Method and device for analyzing authenticity based on environmental sound features, equipment and medium

Abstract

The present application relates to the technical field of speech processing, and can be applied to business scenarios such as financial technology and medical health, and discloses a realness analysis method and device based on environmental sound features, equipment and a medium, comprising: obtaining original speech data containing environmental sound, performing speech separation processing on the original speech data to generate environmental sound data and pure speech data, analyzing the environmental sound data and retrieving associated knowledge information, identifying the content of the pure speech data to generate dialogue text data, inputting the environmental sound features, knowledge information, dialogue text data and user declaration information into an analysis model, and outputting a realness analysis result. The present application can improve the accuracy of realness judgment and the adaptability to complex scenarios by separating the environmental sound data and the speech content and extracting their available features respectively, and combining background knowledge and user declaration information for fusion reasoning in a unified analysis model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speech processing, and in particular to a realness analysis method and device based on environmental sound features, equipment and storage medium. BACKGROUND

[0002] In the existing position verification related technology, although some schemes have tried to introduce environmental information as auxiliary features, most of these methods are still at the shallow level of feature extraction and matching stage, and lack the deep structured understanding and cross-modal correlation analysis capability of environmental data. In practical applications, environmental sounds often have high diversity and heterogeneity, such as bird songs, dialects, traffic alarm sounds, etc. in the background, which may have mutual contradictions or ambiguous directions. The existing technology is difficult to cope with such complexity. Especially when there are multiple conflicting environmental signals, the existing method lacks a systematic conflict detection and explanation mechanism, and usually cannot make accurate judgments.

[0003] In the field of financial technology business, remote account opening, online loan application, identity verification and other scenarios are increasingly popular. After the user submits the declaration information through the voice channel, the platform needs to determine whether the user is truly at the claimed location to avoid fraud caused by location forgery. However, the current mainstream technology still mainly relies on IP address, device positioning or basic features of voice signals, and lacks effective use of environmental sounds contained in the call process, especially cannot make credible inferences combining voice content and background information, resulting in limited fraud identification capability.

[0004] In the field of medical health business, in scenarios such as remote consultation, voice recording of medical history or online emergency dispatch, if the location information stated by the patient or the person seeking help does not match the real environment, it may cause delay in dispatch or mismatch of resources. The existing technology mainly relies on speech recognition and general positioning methods, and has not yet built an effective mechanism to use environmental acoustic signals to assist in determining the real location of the patient. In particular, in the absence of GPS or high-precision positioning means, the lack of judgment ability is particularly prominent in emergency scenarios.

[0005] In summary, the existing technology has significant deficiencies in multi-source environmental information fusion analysis, conflict information analysis and spatial consistency inference, and is difficult to meet the intelligent verification needs of application scenarios such as financial risk control and medical dispatch that require high location authenticity. SUMMARY

[0006] The main purpose of the present application is to provide a realness analysis method and device based on environmental sound features, equipment and storage medium, aiming to solve the technical problems that the existing verification method relies on a single data source and lacks collaborative analysis of environmental information and semantic logic, resulting in the inability to effectively identify the contradiction between acoustic environment and semantic content in the fake scenario, and the weak anti-fraud ability and low verification accuracy.

[0007] To achieve the above object, the present application provides a method for analyzing authenticity based on environmental sound features, comprising:

[0008] obtaining original speech data containing environmental sound;

[0009] performing speech separation processing on the original speech data to generate environmental sound data and clean speech data;

[0010] analyzing the environmental sound data, extracting environmental sound features, and retrieving knowledge information associated with the environmental sound features;

[0011] performing speech content recognition on the clean speech data to generate dialogue text data;

[0012] inputting the environmental sound features, knowledge information, dialogue text data, and user declaration information into an analysis model to output authenticity analysis results.

[0013] Further, to achieve the above object, the present application provides a device for analyzing authenticity based on environmental sound features, comprising:

[0014] a speech collection module for obtaining original speech data containing environmental sound;

[0015] a speech separation module for performing speech separation processing on the original speech data to generate environmental sound data and clean speech data;

[0016] an environmental perception analysis module for analyzing the environmental sound data, extracting environmental sound features, and retrieving knowledge information associated with the environmental sound features;

[0017] a speech recognition module for performing speech content recognition on the clean speech data to generate dialogue text data;

[0018] an authenticity analysis module for inputting the environmental sound features, knowledge information, dialogue text data, and user declaration information into an analysis model to output authenticity analysis results.

[0019] Further, to achieve the above object, the present application also provides a computer device comprising a memory, a processor, and a program for analyzing authenticity based on environmental sound features stored on the memory and executable on the processor, which, when executed by the processor, implements the steps of the method for analyzing authenticity based on environmental sound features as described above.

[0020] Further, to achieve the above object, the present application also provides a computer readable storage medium, wherein the storage medium stores a realness analysis program based on environmental sound features, and the realness analysis program based on environmental sound features, when executed by a processor, implements the steps of the realness analysis method based on environmental sound features.

