The currently widely used voice service management systems mainly rely on basic voice recognition technology and data analysis methods. They can automatically record the content of prisoners' calls, analyze the emotions and key information in the voice, and help managers monitor the behavior and psychological state of prisoners. However, these systems are not adaptable to environmental factors such as accents and noise, resulting in low accuracy of voice recognition and poor ability to interpret complex psychological states.

[0004] Therefore, there is an urgent need for a technical solution that can improve recognition accuracy and enhance the ability to interpret psychological states. Summary of the invention

[0005] In order to solve the deficiencies of the prior art, the embodiments of the present application provide a natural language understanding driven intelligent voice service management system and method. The present application solves the technical problems of the existing management system such as low accuracy of voice recognition and poor ability to interpret complex psychological states.

[0006] An intelligent voice service management system driven by natural language understanding includes: a voice receiving module for receiving demands made by drug addicts through voice; wherein, after receiving the demands, the module also includes data preprocessing of the demands: Wherein, X represents the preprocessed data, x(n) represents the frame signal, N represents the number of transformation points, k represents the transformation coefficient, and n represents the sampling point index in the time domain; a semantic understanding module is used to convert the demand into text information to extract the semantic information of the text information; a dialogue management module is used to determine the next action according to the semantic information to obtain a reply text; wherein, if the semantic information belongs to the category of knowledge question and answer, the knowledge base question and answer module is called to retrieve the answer; if the semantic information belongs to the task-based demand, the personal profile database is accessed to provide personalized task guidance; a speech synthesis module is used to convert the reply text into speech to feed back to the drug addict; and a dialogue log database is used to record each round of dialogue for supervision and subsequent analysis.

[0007] In a possible implementation, converting the demand into text information to extract semantic information of the text information includes: extracting time-frequency features of the demand, the time-frequency features are used to input a speech decoding model; using the speech decoding model to iteratively scale noise prediction on the time-frequency features to obtain a speech mask sequence; encoding the speech mask sequence to obtain a target speech sequence with background noise removed; using a whisper model to convert the target speech sequence into text to obtain a text sequence; performing intent recognition on the text sequence to obtain first semantic information; and filling slots in the text sequence to obtain second semantic information.

[0008] In a possible implementation, extracting the time-frequency features of the demand includes: Wherein, Y represents the time-frequency characteristics, x(τ) represents the required signal, w(τ-t) represents the window function, ω represents the angular frequency, τ represents the independent variable in the time domain, and t represents the translation parameter of the window function, which is used to adjust the position of the window in the time domain.

[0009] In a possible implementation, the iterative scaling noise prediction of the time-frequency features using the speech decoding model to obtain the speech mask sequence further includes: performing attenuation calculation on the output of the speech decoding model after each round of iteration to increase the attention weight of the long-term dependency relationship of the context of the speech mask sequence; wherein the attenuation calculation includes: Among them, X' t represents the result of attenuation calculation, represents the weight of the Gaussian distribution, σ represents the control weight, T represents the maximum time span, τ represents the variable of the time span, and t represents the current time point.

[0010] In a possible implementation, performing intent recognition on the text sequence to obtain the first semantic information includes: constructing a rule template and category information to obtain a main domain list under each category information; using the main domain list to count the frequency distribution of keywords in the text sequence; and combining and arranging the keywords that reach the target frequency threshold with the corresponding intent according to the intent strength score to obtain the first semantic information.

[0011] In one possible implementation, filling the slots of the text sequence to obtain the second semantic information includes: using a subword segmentor to segment the text sequence to obtain a segmentation sequence; encoding the segmentation sequence using a BERT model to obtain a segmentation feature vector; and using a conditional random field to model the context information of the segmentation feature vector, and the modeled features are used to input a classification neural network to obtain the second semantic information.

[0012] In a possible implementation, when the semantic information belongs to the category of knowledge question and answer, determining the next action based on the semantic information to obtain a reply text includes: using structured query language to construct a matching statement; using the knowledge base question and answer module to match the first semantic information and the second semantic information with the header of the knowledge base through the matching statement to retrieve the answer; and using the retrieved answer as the reply text to pass it to the speech synthesis module.

