Voice single intention recognition method and device for remote call scene
Through the integration of speech preprocessing and multimodal feature, the impact of background noise and intelligent paging scene specificity in remote call scenarios are solved, and the accuracy of speech intention recognition and decision reference value are improved.
Patent Information
- Application Number
- CN202411237724.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-04
- Publication Date
- 2024-12-10
AI Technical Summary
In remote call scenarios, the uncertainty of the scene where the real person is in leads to uncontrollable call background noise, affecting the accuracy of voice intention recognition. The existing semantic intention recognition methods do not fully consider the specific factors of the intelligent paging scenario, and it is difficult to make subsequent decisions Provides fine-grained reference value.
Voice preprocessing technology is used to denoise and text extraction to extract speech, text and speech duration features. Through single-round multi-modal feature fusion and multiple-round feature fusion, combined with attention and dimensionality reduction network, the probability of single-target intention tendency is identified.
It improves the accuracy of speech intention recognition and the integrity of feature information, provides fine-grained decision-making reference value, and enhances the generalization and usability of speech intention recognition.
Smart Images
Figure CN119107941A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field related to speech intention recognition, and in particular to a method and device for single speech intention recognition in a remote call scenario. Background Art
[0002] As the Internet drives the rapid growth of demand for the service industry, voice customer service robots are used to communicate with real people to meet the communication needs of large numbers of users. In the existing technology, voice recognition is usually converted into text to perform semantic recognition and understanding of the text content, obtain the customer's intention and generate a dialogue strategy.
[0003] In remote call scenarios, the uncertainty of the scene in which the real person is located makes the background noise of the call uncontrollable. Although the communication equipment itself has a certain background noise reduction function, there is still some background interference that affects the recognition accuracy of the voice intent recognition model; the existing semantic intent recognition method mainly uses voice features and text features to recognize the single intention of the real person, and does not fully consider the positive and negative impact of specific factors in the intelligent paging scenario on intent recognition; and in the existing semantic intent recognition method, voice intent recognition is classified as a classification problem, which is difficult to provide fine-grained reference value for subsequent decision-making. Summary of the invention
[0004] The purpose of the present invention is to provide a method and device for voice single intention recognition in a remote call scenario to solve the problems raised in the above background technology.
[0005] To achieve the above object, the present invention provides the following technical solutions:
[0006] A method for recognizing a single voice intention in a remote call scenario, comprising:
[0007] Performing speech preprocessing, speech denoising and text extraction, wherein the speech preprocessing specifically includes speech duration extraction, speech denoising and text extraction performed in sequence;
[0008] Extracting feature vectors of each modal information; the extraction includes speech feature extraction, speech text feature extraction, text feature extraction, and speech duration feature extraction;
[0009] Perform single-round multi-modal feature vector fusion and multi-round feature vector fusion;
[0010] The probability of identifying a single-target intention tendency is expressed as the probability possibility that it can be regarded as a single target.
[0011] As a further solution of the present invention: the feature fusion includes single-round multimodal feature fusion and multi-round feature fusion in sequence;
[0012] The single-round multimodal feature fusion sequentially includes a first feature splicing network, a first attention network, and a first dimensionality reduction network; the feature fusion module sequentially includes a second feature splicing network, a second attention network, and a second dimensionality reduction network.
[0013] As a further solution of the present invention: the speech preprocessing, speech denoising and text extraction, the speech preprocessing specifically includes the steps of speech duration extraction, speech denoising and text extraction performed in sequence specifically including:
[0014] Acquire multi-round conversation data in a smart paging scenario, the multi-round conversation data consisting of one or more single-round conversation sub-data arranged in time sequence, the single-round conversation sub-data including: a first text and a second voice;
[0015] Inputting the second speech in the single-round conversation sub-data into the speech preprocessing unit: obtaining the second speech duration attribute of the second speech by speech duration extraction;
[0016] By removing the environmental noise of the second voice through voice denoising, high-quality second voice information is obtained to improve voice usability;
[0017] The obtained high-quality second voice information is input into a text extraction module to obtain second text information corresponding to the second voice.
[0018] As a further solution of the present invention: the extracting of feature vectors of each modal information; the extracting respectively includes the steps of speech feature extraction, speech text feature extraction, text feature extraction, and speech duration feature extraction, specifically including:
[0019] Performing text feature extraction, voice feature extraction, voice text feature extraction, and voice duration feature extraction on the first text, the second voice information, the second text, and the second voice duration attribute, respectively, to obtain corresponding intermediate feature vectors;
[0020] The intermediate feature vectors respectively include: a first text feature vector, a second voice feature vector, a second text feature vector, and a second voice duration attribute feature vector.