[0021] Beneficial effects: The present application relates to the technical field of speech processing, and can be applied to business scenarios such as financial technology and medical health. The present application discloses a realness analysis method and device based on environmental sound features, equipment and medium, which comprises the following steps: obtaining original speech data containing environmental sound; performing speech separation on the original speech data to generate environmental sound data and pure speech data; analyzing the environmental sound data to generate environmental sound features and retrieving knowledge information associated with the environmental sound features; performing speech content recognition on the pure speech data to generate dialogue text data; inputting the environmental sound features, knowledge information, dialogue text data and user declaration information into an analysis model to generate a realness judgment result. The present application separates the environmental sound and pure speech in the original speech data, extracts environmental features and semantic content respectively, and performs unified modeling and analysis in combination with knowledge information and user declaration information. In the case of multiple source heterogeneous environmental information, reasoning can be performed based on dimensions such as semantics, geography and time, so as to realize accurate judgment of the realness of the speaker declaration and improve the discrimination accuracy and robustness in complex scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0022] The present application will be further described below in combination with the drawings and embodiments. In the drawings:

[0023] Figure 1 An application environment schematic diagram of the realness analysis method based on environmental sound features in an embodiment of the present application;

[0024] Figure 2 A flow schematic diagram of the realness analysis method based on environmental sound features in an embodiment of the present application;

[0025] Figure 3 A functional module schematic diagram of the realness analysis device based on environmental sound features in a preferred embodiment of the present application;

[0026] Figure 4 A structure schematic diagram of a computer device in an embodiment of the present application;

[0027] Figure 5 Another structure schematic diagram of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0028] It should be understood that the specific embodiments described herein are merely intended to explain the present application and are not intended to limit the present application.

[0029] The method for analyzing authenticity based on environmental sound features provided by the embodiments of the present application can be applied in application environments such as Figure 1 , wherein a user terminal communicates with a server through a network. The server can obtain original speech data containing environmental sound through the user terminal, perform speech separation on the original speech data to generate environmental sound data and clean speech data, analyze the environmental sound data to generate environmental sound features and retrieve knowledge information associated with the environmental sound features, perform speech content recognition on the clean speech data to generate dialogue text data, and input the environmental sound features, the knowledge information, the dialogue text data and user declaration information into an analysis model to generate an authenticity judgment result. By separating the environmental sound and the clean speech in the original speech data, extracting environmental features and semantic content respectively, and performing unified modeling and analysis in combination with knowledge information and user declaration information, the present application can make reasoning based on dimensions such as semantics, geography and time in the presence of multi-source heterogeneous environmental information, so as to realize accurate judgment of speaker declaration authenticity and improve the discrimination accuracy and robustness in complex scenarios. The user terminal can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices. The server can be implemented by an independent server or a server cluster composed of multiple servers. The present application will be described in detail below through specific embodiments.

[0030] Please refer to Figure 2 , Figure 2 The flowchart of an embodiment of the method for analyzing authenticity based on environmental sound features provided by the present application is shown in the figure. It should be noted that although a logical sequence is shown in the flowchart, the steps shown or described can be executed in an order different from that herein in some cases.

[0031] As shown in Figure 2 , the method for analyzing authenticity based on environmental sound features provided by the present application includes the following steps:

[0032] S10, obtaining original speech data containing environmental sound;

[0033] In this embodiment, the operation of obtaining original speech data containing environmental sound is essentially to collect an audio signal with speech and environmental background mixed characteristics. In this process, the target is not only to collect speech content, but also to consciously retain auxiliary sound source information naturally existing in the environment. These environmental sounds can include background conversations, distant sirens, traffic noise, animal sounds commonly heard in certain areas, construction equipment operation sounds, etc. The introduction of such sounds is not error or redundant information in noise filtering processing, but an important part of the data structure for subsequent analysis.

[0034] In the implementation of this operation, attention should be paid to the wide frequency response coverage of the collected signal to ensure that different characteristic sounds such as high-frequency sounds such as alarm sounds, bird songs, medium-frequency sounds such as human voices, and low-frequency sounds such as mechanical equipment vibration sounds can be effectively recorded. In addition, the acquisition of this audio data should preserve the time continuity and intensity dynamics, and should not smooth or excessively denoise the waveform features in the audio that may carry indicative significance.

[0035] Due to the regional, temporal, and probabilistic characteristics of environmental sound, the collection process should not interfere or change the natural sound field as much as possible. For example, spaces with strong interference sources or background music should be avoided to avoid masking the spatial distribution structure of natural environmental sound. In some scenarios, it is also possible to optimize the completeness and clarity of environmental sound by setting physical structure parameters such as recording distance or sound collection direction, thereby providing more auxiliary information for subsequent semantic recognition and regional feature judgment.

[0036] The acquired audio data should be original, i.e. waveform data saved directly without semantic extraction, sound source separation, content filtering, etc. It not only carries speech content, but also provides a basis for expressing time continuity, sound field integrity, and unprocessed information, which is an indispensable raw input in multi-level analysis path.

[0037] Sound acquisition logic rules can be set to enhance the complete recording of environmental sound, such as prioritizing data acquisition in open or semi-open space scenarios to improve the probability of natural sound entering. Audio dynamic range control mechanisms can also be configured to store only when the overall sound pressure level is within a threshold range to avoid inapplicable data in extreme silence or extreme noise conditions. Time tags, location tags, or environmental classification tags can be added before or after acquisition to enable the audio data to carry more rich context information when sent to subsequent speech separation, content recognition, and authenticity inference. For different environmental backgrounds, the collection duration can also be adjusted to ensure that typical background sound events have a recordable time window, such as extending the collection duration to cover periodic sound signals such as timed broadcasts or fixed-point prompt sounds. In some cases, concurrent acquisition mechanisms can also be used to record multiple sound source directions in the same scene simultaneously to expand the sound source range and sound field structure depth of environmental features.