[0013] In a possible implementation, when the semantic information belongs to a task-type requirement, determining the next action according to the semantic information to obtain a reply text includes: obtaining a personal file database of the drug addict; filling the personal file database, the first semantic information and the second semantic information into a preset prompt word template to obtain a target prompt word; inputting the target prompt word into multiple large language models to obtain multiple indication outputs; and integrating the multiple indication outputs, and the integrated result is used as a reply text to be passed to a speech synthesis module.

[0014] The embodiment of the present application also provides an intelligent voice service management method driven by natural language understanding, including: receiving demands made by drug addicts through voice; converting the demands into text information to extract semantic information of the text information; determining the next action based on the semantic information to obtain a reply text; converting the reply text into voice to provide feedback to the drug addicts; and recording each round of conversation for supervision and subsequent analysis.

[0015] An embodiment of the present application also provides a device for natural language understanding-driven intelligent voice service management, comprising: a processor, a memory, and a system bus; wherein the processor and the memory are connected via the system bus; the memory is used to store one or more programs, and the one or more programs include instructions, which, when executed by the processor, enable the processor to execute any one of the systems or methods described above.

[0016] Through the natural language understanding driven intelligent voice service management system and method provided above, the embodiment of the present application can improve the recognition accuracy by means of data preprocessing and generating voice mask sequences. Furthermore, in some embodiments, by introducing the integrated results of multiple large language models, the system's ability to interpret complex psychological states can be enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A schematic diagram of a natural language understanding-driven intelligent voice service management system provided in an embodiment of the present application;

[0019] Figure 2 An exemplary flowchart for extracting semantic information provided in an embodiment of the present application. DETAILED DESCRIPTION

[0020] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present disclosure unless otherwise specifically stated.

[0021] Those skilled in the art can understand that the terms "first", "second" and the like in the embodiments of the present disclosure are only used to distinguish different steps, devices or modules, etc., and neither represent any specific technical meaning nor represent the necessary logical order between them. It should also be understood that in the embodiments of the present disclosure, "multiple" can refer to two or more, and "at least one" can refer to one, two or more. It should also be understood that for any component, data or structure mentioned in the embodiments of the present disclosure, in the absence of explicit limitation or contrary revelation given in the context, it can generally be understood as one or more. In addition, the term "and / or" in the present disclosure is only a description of the association relationship of the associated objects, indicating that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in the present disclosure generally indicates that the associated objects before and after are an "or" relationship. It should also be understood that the description of each embodiment in the present disclosure emphasizes the differences between the embodiments, and the same or similar parts can refer to each other. For the sake of brevity, they will not be repeated one by one.

[0022] At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present disclosure and its application or use. The techniques, methods and devices known to ordinary technicians in the relevant fields may not be discussed in detail, but where appropriate, the techniques, methods and devices should be considered part of the specification. It should be noted that similar numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0023] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0024] Figure 1 A schematic diagram of a natural language understanding-driven intelligent voice service management system provided for an embodiment of the present application. It should be understood that the system shown in the figure is exemplary and non-restrictive. This means that the module architecture involved is not limited to a specific form or design, but is presented as an example. In other words, the architecture shown in the figure can be regarded as a way of expression to clearly describe related concepts and relationships, and does not exclude other forms of architecture. Therefore, when interpreting the architecture in the picture, it should be understood that the module has flexibility and diversity, and its purpose is to provide an exemplary description rather than a restrictive provision on a specific form.

[0025] like Figure 1 As shown, at the voice receiving module 101, the module is used to receive the needs of drug addicts through voice. Specifically, it may include a sound collection device (such as a microphone), a sound converter, etc., which is used to convert the sound signal into an electrical signal for further processing. The received sound signal will contain various information, such as the speaker's tone, volume, and language content. Among them, after receiving the demand, it also includes data preprocessing of the demand:

[0026]

[0027] Among them, X represents the preprocessed data, x(n) represents the frame signal, N represents the number of transformation points, k represents the transformation coefficient, and n represents the sampling point index in the time domain. It should be understood that data preprocessing is the first step in signal processing, and its purpose is to improve the quality of useful information in the signal and reduce the influence of noise and irrelevant information.