[0021] As a further solution of the present invention: the steps of performing single-round fusion of multimodal feature vectors and fusion of multiple rounds of feature vectors specifically include:
[0022] Simultaneously performing a single-round multimodal feature fusion on the plurality of intermediate feature vectors, deeply fusing the feature vectors of each modality through a first feature concatenation network and a first attention network, and obtaining the corresponding single-round multimodal feature vector after dimensionality reduction through a first dimensionality reduction network;
[0023] According to the time sequence of the single-round dialogue sub-data in the multi-round dialogue data and based on multi-step cyclic training, a number of single-round multimodal feature vectors arranged according to the time sequence are obtained;
[0024] One or several single-round multimodal feature vectors arranged in time sequence are simultaneously subjected to multi-round feature fusion, each single-round multimodal feature vector is deeply fused through the second feature splicing network and the second attention network, and the reduced single-intention feature vector is obtained through the second dimensionality reduction network.
[0025] The embodiment of the present invention aims to provide a voice single intention recognition device for a remote call scenario, comprising:
[0026] A speech preprocessing module is used to perform speech preprocessing, speech denoising and text extraction, wherein the speech preprocessing specifically includes speech duration extraction, speech denoising and text extraction performed in sequence;
[0027] A multi-feature extraction module is used to extract feature vectors of each modal information; the extraction includes speech feature extraction, speech text feature extraction, text feature extraction, and speech duration feature extraction;
[0028] A feature fusion module, used to perform single-round multi-modal feature vector fusion and multi-round feature vector fusion;
[0029] The single intention recognition module is used to identify the probability of a single target intention tendency. The specific output result is expressed as the probability possibility of being regarded as a single target.
[0030] As a further solution of the present invention: the feature fusion module sequentially comprises a single-round multimodal feature fusion unit and a multi-round feature fusion unit;
[0031] The single-round multimodal feature fusion module sequentially includes a first feature splicing network, a first attention network, and a first dimensionality reduction network; the feature fusion module sequentially includes a second feature splicing network, a second attention network, and a second dimensionality reduction network.
[0032] As a further solution of the present invention: the speech preprocessing module comprises:
[0033] A data acquisition unit is used to acquire multi-round conversation data in an intelligent paging scenario, wherein the multi-round conversation data is composed of one or more single-round conversation sub-data arranged in time sequence, and the single-round conversation sub-data includes: a first text and a second voice;
[0034] The preprocessing unit is used to input the second speech in the single-round dialogue sub-data into the speech preprocessing unit: obtain the second speech duration attribute of the second speech by speech duration extraction;
[0035] A denoising processing unit, configured to remove environmental noise of the second voice through voice denoising, and obtain high-quality second voice information to improve voice usability;
[0036] The text processing unit is used to input the acquired high-quality second voice information into the text extraction module to acquire the second text information corresponding to the second voice.
[0037] As a further solution of the present invention: the multi-feature extraction module includes:
[0038] The sequential processing extraction unit is used to perform text feature extraction, voice feature extraction, voice-text feature extraction, and voice duration feature extraction on the first text, the second voice information, the second text, and the second voice duration attribute, respectively, to obtain corresponding intermediate feature vectors; the intermediate feature vectors respectively correspond to: the first text feature vector, the second voice feature vector, the second text feature vector, and the second voice duration attribute feature vector.
[0039] As a further solution of the present invention: the single intention recognition module includes:
[0040] A feature fusion unit, used to simultaneously perform a single-round multimodal feature fusion on the plurality of intermediate feature vectors, deeply fuse the feature vectors of each modality through a first feature concatenation network and a first attention network, and obtain the corresponding single-round multimodal feature vector after dimensionality reduction through a first dimensionality reduction network;
[0041] A cyclic training unit, for obtaining a plurality of single-round multimodal feature vectors arranged in time sequence based on the time sequence of single-round dialogue sub-data in the multi-round dialogue data and based on multi-step cyclic training;
[0042] The fusion processing unit is used to simultaneously perform multi-round feature fusion on one or several single-round multimodal feature vectors arranged in time sequence, deeply fuse each single-round multimodal feature vector through the second feature splicing network and the second attention network, and obtain the reduced single-intention feature vector through the second dimensionality reduction network.