[0038] Example: In the medical health business, the user submits a description of the illness through a remote terminal. The system records the environment sound at the same time as receiving the voice input. For example, the audio naturally contains elevator voice prompts, background medical device prompt sounds, and medical staff conversation sounds, which can be used as key evidence when judging whether the user is in a hospital environment.

[0039] In the financial technology business scenario, when a user completes remote account opening or identity verification and uploads voice data, the system can assist in determining whether the statement is consistent with the actual environment by analyzing whether there are sound features with regional orientation (such as specific dialects and trading hall broadcasts) in the original audio. In this way, environmental sound is no longer an interference item in the recognition task, but an important element of the overall trusted analysis system.

[0040] The embodiment retains the environmental sound component in the original voice data, which not only provides context for subsequent voice recognition and semantic understanding, but also provides auxiliary judgment conditions for more complex region recognition, behavior verification, or scene matching. This data acquisition method retains physical parameters such as time structure, spectral composition, and sound field structure, allowing downstream analysis modules to fully exploit non-semantic dimension information in the audio and perform spatial positioning and authenticity judgment based on these structured or unstructured signals, thereby improving the generalization ability and abnormal recognition ability of the entire process in complex situations.

[0041] S20, performing voice separation processing on the original voice data to generate environmental sound data and pure voice data;

[0042] In this embodiment, the original voice data is processed by voice separation, which aims to effectively distinguish and reconstruct the parts of the recorded composite audio signal that have different sound source characteristics. The original voice data is usually composed of multiple sounds, including the main voice target (i.e., the speaker's voice) and the accompanying environmental sound components (such as traffic noise, natural sound, or other people's conversation, etc.). These sounds have significant differences in frequency spectrum, energy, and timing structure, so they can be separated by a computational model.

[0043] The processing method generally includes four stages: sound source modeling, feature extraction, mask generation, and signal reconstruction. First, the original voice data is input into the acoustic modeling framework, and a neural network model with voice separation capability is used to learn the internal sound source representation of the mixed signal under unsupervised or semi-supervised mechanism. This representation process takes the time-frequency graph of the original voice data as input, and extracts time-frequency domain representations containing voice dynamics, formant structure, background noise texture, etc. through hierarchical convolution structure.

[0044] After completing the representation extraction, the sound source features that may belong to the background environment can be clustered, labeled, or enhanced, and then a sound source mask matrix is constructed to indicate the signal distribution of the main speaker or environmental sound in a specific frequency band. The mask matrix can be binary, used to forcibly block non-target frequency signals, or probabilistic, used to represent the degree of mixing and signal proportion.

[0045] Next, the frequency domain representation of the original speech data is operated with the generated mask matrix on a frequency-by-frequency basis to reconstruct two audio signals representing different sound source channels. Finally, inverse Fourier transform or inverse short-time Fourier transform is performed on these frequency domain signals to generate time domain audio signals, which are output as environmental sound data and clean speech data, respectively.

[0046] During the separation process, the integrity of the speech boundary and the continuity of the environmental structure should be maintained to prevent speech disruption or excessive background cleaning caused by mask operation. It is also necessary to ensure that the separation model has the generalization ability to unknown environmental sounds to adapt to speech input data from different devices, scenes or areas.

[0047] Structured neural network models can be used to implement speech separation, such as using a convolutional time domain network-based separation model (such as Conv-TasNet) to learn the mapping relationship between mixed speech and separation targets through end-to-end training. Time-frequency domain methods based on mask learning can also be used, such as inputting the original speech signal into a deep neural network (such as Dual-Path RNN) after short-time Fourier transform to extract the soft mask matrix corresponding to the sound source, and then reconstructing the separated audio track by inverse transform. Attention mechanisms and sound source representation clustering strategies can also be combined to model the changing background sound field in speech, further improving the separation model's ability to distinguish complex background sounds.

[0048] In the model input design, background silence segments, spectral contrast segments or known background samples synchronized with data collection can be added to help the model establish the boundaries of sound source features. In the output stage, energy constraint balancing strategies can be applied to the separated audio tracks to avoid complete truncation of weak environmental sounds or marginal speech.

[0049] Example: In the medical health scene, remote voice communication between doctors and patients is often accompanied by background sounds such as ward announcements, instrument prompt sounds or other patient conversation sounds. Through speech separation processing, the doctor and patient conversation content can be clearly extracted for semantic analysis, while the environmental sound is independently preserved as an auxiliary clue to determine whether the communication actually occurred in a medical environment.

[0050] In the financial technology business, customers often upload audio during remote identity verification, accompanied by counter environment, crowd noise or background broadcast sound. Speech separation can preserve high-quality customer speech for system recognition, and the independently separated environmental sound can be used as auxiliary information to verify whether the speech was collected in a specific financial service site. Through speech signal separation processing, the system's ability to interpret complex input data and decision reliability is further enhanced.