[0028] Next, at the semantic understanding module 102, the module is used to convert the demand into text information to extract semantic information of the text information. Figure 2 The following is an exemplary flow chart of extracting semantic information provided by the embodiment of the present application. Specifically, Figure 2 As shown, at step S201, firstly, the time-frequency features of the requirement are extracted, and the time-frequency features are used to input the speech decoding model. The extraction of the time-frequency features of the requirement includes:

[0029]

[0030] Among them, Y represents the time-frequency characteristics, x(τ) represents the required signal, w(τ-t) represents the window function, ω represents the angular frequency, τ represents the independent variable in the time domain, and t represents the translation parameter of the window function, which is used to adjust the position of the window in the time domain. Through Fourier transform, the signal in the time domain is converted into a representation in the frequency domain, thereby revealing the relationship between the frequency component and time distribution of the signal. The application of the window function is to reduce the boundary effect. By taking a small time window near each time point for local Fourier transform, the time-frequency characteristics can simultaneously reflect the information of time and frequency.

[0031] After obtaining the time-frequency features in step S202, the speech decoding model is used to iteratively scale the noise prediction of the time-frequency features to obtain a speech mask sequence. The speech mask sequence is a representation method that indicates the probability of the existence of a sound signal at a specific time-frequency point through a series of values. By applying the speech mask, the model can distinguish between speech and non-speech parts, thereby effectively extracting the target speech in a complex noise environment. Among them, after each round of iteration, the output of the speech decoding model can be attenuated to increase the attention weight of the long-term dependency of the context of the speech mask sequence. In one embodiment, the attenuation calculation includes:

[0032]

[0033] Among them, X' t represents the result of attenuation calculation, represents the weight of the Gaussian distribution, σ represents the control weight, T represents the maximum time span, τ represents the variable of the time span, and t represents the current time point. This Gaussian weight-based decay calculation method can effectively adjust the model's attention to contextual information at different time scales, thereby improving the ability to handle long-term dependencies.

[0034] Next, at step S203, the speech mask sequence is encoded to obtain a target speech sequence with background noise removed. Then, at step S204, the target speech sequence is converted into text using a whisper model to obtain a text sequence.

[0035] Combined with the above content, it can be seen that the voice mask sequence is a sequence used to represent the probability of the existence of a sound signal in the time-frequency domain. It is generated by a certain mathematical model and can indicate the existence of the target voice signal at a specific time-frequency point. In the encoding process, the voice mask sequence is transformed by a series of complex algorithms to extract and separate the target voice sequence from it. This process can include noise suppression, echo cancellation and other signal processing techniques to ensure that the clear target voice is accurately separated from the background noise. Subsequently, the encoded target voice sequence is input into the whisper model. The whisper model is a speech recognition technology that can convert a voice signal into a corresponding text sequence. This conversion process involves a deep neural network, which can learn the complex mapping relationship between voice signals and text. In this way, the whisper model converts the target voice sequence into an accurate text sequence, thereby achieving high-accuracy speech recognition.

[0036] After obtaining the text sequence output by the whisper model in step S205, the text sequence is subjected to intent recognition to obtain the first semantic information. In one embodiment, a rule template and category information are first constructed to obtain a main domain list under each category information. In this step, it is necessary to define and summarize the intent categories that may be involved in the text sequence. For example, the intent category may be an inquiry, a command, a request, etc. Under each intent category, a main domain list is further defined, and the main domain refers to the most core and frequently occurring subject vocabulary in the intent category, such as "purchase", "booking", "query", etc. The construction of these main domain lists usually relies on domain knowledge and typical use case analysis.

[0037] Then, the frequency distribution of keywords in the text sequence is counted using the main domain list. The statistics of keyword frequency not only need to calculate the number of times each keyword appears in the text sequence, but also need to consider its distribution in the entire data set to evaluate the importance and relevance of the keyword. Here, mathematical models such as TF-IDF (Term Frequency-Inverse Document Frequency) can be applied to optimize the keyword screening and scoring process.

[0038] And the keywords reaching the target frequency threshold are combined and arranged with the corresponding intents according to the intent strength scores to obtain the first semantic information. The intent strength score can be calculated by a variety of algorithms, for example, it can be comprehensively evaluated based on factors such as the correlation between the keyword and the main domain, the location information of the keyword in the text, and the contextual relationship.