[0043] Compared with the prior art, the beneficial effects of the present invention are: in view of the strong correlation between voice duration and intention in multi-round intelligent paging scenarios, call duration feature information is introduced to improve the effectiveness of feature information; in view of the uncontrollable problem of call background noise caused by the uncertainty phenomenon of the scene in which real people are located, a speech denoising algorithm is introduced in the data preprocessing link to improve the usability of call voice and the scenario versatility of the method; considering the causal relationship between the output text of the intelligent robot and the single intent recognition in the multi-round intelligent paging scenario, the first text feature is introduced in the data feature extraction link to improve the integrity of the feature information. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 The present invention is a flowchart of a method for single-intention voice recognition in a remote call scenario.
[0045] Figure 2 The present invention is a flowchart of speech preprocessing in a speech single intent recognition method for a remote call scenario.
[0046] Figure 3 The present invention is a block diagram of a voice single-intent recognition device for a remote call scenario. DETAILED DESCRIPTION
[0047] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0048] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.
[0049] like Figure 1 The present invention provides a method for recognizing a single voice intention in a remote call scenario according to an embodiment of the present invention, comprising the following steps:
[0050] Step S1, performing speech preprocessing, speech denoising and text extraction, wherein the speech preprocessing specifically includes speech duration extraction, speech denoising and text extraction performed in sequence;
[0051] Step S2, extracting feature vectors of each modal information; the extraction includes speech feature extraction, speech text feature extraction, text feature extraction, and speech duration feature extraction;
[0052] Step S3, performing single-round multi-modal feature vector fusion and multi-round feature vector fusion;
[0053] Step S4, identifying the probability of a single target intention tendency, and the specific output result is expressed as the probability possibility that it can be regarded as a single target.
[0054] In this embodiment, a method for single-intention recognition of speech is proposed. The method reduces the negative impact of background noise through speech preprocessing and improves the availability of sound sources. In view of the specificity of intelligent paging scenarios, the method integrates the text features of intelligent machines and the speech duration features of real people on the basis of traditional speech features and text features to enhance the information richness of data features. Single-intention recognition is regarded as a regression task, and the availability of the judgment result in fine-grained reference value is improved by outputting the result probability. First, a deep learning speech separation method is introduced in the speech preprocessing stage to extract relatively clean speech information and improve the availability of call speech. Secondly, for this call scenario, in the traditional speech intention recognition method, On this basis (voice and text features), call duration features and text features of the intelligent robot's reply are introduced to improve the richness of intent features; finally, a regression method is used rather than a simple classification method to output the results of single intent recognition, so as to improve the reference value of the output results; different from the conventional method that regards intent recognition as a classification task, the task mechanism of the invention converts intent recognition into a regression task, and the final output result is a probability value that can be regarded as a single intent, which is specifically expressed as a continuous value between 0% and 100% (accurate to the fourth decimal place). The transformation of the task mechanism enables the method of the invention to provide fine-grained data reference, and provide a more scientific and sufficient decision-making basis for downstream task decisions.
[0055] As another preferred embodiment of the present invention, the feature fusion sequentially includes single-round multimodal feature fusion and multi-round feature fusion;
[0056] The single-round multimodal feature fusion module sequentially includes a first feature splicing network, a first attention network, and a first dimensionality reduction network; the feature fusion module sequentially includes a second feature splicing network, a second attention network, and a second dimensionality reduction network.
[0057] like Figure 2 As shown, as another preferred embodiment of the present invention, the speech preprocessing, speech denoising and text extraction are performed, and the speech preprocessing specifically includes the steps of speech duration extraction, speech denoising and text extraction performed in sequence, specifically including:
[0058] S21, obtaining multi-round conversation data in a smart paging scenario, the multi-round conversation data consisting of one or more single-round conversation sub-data arranged in time sequence, the single-round conversation sub-data including: a first text and a second voice;
[0059] S22, inputting the second speech in the single-round dialogue sub-data into the speech preprocessing unit: obtaining a second speech duration attribute of the second speech by speech duration extraction;
[0060] S23, removing environmental noise of the second voice through voice denoising, and obtaining high-quality second voice information to improve voice usability;
[0061] S24, inputting the obtained high-quality second voice information into a text extraction module to obtain second text information corresponding to the second voice.