[0051] The embodiment can effectively distinguish the environmental sound interlaced in the original voice data from the pure voice signal, on the one hand, can provide clearer and less disturbed main voice input for subsequent voice recognition, and improve the text generation accuracy; on the other hand, the separated environmental sound data can be independently used for sound source recognition, regional feature analysis or space background modeling, and enhance the understanding ability of the whole analysis process to the speaking environment. Through the processing mechanism, not only the semantic and context are decoupled, but also the depth and accuracy of multi-dimensional utilization of input audio content are expanded.

[0052] S30, analyzing the environmental sound data, extracting environmental sound features, and retrieving knowledge information associated with the environmental sound features;

[0053] In the embodiment, the environmental sound data is analyzed, the goal is to extract environmental sound features with geographical, semantic or scene recognition value, and further access external knowledge resources associated with the features. The environmental sound data usually contains audio patterns related to specific space-time positions, such as bird chirping, crowd dialect, vehicle horn, mechanical operation, etc., which have quantifiable acoustic structures and can be used to build reasoning channels between sound and place, species or event.

[0054] First, the acoustic event detection operation is performed on the environmental sound data, that is, the acoustic segments containing independent meanings are identified and divided in the time dimension. This process usually combines short-time energy, spectral change, signal-to-noise ratio and other indicators, and can also use a trained acoustic event detection model to extract a candidate acoustic event set. Each acoustic event contains start time, end time, frequency structure and energy parameters and other elements.

[0055] Then, the acoustic events are classified into multiple categories. Part of the acoustic events belong to biological acoustic events, which can be identified by species classification model to identify corresponding bird, insect or animal species. The model usually uses convolutional neural network or transfer learning method, and introduces global or regional bird chirping dataset in the training process to identify specific species labels, regional characteristics, occurrence frequency and other indicators. Another part of the acoustic events belong to non-biological acoustic events, such as ambulance siren, subway passing, school broadcast, etc., which can be classified and identified by voiceprint template matching, MFCC feature clustering, time sequence alignment and other methods.

[0056] The above classification results are integrated into a unified set of environmental sound features. This set retains information such as sound source type, classification confidence, and time location in a data structure, and can be used as structured input for downstream analysis. Based on the feature set, external retrieval is further performed through a knowledge graph or embedded index system. Knowledge information usually includes established entity relationships such as sound source-place, sound source-time, and sound source-ecological chain. For example, if the detected bird chirping signal matches "white-headed pheasant", the relevant knowledge such as its high activity in urban green belts, specific latitude range, and early morning activity can be retrieved; if the "ambulance siren" is detected, the corresponding information such as the dense area of urban hospitals, specific road types, and time period rules can be retrieved.

[0057] Finally, the generated environmental sound features and their associated knowledge information form the basis for understanding the structure of the current audio environment, providing interpretable support for subsequent comprehensive judgment.

[0058] Acoustic event recognition can be performed in a multi-model collaborative manner, for example, a YAMNet model is used in the first stage to perform coarse-grained event detection on environmental sound data, and preliminary acoustic segments are extracted; in the second stage, a bird classification model and a non-biological event soundprint matching model are called respectively to perform fine-grained classification operation on the candidate segments. For biological acoustic event classification, a sound recognition model with pre-training capability can be trained based on bird sound resources with rich species annotation and geographical distribution characteristics, so that the model can further enhance the ability to distinguish regional species distribution characteristics based on the identification of different types of birds in the audio, thereby achieving more accurate biological acoustic event classification and feature extraction in different geographical backgrounds. In non-biological sound recognition, multiple recognizers based on MFCC templates can also be constructed, respectively for traffic, construction, and campus in typical urban environments.

[0059] Knowledge information retrieval can rely on a locally constructed environmental sound knowledge graph, which uses RDF format or embedded vector library to implement time distribution, spatial distribution, and ecological chain information indexing for each type of sound source. A pre-trained language model can also be used to perform a RAG (retrieval-augmented generation) process with the sound source label as input to obtain reliable associated knowledge from literature and city information library, and output structured spatial distribution data.

[0060] Example: In the medical health scene, a remote diagnosis and treatment platform can identify hospital background broadcast sound, call bell prompt sound, and elevator arrival sound in the patient uploaded voice as environmental sound features, and then combine the sound source distribution information of the specific environment of the hospital in the knowledge base to assist in judging whether the voice collection actually occurred in a medical place, thereby improving the credibility of remote medical interaction.

[0061] In a financial service scenario, when a customer uploads voice content through a mobile terminal, the background system can detect the dialect language, bank call system voice prompts, printer working sound, etc. in the background, compare them with known financial business site environment characteristics, combine financial institution geographic layout knowledge, improve the discrimination ability of the authenticity of the customer's behavior location, and help prevent remote fraud and identity fraud risks.

[0062] Through systematic analysis of environmental sound data, the embodiment can mine label information with spatiotemporal significance from the audio segment, and then combine structured knowledge resources to conduct in-depth analysis on the consistency between the environment and the speaker declaration. Compared with the single sound source detection or the method of only doing spectral analysis, this processing method can detect and interpret multiple types of acoustic events, effectively improve the level of understanding of the audio environment background, and enhance the reliability of the decision basis of the downstream judgment module.