[0039] While obtaining the first semantic information in step S206, the text sequence can be slot-filled to obtain the second semantic information. In one embodiment, the text sequence is first segmented using a subword tokenizer to obtain a segmentation sequence. The use of a subword tokenizer is based on a finer-grained analysis of the text sequence in order to more accurately capture the semantic information. Subword segmentation technology can split complex or unknown words into smaller, meaningful units (such as subwords, characters, etc.), which helps to process vocabulary diversity and new words in a language. For example, using the Byte Pair Encoding (BPE) algorithm or the WordPiece algorithm, a text sequence can be effectively cut into a series of subwords.

[0040] Secondly, the word segmentation sequence is encoded by the BERT model to obtain the word segmentation feature vector. The BERT model learns the contextual representation of word segmentation through the bidirectional encoder of Transformer. The feature vector of each word segmentation not only contains its own information, but also integrates its contextual information in the sentence. The training process of the BERT model includes two parts: Masked Language Model (MLM) and Next Sentence Prediction (NSP), which enables the model to capture deep semantic associations.

[0041] And use conditional random fields to model the context information of the word segmentation feature vector, and the modeled features are used to input the classification neural network to obtain the second semantic information. CRF is a sequence labeling model that can consider the dependencies between adjacent labels and model the optimal label combination of the entire sequence. In this example, the CRF layer will be used to capture and model the correlations and constraints in the word segmentation sequence, providing a more refined feature representation for the classification neural network. The features output by the CRF layer are sent to the classification neural network, and the network structure can be a fully connected layer, a convolutional neural network (CNN) or a recurrent neural network (RNN), etc. The goal of this layer is to make a final category judgment based on the extracted features, obtain the label of each word segmentation (such as entity name, attribute value, etc.), thereby completing the slot filling and obtaining the second semantic information.

[0042] Back to Figure 1 As shown in the content, at the dialogue management module 103, the module is used to determine the next action based on the semantic information to obtain a reply text; wherein, if the semantic information belongs to the category of knowledge question and answer, the knowledge base question and answer module is called to retrieve the answer; if the semantic information belongs to task-based requirements, the personal profile database is accessed to provide personalized task guidance.

[0043] In an implementation scenario, when the semantic information belongs to the category of knowledge question and answer, determining the next action according to the semantic information to obtain the reply text includes: first constructing a matching statement using structured query language. It should be clear that the semantic information extraction refers to extracting key information (such as keywords, entities, concepts, etc.) from the user's query through natural language processing technology, and converting it into structured information that can be processed by the computer. In the scenario of knowledge question and answer, this step is crucial because it determines whether the system can accurately understand the user's query intention and retrieve relevant answers accordingly. The system first constructs a matching statement using structured query language (SQL). Specifically, one or more query statements can be formed by converting the extracted semantic information into SQL query conditions. For example, if the user queries "Newton's second law", the system may extract "Newton" and "second law" as keywords, and then convert these keywords into query conditions in the SQL statement.

[0044] Secondly, the knowledge base question-answering module is used to match the first semantic information and the second semantic information with the header of the knowledge base through a matching statement to retrieve the answer. This step involves database operations, including but not limited to multi-table join query, subquery, etc., to ensure that the answer that meets the query conditions can be accurately retrieved from the knowledge base. After obtaining the reply text, the retrieved answer is used as the reply text to be passed to the speech synthesis module.

[0045] In another implementation scenario, when the semantic information belongs to a task-based requirement, determining the next action according to the semantic information to obtain a reply text includes: first obtaining a personal file database of the drug addict. This database includes but is not limited to basic personal information, drug rehabilitation history, treatment response, etc.

[0046] Then the personal profile database, the first semantic information and the second semantic information are filled into a preset prompt word template to obtain a target prompt word. These semantic information may include recent behavior patterns, language usage habits, emotional changes, etc. These data are filled into the preset prompt word template together with the personal profile.