[0062] In this embodiment, step S1 is further expanded, wherein the specific data format of the single conversation sub-data is [[first text][second voice]], the [first text] in the single conversation sub-data is predefined template text data or text data output by an intelligent robot, and the [second voice] in the single conversation sub-data is actual voice data of a real person in an intelligent paging scenario.
[0063] As another preferred embodiment of the present invention, the extracting of feature vectors of each modal information; the extracting respectively includes the steps of speech feature extraction, speech text feature extraction, text feature extraction, and speech duration feature extraction, specifically including:
[0064] Performing text feature extraction, voice feature extraction, voice text feature extraction, and voice duration feature extraction on the first text, the second voice information, the second text, and the second voice duration attribute, respectively, to obtain corresponding intermediate feature vectors;
[0065] The intermediate feature vectors respectively include: a first text feature vector, a second voice feature vector, a second text feature vector, and a second voice duration attribute feature vector.
[0066] Furthermore, the steps of performing single-round fusion of multimodal feature vectors and multi-round fusion of feature vectors specifically include:
[0067] Simultaneously performing a single-round multimodal feature fusion on the plurality of intermediate feature vectors, deeply fusing the feature vectors of each modality through a first feature concatenation network and a first attention network, and obtaining the corresponding single-round multimodal feature vector after dimensionality reduction through a first dimensionality reduction network;
[0068] According to the time sequence of the single-round dialogue sub-data in the multi-round dialogue data and based on multi-step cyclic training, a number of single-round multimodal feature vectors arranged according to the time sequence are obtained;
[0069] One or several single-round multimodal feature vectors arranged in time sequence are simultaneously subjected to multi-round feature fusion, each single-round multimodal feature vector is deeply fused through the second feature splicing network and the second attention network, and the reduced single-intention feature vector is obtained through the second dimensionality reduction network.
[0070] In this embodiment, specifically, the [first text], [second voice information], [second text], and [second voice duration attribute] are respectively input into the text feature extraction module, the voice feature extraction module, the voice-text feature extraction module, and the voice duration feature extraction module in the multi-feature extraction unit to obtain the [first text feature vector], [second voice feature vector], [second text feature vector], and [second voice duration attribute feature vector].
[0071] The [first text feature vector], [second speech feature vector], [second text feature vector], and [second speech duration attribute feature vector] are simultaneously input into the single-round multimodal feature fusion module in the feature fusion unit, and the feature vectors of each modality are deeply fused through the first feature concatenation network and the first attention network, and the reduced [single-round multimodal feature vector] is obtained through the first dimensionality reduction network.
[0072] According to the time sequence of the single-round dialogue sub-data in the multi-round dialogue data, a loop is executed to obtain one or more [single-round multimodal feature vectors] arranged according to the time sequence.
[0073] One or several [single-round multimodal feature vectors] obtained in time sequence are simultaneously input into the multi-round feature fusion module in the feature fusion unit, and each [single-round multimodal feature vector] is deeply fused through the second feature splicing network and the second attention network, and the reduced-dimensional [single-intention feature vector] is obtained through the second dimensionality reduction network.
[0074] The [single intent feature vector] is input into the single intent recognition unit, and the regression probability of multi-round conversation data in the smart paging scenario on the single intent is obtained through the regression algorithm.
[0075] like Figure 3 As shown, the present invention also provides a voice single intention recognition device for a remote call scenario, which comprises:
[0076] The speech preprocessing module 100 is used to perform speech preprocessing, speech denoising and text extraction, wherein the speech preprocessing specifically includes speech duration extraction, speech denoising and text extraction performed in sequence;
[0077] A multi-feature extraction module 200 is used to extract feature vectors of each modal information; the extraction includes speech feature extraction, speech text feature extraction, text feature extraction, and speech duration feature extraction;
[0078] A feature fusion module 300 is used to perform single-round fusion of multi-modal feature vectors and fusion of multiple rounds of feature vectors;
[0079] The single intention recognition module 400 is used to identify the probability of a single target intention tendency, and the specific output result is expressed as the probability possibility that it can be regarded as a single target.
[0080] As another preferred embodiment of the present invention, the feature fusion module sequentially comprises a single-round multimodal feature fusion unit and a multi-round feature fusion unit;
[0081] The single-round multimodal feature fusion module sequentially includes a first feature splicing network, a first attention network, and a first dimensionality reduction network; the feature fusion module sequentially includes a second feature splicing network, a second attention network, and a second dimensionality reduction network.