[0063] S40, performing voice content recognition on the pure voice data to generate dialogue text data;

[0064] In this embodiment, the pure voice data is taken as input, which needs to be frame-level segmented and window function weighted first to form a set of continuous but limited time length audio frames. This process can be completed by using a fixed length sliding window, for example, setting the frame length to 25 milliseconds and the frame shift to 10 milliseconds, and applying a window function on each frame to reduce the spectral leakage effect, so that the signal has better locality in frequency domain analysis. Feature vectors reflecting acoustic properties are extracted on each frame, one of the commonly used representations is the mel frequency cepstral coefficient, which is obtained by performing short-time Fourier transform, mel filter bank processing and logarithmic compression on the voice signal, and then performing discrete cosine transform extraction, which can effectively simulate the distribution characteristics of human ear perception spectrum. These feature vectors are sent to the speech recognition model as input for modeling.

[0065] The speech recognition model can use a multi-layer acoustic modeling network to model the context relationship in the acoustic feature sequence through a time modeling structure such as a convolutional neural network or a structure with an attention mechanism. The model outputs a probability distribution sequence of phonemes or sub-word units, which can be generated into preliminary text content through decoding algorithms such as optimal path search based on conditional random field or connection time sequence classification (CTC) mechanism. In order to improve the readability and semantic integrity of the text content, the initial output needs to be input into a semantic layer language model for correction. The language model can capture syntax structure, semantic association and context coherence, realize automatic error correction, syntax completion and format standardization, and finally form dialogue text data with complete syntax and clear meaning.

[0066] In implementation, a deep model based on attention mechanism can be used to model the clean speech data, such as inputting frame-level features into a bidirectional encoder network, and combining position encoding and attention layer to reweight the context information. Phoneme decoding can combine language priori and acoustic likelihood for joint decoding to improve the robustness of the results. The generation of dialogue text can use an end-to-end speech recognition architecture to embed acoustic modeling and language model into a unified network, and simultaneously optimize recognition accuracy and language consistency through multi-task training. A post-language model can also be used to correct the preliminary recognized text, and a context-based neural network generator can be used to identify timing errors, improper word forms, and syntax defects.

[0067] In scenarios where different languages or pronunciation differences are significant, model transfer learning or parameter tuning training can be performed to be compatible with multiple speech input features. For example, fine-tuning speech data with accent variants, specific dialects, and mixed languages enables the model to have good recognition ability for phoneme boundary ambiguity and different syntax structures.

[0068] Example: In the medical health business scenario, users make remote condition statements through mobile terminals, and the speech is often accompanied by environmental noise and accompanying personnel interference. The system first removes the background sound from the overall speech data and only retains the user's subjective expression content as clean speech data, and then performs speech content recognition. This step can accurately generate dialogue text from expressions with accents or mixed medical terms through sequence modeling and language structure repair of clean speech data, so that the patient's description of the symptoms can be standardized, providing high-quality structured input for doctors or inquiry systems.

[0069] In the financial technology business scenario, users submit claim statements or location verification information through the voice channel, and the system performs speech recognition processing after extracting clean speech data. This step can accurately identify key content such as time, location, and event process in user pronunciation, especially in the context of dialects, uneven speech rates, or disordered information expression sequences, and still restore semantic order and complete missing words through language modeling, thereby outputting dialogue text data for subsequent verification and analysis, supporting consistency verification of statement content and background information in complex scenarios.

[0070] This embodiment can accurately extract the speech content by performing multi-stage processing on clean speech data, from low-level acoustic feature extraction to high-level semantic calibration through language model correction, effectively avoiding recognition bias caused by environmental interference, and improving the accuracy and semantic expression consistency of dialogue text. Combined with time modeling structure and language model, the adaptability and generalization ability of the speech recognition system in natural context can be significantly enhanced.

[0071] S50, input the environmental sound features, knowledge information, dialogue text data and user declaration information into an analysis model, and output a authenticity analysis result.

[0072] In the embodiment, the operation of integrating the environmental sound features, knowledge information, dialogue text data and user declaration information into the analysis model aims to complete deep joint modeling of multi-source data to realize reasoning analysis of potential consistency or contradictory relationship between information. The environmental sound features are generally composed of acoustic label information extracted in the previous step, such as a set of multi-dimensional vectors representing a specific biological species or a non-biological event, which usually has regional distribution sensitivity and time occurrence regularity. The knowledge information is derived from structured data corresponding to the above-mentioned environmental features, which can specifically represent the corresponding probability of a certain sound and a geographical location or time period, a known appearing area or a limitation condition, etc. The information is mapped into a semantic graph or a knowledge vector through a pre-trained embedding method.

[0073] The dialogue text data is derived from the pure speech content recognition step, contains the content expressed by the user subjectively, and is semantically represented through language understanding mechanism to form a semantic vector reflecting the space-time direction, behavior description or event state. The user declaration information is the subjective declaration submitted by the user related to the task target, which can include geographical location, time point or identity attribute, etc. The processing method needs to embed the information into a structured and readable numerical representation after standardization, which is usually completed by using a geographic coordinate embedder, a position coding network or a space-time index model.