[0047] This template is carefully designed to ensure that useful and targeted target prompt words can be extracted from it. For example, if the template is designed as: "Taking into account [first semantic information] and [second semantic information], it is recommended to take the following measures:", then after filling in the corresponding information, the target prompt words that may be obtained are "Taking into account the patient's recent mood swings and past drug rehabilitation experience, it is recommended to adopt emotional stability training and cognitive behavioral therapy." The target prompt words are input into multiple large language models to obtain multiple instruction outputs. And the multiple instruction outputs are integrated, and the integrated results are used as reply text to be passed to the speech synthesis module. These models can be trained using different algorithms or data sets, and can generate a variety of possible instruction outputs based on the input prompt words. For example, one model may emphasize emotional support, while another model may emphasize drug treatment.

[0048] Finally, the multiple indicative outputs generated by these models are integrated. This process can be implemented through mathematical models such as weighted average, voting mechanism, or more complex algorithms (such as model fusion technology based on deep learning) to obtain the final response text that takes all aspects into consideration. For example, if two models recommend "increasing the number of psychological counseling" and "adjusting the dosage of medication" respectively, the integration process may combine these suggestions to form a more comprehensive treatment plan.

[0049] In this way, the final response text is not only based on individual detailed profile information, but also combines the intelligent analysis and suggestions of multiple professional models, and is converted into speech output through the speech synthesis module, providing drug addicts with more accurate and personalized support plans.

[0050] At the speech synthesis module 104, the module is used to convert the reply text into speech to feed back to the drug addict.

[0051] At the conversation log database 105, the database is used to record each round of conversation for supervision and subsequent analysis.

[0052] Furthermore, an embodiment of the present application also provides a device for natural language understanding-driven intelligent voice service management, comprising: a processor, a memory, and a system bus; the processor and the memory are connected via the system bus; the memory is used to store one or more programs, and the one or more programs include instructions, which, when executed by the processor, enable the processor to execute any one of the implementation methods of the above-mentioned processing method.

[0053] Furthermore, an embodiment of the present application also provides a computer program product, which, when executed on a terminal device, enables the terminal device to execute any one of the implementation methods of the above-mentioned processing method.

[0054] It can be known from the description of the above implementation mode that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., including several instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0055] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0056] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0057] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent voice service management system driven by natural language understanding, It is characterized in that include: The voice receiving module is used to receive the demands made by drug addicts through voice; after receiving the demands, the module also includes data preprocessing of the demands: Wherein, X represents the preprocessed data, x(n) represents the frame signal, N represents the number of transformation points, k represents the transformation coefficient, and n represents the sampling point index in the time domain; The semantic understanding module is used to convert the demand into text information to extract the semantic information of the text information, specifically including: extracting the time-frequency features of the demand, the time-frequency features are used to input the speech decoding model; using the speech decoding model to iteratively scale the noise prediction of the time-frequency features to obtain a speech mask sequence; encoding the speech mask sequence to obtain a target speech sequence with background noise removed; using a whisper model to convert the target speech sequence into text to obtain a text sequence; performing intent recognition on the text sequence to obtain first semantic information; and slot filling the text sequence to obtain second semantic information; wherein, using the speech decoding model to iteratively scale the noise prediction of the time-frequency features to obtain the speech mask sequence also includes: performing attenuation calculation on the output of the speech decoding model after each round of iteration to increase the attention weight of the long-term dependency relationship of the context of the speech mask sequence; wherein the attenuation calculation includes: Among them, X′ t represents the result of attenuation calculation, represents the weight of Gaussian distribution, σ represents the control weight, T represents the maximum time span, τ represents the variable of time span, and τ represents the current time point; A dialogue management module is used to determine the next action according to the semantic information to obtain a reply text; if the semantic information belongs to the category of knowledge question and answer, the knowledge base question and answer module is called to retrieve the answer; if the semantic information belongs to task-based requirements, the personal profile database is accessed to provide personalized task guidance; A speech synthesis module, used for converting the reply text into speech to provide feedback to the drug addict; and The conversation log database is used to record each round of conversation for supervision and subsequent analysis.

2. The system according to claim 1, It is characterized in that The time-frequency features of the demand are extracted as follows: Wherein, Y represents the time-frequency characteristics, x(τ) represents the required signal, w(τ-t) represents the window function, ω represents the angular frequency, τ represents the independent variable in the time domain, and t represents the translation parameter of the window function, which is used to adjust the position of the window in the time domain.