[0082] As another preferred embodiment of the present invention, the speech preprocessing module includes:
[0083] A data acquisition unit is used to acquire multi-round conversation data in an intelligent paging scenario, wherein the multi-round conversation data is composed of one or more single-round conversation sub-data arranged in time sequence, and the single-round conversation sub-data includes: a first text and a second voice;
[0084] The preprocessing unit is used to input the second speech in the single-round dialogue sub-data into the speech preprocessing unit: obtain the second speech duration attribute of the second speech by speech duration extraction;
[0085] A denoising processing unit, configured to remove environmental noise of the second voice through voice denoising, and obtain high-quality second voice information to improve voice usability;
[0086] The text processing unit is used to input the acquired high-quality second voice information into the text extraction module to acquire the second text information corresponding to the second voice.
[0087] As another preferred embodiment of the present invention, the multi-feature extraction module includes:
[0088] The sequential processing extraction unit is used to perform text feature extraction, voice feature extraction, voice-text feature extraction, and voice duration feature extraction on the first text, the second voice information, the second text, and the second voice duration attribute, respectively, to obtain corresponding intermediate feature vectors; the intermediate feature vectors respectively correspond to: the first text feature vector, the second voice feature vector, the second text feature vector, and the second voice duration attribute feature vector.
[0089] As another preferred embodiment of the present invention, the single intent recognition module includes:
[0090] A feature fusion unit, used to simultaneously perform a single-round multimodal feature fusion on the plurality of intermediate feature vectors, deeply fuse the feature vectors of each modality through a first feature concatenation network and a first attention network, and obtain the corresponding single-round multimodal feature vector after dimensionality reduction through a first dimensionality reduction network;
[0091] A cyclic training unit, for obtaining a plurality of single-round multimodal feature vectors arranged in time sequence based on the time sequence of single-round dialogue sub-data in the multi-round dialogue data and based on multi-step cyclic training;
[0092] The fusion processing unit is used to simultaneously perform multi-round feature fusion on one or several single-round multimodal feature vectors arranged in time sequence, deeply fuse each single-round multimodal feature vector through the second feature splicing network and the second attention network, and obtain the reduced single-intention feature vector through the second dimensionality reduction network.
[0093] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
[0094] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the disclosure in the specification and examples. This application is intended to cover any variations, uses or adaptations of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.
[0095] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.
Claims
1. A method for recognizing a single voice intention in a remote call scenario, characterized in that: Include: Performing speech preprocessing, speech denoising and text extraction, wherein the speech preprocessing specifically includes speech duration extraction, speech denoising and text extraction performed in sequence; Extracting feature vectors of each modal information; the extraction includes speech feature extraction, speech text feature extraction, text feature extraction, and speech duration feature extraction; Perform single-round multi-modal feature vector fusion and multi-round feature vector fusion; The probability of identifying a single-target intention tendency is expressed as the probability possibility that it can be regarded as a single target.
2. According to the method for recognizing a single voice intention in a remote call scenario according to claim 1, it is characterized in that: The feature fusion includes single-round multimodal feature fusion and multi-round feature fusion in sequence; The single-round multimodal feature fusion sequentially includes a first feature splicing network, a first attention network, and a first dimensionality reduction network; the feature fusion module sequentially includes a second feature splicing network, a second attention network, and a second dimensionality reduction network.
3. The method for recognizing a single voice intention in a remote call scenario according to claim 2, characterized in that: The speech preprocessing, speech denoising and text extraction, the speech preprocessing specifically includes the steps of speech duration extraction, speech denoising and text extraction performed in sequence specifically including: Acquire multi-round conversation data in a smart paging scenario, the multi-round conversation data consisting of one or more single-round conversation sub-data arranged in time sequence, the single-round conversation sub-data including: a first text and a second voice; Inputting the second speech in the single-round conversation sub-data into the speech preprocessing unit: obtaining the second speech duration attribute of the second speech by speech duration extraction; By removing the environmental noise of the second voice through voice denoising, high-quality second voice information is obtained to improve voice usability; The obtained high-quality second voice information is input into a text extraction module to obtain second text information corresponding to the second voice.
4. The method for recognizing a single voice intention in a remote call scenario according to claim 3, characterized in that: The extracting of feature vectors of each modal information; the extracting respectively includes the steps of speech feature extraction, speech text feature extraction, text feature extraction, and speech duration feature extraction, specifically including: Performing text feature extraction, voice feature extraction, voice text feature extraction, and voice duration feature extraction on the first text, the second voice information, the second text, and the second voice duration attribute, respectively, to obtain corresponding intermediate feature vectors; The intermediate feature vectors respectively include: a first text feature vector, a second voice feature vector, a second text feature vector, and a second voice duration attribute feature vector.