[0074] The above four types of data need to be processed by time alignment and space normalization before input to ensure their consistency in representation scale and analysis interval. The processing method can include using sliding window technology to synchronize event granularity, using geographic coordinate mapping or regional label to unify the space reference system, or adjusting the corresponding relationship between sentences and sound events based on semantic annotation. The aligned data constitute a unified input set, in which each sub-vector represents environmental features, knowledge semantics, text semantics and declaration coordinates.

[0075] After the set is input into the analysis model, the correlation matrix between different modalities is first calculated through the multi-modal attention mechanism to extract suspicious conflict or highly consistent feature regions, and then integrated into a unified multi-modal representation vector through the fusion module. In the reasoning stage, the vector passes through the structured reasoning unit containing hierarchical logic judgment path to evaluate whether there is a semantic / geographical / time conflict between information or whether it shows a high degree of synergistic consistency. The final model outputs the authenticity analysis result, which is represented as a structured label (such as true or false, credibility score) and an explainable information supporting the logical reason for the result.

[0076] In different implementations, the multi-modal fusion method can be selected according to actual system resources and application scenarios. A soft fusion strategy guided by attention can be used to integrate information from different sources through a weighting method, or a graph neural network structure can be constructed to propagate information based on node relationships by constructing a feature node graph. The encoding method of user declaration information can also be adjusted as needed, for example, using city label embedding in city-level judgment, and using latitude and longitude encoding and aligning with geographic nodes in the knowledge graph in fine location scenarios. The processing of dialogue text data can introduce a language context enhancement mechanism, such as introducing a historical dialogue window or an external time axis structure to assist in identifying the subject's spatio-temporal attributes. The vectorization of environmental sound features can be based on a label set frequency histogram, event duration vector, or event confidence distribution graph, and different encoding methods will affect the subsequent semantic alignment effect.

[0077] Example: In a financial technology business, a user submits an address information through voice to assist risk verification, and the system extracts a bird chirp mark specific to a Southeast Asian region from the background sound features in the user's surrounding environment. At the same time, the user indicates that he is currently located in the European region during the conversation, and the submitted declaration location also points to a European city. The system aligns the inputs in time and space, finds that there is a significant spatial contradiction between the bird sound source and the European declaration location, and after combining semantic analysis, the system model outputs a low credibility result for the authenticity analysis, with a contradiction analysis explanation.

[0078] In a medical health business scenario, a remote consultation user claims to be at home resting, while expressing "it has been very noisy nearby, with a fire truck passing by" in the conversation. After analyzing the pure voice content and the background environment, the system identifies the repeated occurrence of a fire truck siren sound, and retrieves from the knowledge information that the user's declared location has no fire alarm record within the past three hours. Through multi-modal fusion analysis, the system determines that there is a contradiction between the description in the user's voice and the actual declaration, and outputs an inconsistency result for the authenticity analysis and points out the reason for the time anomaly. This mechanism provides auxiliary support for environmental consistency and user state authenticity judgment in remote consultation.

[0079] This embodiment can comprehensively analyze the spatio-temporal or semantic conflicts implied in different modalities by jointly modeling environmental sound features, knowledge information, dialogue text data, and user declaration information. This not only allows the identification of explicit inconsistent behaviors (such as contradictions between declared regions and species distribution), but also captures fine-grained reasoning relationships between cross-modalities (such as inconsistencies between declared time and ambulance sound occurrence period). This mechanism enables the system to derive complex relationships from multi-dimensional clues, significantly improving the accuracy and robustness of authenticity determination, and is particularly suitable for handling information environments with ambiguous, contradictory, and complex semantics in the real world.

[0080] The application relates to the technical field of voice processing, can be applied to business scenes such as financial technology and medical health, and discloses a realness analysis method and device based on environmental sound features, equipment and a medium, which comprises the following steps: obtaining original voice data containing environmental sound; performing voice separation on the original voice data to generate environmental sound data and pure voice data; analyzing the environmental sound data to generate environmental sound features and retrieving knowledge information associated with the environmental sound features; performing voice content recognition on the pure voice data to generate dialogue text data; and inputting the environmental sound features, the knowledge information, the dialogue text data and user declaration information into an analysis model to generate a realness judgment result. The environmental sound and the pure voice in the original voice data are separated for processing, environmental features and semantic content are extracted respectively, and unified modeling and analysis are performed in combination with knowledge information and user declaration information, so that reasoning can be performed based on dimensions such as semantics, geography and time under the condition of multiple source heterogeneous environmental information, the accuracy of judgment of the speaker declaration realness is improved, and the discrimination accuracy and robustness in a complex scene are improved.

[0081] In one embodiment, the step S10 comprises:

[0082] S101, collecting original voice signals containing environmental sound through a microphone module of a terminal device;

[0083] S102, performing environmental sound frequency band selective gain processing on the original voice signals to generate original voice data containing enhanced environmental sound;

[0084] S103, converting the original voice data into a standardized audio format file;

[0085] S104, storing the standardized audio format file into a storage area and recording the collection time and the collection device identifier of the original voice data.

[0086] In the embodiment, the original voice data containing environmental sound is collected, so as to construct a real and multi-dimensional acoustic input source and ensure that the environmental elements have sufficient expression basis in subsequent voice analysis. This process can be divided into a plurality of continuous and associated processing links.