3. The system according to claim 1, It is characterized in that in, The performing intent recognition on the text sequence to obtain first semantic information includes: Construct rule templates and category information to obtain a list of primary domains under each category information; Using the main domain list to count the frequency distribution of keywords in the text sequence; and The keywords reaching the target frequency threshold are combined and arranged with the corresponding intents according to the intent strength scores to obtain the first semantic information.

4. The system according to claim 1, It is characterized in that in, The step of filling the slots of the text sequence to obtain the second semantic information includes: Using a subword segmenter to segment the text sequence to obtain a segmentation sequence; Encoding the word segmentation sequence using a BERT model to obtain a word segmentation feature vector; and The context information of the word segmentation feature vector is modeled using a conditional random field, and the modeled features are used to input into a classification neural network to obtain second semantic information.

5. The system according to claim 1, It is characterized in that in, If the semantic information belongs to the category of knowledge question and answer, the next action is determined according to the semantic information to obtain a reply text including: Use structured query language to construct matching statements; Using the knowledge base question-answering module to match the first semantic information and the second semantic information with a header of the knowledge base through a matching statement to retrieve an answer; and The retrieved answer is used as a reply text to be passed to a speech synthesis module.

6. The system according to claim 1, It is characterized in that in, When the semantic information belongs to a task-based requirement, determining the next action according to the semantic information to obtain a reply text includes: Obtaining a personal file database of the drug addict; Filling the personal profile database, the first semantic information and the second semantic information into a preset prompt word template to obtain a target prompt word; Inputting the target prompt word into a plurality of large language models to obtain a plurality of indication outputs; and The multiple indication outputs are integrated, and the integrated result is used as a reply text to be passed to the speech synthesis module.

7. A natural language understanding driven intelligent voice service management method, It is characterized in that include: Receiving the demands made by the drug addicts through voice; wherein, after receiving the demands, the process further includes performing data preprocessing on the demands: Wherein, X represents the preprocessed data, x(n) represents the frame signal, N represents the number of transformation points, k represents the transformation coefficient, and n represents the sampling point index in the time domain; The requirement is converted into text information to extract the semantic information of the text information, specifically including: extracting the time-frequency features of the requirement, the time-frequency features are used to input the speech decoding model; using the speech decoding model to iteratively scale the noise prediction of the time-frequency features to obtain a speech mask sequence; encoding the speech mask sequence to obtain a target speech sequence with background noise removed; using a whisper model to convert the target speech sequence into text to obtain a text sequence; performing intent recognition on the text sequence to obtain first semantic information; and slot filling the text sequence to obtain second semantic information; wherein, using the speech decoding model to iteratively scale the noise prediction of the time-frequency features to obtain the speech mask sequence also includes: performing attenuation calculation on the output of the speech decoding model after each round of iteration to increase the attention weight of the long-term dependency relationship of the context of the speech mask sequence; wherein the attenuation calculation includes: Among them, X′ t represents the result of attenuation calculation, represents the weight of Gaussian distribution, σ represents the control weight, T represents the maximum time span, τ represents the variable of time span, and t represents the current time point; Determine the next action according to the semantic information to obtain a reply text; if the semantic information belongs to the knowledge question and answer category, call the knowledge base question and answer module to retrieve the answer; if the semantic information belongs to the task-based requirement, access the personal profile database to provide personalized task guidance; Convert the reply text into speech to provide feedback to the drug rehabilitation personnel; and Each conversation was recorded for monitoring and subsequent analysis.

8. A device for natural language understanding driven voice service management, It is characterized in that include: A processor, a memory, and a system bus; wherein the processor and the memory are connected via the system bus; The memory is used to store one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by the processor, the processor is caused to perform the method of claim 7.

Citation Information

Patent Citations

  • Voice question-answering system based on knowledge graph and application method thereof

    CN111597308A

  • Man-machine conversation method and device, electronic device and storage medium

    CN112365892A

  • Speech enhancement system, method, device and equipment

    CN115472153A

  • Robot dialogue method and device, electronic equipment and storage medium

    CN115617972A

  • Single-channel speech enhancement model training method and device, equipment and storage medium

    CN117558288A