5. The method for recognizing a single voice intention in a remote call scenario according to claim 4, characterized in that: The steps of performing single-round fusion of multimodal feature vectors and multi-round fusion of feature vectors specifically include: Simultaneously performing a single-round multimodal feature fusion on the plurality of intermediate feature vectors, deeply fusing the feature vectors of each modality through a first feature concatenation network and a first attention network, and obtaining the corresponding single-round multimodal feature vector after dimensionality reduction through a first dimensionality reduction network; According to the time sequence of the single-round dialogue sub-data in the multi-round dialogue data and based on multi-step cyclic training, a number of single-round multimodal feature vectors arranged according to the time sequence are obtained; One or several single-round multimodal feature vectors arranged in time sequence are simultaneously subjected to multi-round feature fusion, each single-round multimodal feature vector is deeply fused through the second feature splicing network and the second attention network, and the reduced single-intention feature vector is obtained through the second dimensionality reduction network.
6. A voice single intention recognition device for a remote call scenario, characterized in that: Include: A speech preprocessing module is used to perform speech preprocessing, speech denoising and text extraction, wherein the speech preprocessing specifically includes speech duration extraction, speech denoising and text extraction performed in sequence; A multi-feature extraction module is used to extract feature vectors of each modal information; the extraction includes speech feature extraction, speech text feature extraction, text feature extraction, and speech duration feature extraction; A feature fusion module is used to perform single-round multi-modal feature vector fusion and multi-round feature vector fusion; The single intention recognition module is used to identify the probability of a single target intention tendency. The specific output result is expressed as the probability possibility of being regarded as a single target.
7. The device for recognizing a single voice intention in a remote call scenario according to claim 6, characterized in that: The feature fusion module sequentially comprises a single-round multimodal feature fusion unit and a multi-round feature fusion unit; The single-round multimodal feature fusion module sequentially includes a first feature splicing network, a first attention network, and a first dimensionality reduction network; the feature fusion module sequentially includes a second feature splicing network, a second attention network, and a second dimensionality reduction network.
8. The device for recognizing a single voice intention in a remote call scenario according to claim 7, characterized in that: The speech preprocessing module comprises: A data acquisition unit is used to acquire multi-round conversation data in an intelligent paging scenario, wherein the multi-round conversation data is composed of one or more single-round conversation sub-data arranged in time sequence, and the single-round conversation sub-data includes: a first text and a second voice; The preprocessing unit is used to input the second speech in the single-round dialogue sub-data into the speech preprocessing unit: obtain the second speech duration attribute of the second speech by speech duration extraction; A denoising processing unit, configured to remove environmental noise of the second voice through voice denoising, and obtain high-quality second voice information to improve voice usability; The text processing unit is used to input the acquired high-quality second voice information into the text extraction module to acquire the second text information corresponding to the second voice.
9. The device for recognizing a single voice intention in a remote call scenario according to claim 8, characterized in that: The multi-feature extraction module comprises: The sequential processing extraction unit is used to perform text feature extraction, voice feature extraction, voice-text feature extraction, and voice duration feature extraction on the first text, the second voice information, the second text, and the second voice duration attribute, respectively, to obtain corresponding intermediate feature vectors; the intermediate feature vectors respectively correspond to: the first text feature vector, the second voice feature vector, the second text feature vector, and the second voice duration attribute feature vector.
10. The device for recognizing a single voice intention in a remote call scenario according to claim 9, characterized in that: The single intent recognition module comprises: A feature fusion unit, used to simultaneously perform a single-round multimodal feature fusion on the plurality of intermediate feature vectors, deeply fuse the feature vectors of each modality through a first feature concatenation network and a first attention network, and obtain the corresponding single-round multimodal feature vector after dimensionality reduction through a first dimensionality reduction network; A cyclic training unit, for obtaining a plurality of single-round multimodal feature vectors arranged in time sequence based on the time sequence of single-round dialogue sub-data in the multi-round dialogue data and based on multi-step cyclic training; The fusion processing unit is used to simultaneously perform multi-round feature fusion on one or several single-round multimodal feature vectors arranged in time sequence, deeply fuse each single-round multimodal feature vector through the second feature splicing network and the second attention network, and obtain the reduced single-intention feature vector through the second dimensionality reduction network.