[0087] Firstly, capturing original voice signals through a terminal device is a prerequisite for obtaining sound information, and the sound signals should have the characteristics of wide coverage, real noise structure and obvious environmental difference. During collection, the terminal device is only required to have audio collection capability and be capable of synchronously capturing human voice and environmental background sound. During the collection process, the original voice signals contain voice content of a user and surrounding natural environmental sound such as traffic noise, crowd conversation, bird chirping and mechanical sound, which embodies the mixed characteristics of voice data in a real scene.

[0088] To improve the recognizability of environmental elements in subsequent analysis, the above-mentioned signal needs to be subjected to frequency band selective gain processing. Gain processing refers to implementing amplitude enhancement on the frequency region containing environmental sound characteristics under the premise of preserving the overall speech structure. Environmental sounds are usually concentrated in non-linguistic main frequency band regions, such as mechanical sounds in the low frequency band, bird songs or ambulance alarms in the high frequency band. The processing method can adjust the spectrum of the original signal by dynamic spectrum separation algorithm, so that the environmental information is more prominent in the spectral energy, and the perception ability in the subsequent separation and identification steps is enhanced. Such gain processing emphasizes the preservation of acoustic structure, does not introduce new speech content or artificial sound sources, and only purposefully adjusts the signal energy distribution.

[0089] The processed audio data needs to be converted into a unified standard to ensure that data under different devices, different collection formats or sampling frequencies has consistent technical expression structure. The conversion process includes operations such as sampling rate unification (such as 48 kHz), bit depth adjustment (such as 16 bit), encoding format standardization (such as WAV, FLAC), etc., so that the audio data adapts to the subsequent signal processing and model input requirements. This conversion also provides a consistency basis for the cross comparison and feature mining of environmental sounds.

[0090] The final speech data is stored in a structured audio database, and the collection time and device identification information need to be recorded. Time information is used to support subsequent time window analysis, such as judging whether there is repetition of a specific event within a specific time period; device identification is used to track potential differences in device categories and collection places, supporting post-processing compensation strategies for hardware capability differences. This recording process emphasizes the structural binding of metadata and audio content to ensure that the data has complete source and context when used and reconstructed.

[0091] The embodiment can significantly improve the recognizability and accuracy of environmental elements in subsequent separation and identification by synchronously acquiring original speech data containing human voice and environmental sound and implementing gain processing on the environmental sound characteristic frequency band. The standardized storage mechanism ensures that the audio maintains consistent expression in cross-device and cross-platform scenarios, and the collection time and device identification record provide complete data chain support for subsequent traceability analysis. This mechanism ensures that the input speech data has a comprehensive, stable and multi-dimensional analysis structure, effectively supporting the input quality of the authenticity evaluation in the downstream task.

[0092] In one embodiment, the above step S20 comprises:

[0093] S201, loading a pre-trained deep neural network-based speech separation model;

[0094] S202, extracting time-frequency domain voiceprint features and background noise energy distribution features of the original speech data through the speech separation model;

[0095] S203, generating a frequency band mask matrix according to a frequency band energy ratio value of the time-frequency domain voiceprint feature and the background noise energy distribution feature, and taking the frequency band mask matrix as a sound source separation threshold;

[0096] S204, performing frequency band mask processing on a frequency domain signal of the original speech data based on the sound source separation threshold through the speech separation model;

[0097] S205, performing inverse short-time Fourier transform on the mask-processed frequency domain signal to generate a time domain reconstructed signal;

[0098] S206, separating the time domain reconstructed signal to generate environmental sound data and clean speech data.

[0099] In the embodiment, the purpose of the speech separation processing on the original speech data is to explicitly decouple the contained human voice information from the environmental background sound, thereby facilitating subsequent independent recognition and reasoning of the human voice content and the environmental context. In this process, a neural network model with strong generalization ability needs to be loaded first for multi-channel component analysis of the speech source. The model should be pre-trained on a wide range of audio corpus, so that it can identify the spectral differences between different sound source characteristics and has the ability to learn complex sound source mixing structures. The pre-trained model usually adopts an encoding-mask-decoding structure, including a time-frequency transformation module, a context modeling module, and an unmixing output module, which can process long time sequences and high-dimensional sound spectrum features.

[0100] After the original speech data enters the model, it will be converted into a frequency domain representation, and its time-frequency spectrum distribution will be obtained through short-time Fourier transform. The model first extracts the time-frequency domain voiceprint feature in the frequency spectrum, i.e., the frequency spectrum structure corresponding to the speaker's voice content, which is usually concentrated in the range of 300Hz to 3.4kHz and has relatively continuous energy peak changes. At the same time, the background noise energy distribution feature is extracted, including but not limited to non-verbal signals such as traffic noise, animal calls, crowd noise, and device operation sound, etc., whose spectral characteristics are often significantly different from human voice components in energy fluctuation and frequency band distribution. The extraction process of the two features can be performed through a multi-channel self-attention mechanism to enhance the model's ability to model the spectral position and time variation.

[0101] In order to achieve efficient separation, the system calculates the energy ratio of each frequency band based on the above two types of features, thereby constructing a frequency band mask matrix. The matrix is used to indicate that each spectral unit should belong to human sound source or environmental sound source in the current time window, and the value is usually between 0 and 1, representing the weight or energy retention probability. The generated frequency band mask matrix can be regarded as a dynamic sound source separation threshold, which integrates multiple parameters such as energy contrast between sound sources, spectral shape difference, and time consistency.

[0102] Subsequently, the voice separation model performs frequency band-by-frequency band mask processing on the frequency domain signal of the original voice data according to the mask matrix, that is, the non-human voice component is suppressed in the human voice frequency band area, the human voice signal is weakened in the environmental sound frequency band area, and the signal part that best represents the characteristics of the target sound source is retained. The mask processing does not destroy the original time sequence structure of the signal, but realizes the component reconstruction in the spectral domain. The spectrum data after mask processing is restored to the time domain signal through inverse short-time Fourier transform, and the voice waveform after frequency band selective retention and suppression is obtained, which has higher target source component purity.

[0103] Finally, the sound source separation is performed based on the reconstructed waveform signal, and the two time domain audio streams are further divided by frequency band clustering, phase re-estimation or independent component analysis, etc. One is environmental sound data, which retains background noise and non-semantic sound sources, and the other is pure voice data, which is used for subsequent voice content understanding and intent recognition. The above process forms a complete closed loop in algorithm structure and operation path, ensuring the stability and generalizability of signal processing.

[0104] The embodiment can accurately disassemble the human voice and environmental sound in the original mixed voice into independent signal channels that do not interfere with each other by loading a neural network model with complex sound source modeling capability and constructing a frequency band mask matrix based on time-frequency domain joint features. The use of the mask matrix makes the separation process adjustable and interpretable, while retaining the time structure and subjective listening of the original voice. The final output of the environmental sound data has high-fidelity background information, and the pure voice data provides a clean input signal for semantic recognition, thereby improving the accuracy and robustness of subsequent location judgment, dialogue analysis, environmental semantic reasoning and other processes. The processing mechanism not only enhances the physical feasibility of sound source separation, but also provides structured input conditions for multi-source collaborative reasoning for downstream model construction.

[0105] In one embodiment, the above step S30 comprises:

[0106] S301, acoustic event detection is performed on the environmental sound data to generate an acoustic event set;

[0107] S302, species classification and identification of biological acoustic events in the acoustic event set are performed to generate a biological acoustic feature set;

[0108] S303, voiceprint pattern matching of non-biological acoustic events in the acoustic event set is performed to generate a mechanical acoustic feature set;

[0109] S304, the biological acoustic feature set and the mechanical acoustic feature set are merged into an environmental sound feature set;

[0110] S305, the spatial distribution information associated with the environmental sound feature set is retrieved from the geographic knowledge graph as the knowledge information.

[0111] In this embodiment, in order to extract environmental sound features with geographical orientation and semantic label capability, first, acoustic event detection processing is performed on the environmental sound data. Acoustic event detection refers to dividing a continuous audio stream into event segments containing independent environmental sound components, and labeling the start and end boundaries and preliminary category attributes of each segment in the time domain. The determination of acoustic events can be based on time-frequency energy mutation, spectral shape change, or through a trained supervised model to complete boundary detection and coarse-grained classification, such as sudden chirping, intermittent traffic sound, and continuous background sound. The output of this process is a structured acoustic event set, each event with a timestamp, spectral structure, and preliminary type label, as an input basis for subsequent classification and identification.

[0112] In the acoustic event set, if the event exhibits biological acoustic properties, such as high frequency concentration, obvious syllable structure, and regular time rhythm, it can be classified as a biological acoustic event. For such events, further species classification and identification are performed. The classification and identification process can be completed through a multi-class label identification model, which includes time convolution network, spectral attention mechanism, and other modules, and supports regional corpus transfer training to adapt to common bird or animal species in different geographical environments. The output biological acoustic feature set of this step can include species name label, sound pattern structure representation, and regional co-occurrence probability, providing a basic index for subsequent association with geographical information.

[0113] Events that do not have biological rhythm characteristics are classified as non-biological acoustic events, such as mechanical equipment sound, emergency vehicle siren, and crowd noise. In these events, features are extracted through voiceprint pattern matching. The matching process is based on the similarity calculation between the spectral envelope, harmonic structure of the event, and the voiceprint vector in the preset template or historical record, and the output result is a mechanical acoustic feature set, including event type (such as alarm, train passing), voiceprint matching confidence, and matching source label.

[0114] Subsequently, the biological acoustic feature set and the mechanical acoustic feature set are structurally merged to obtain an environmental sound feature set in a unified format. This set takes events as the basic unit and contains event type label, occurrence time, spectral description, category probability distribution, and other content, forming a high-order audio representation with information density and structural versatility.

[0115] After the environmental sound feature set is generated, the label field (such as a species label, an event type) thereof is used as a retrieval key value to access a structured constructed geographic knowledge graph system. The graph is constructed based on a pre-defined knowledge triple structure, such as "White-headed Laughing Thrush - distributed in - South China", "ambulance siren - high frequency occurrence in - central city", each node represents a geographic entity or environmental event, and each edge represents a semantic relationship. Through matching calculation (such as cosine similarity, embedding vector neighbor search, etc.) between the label vector and the entity nodes of the knowledge graph, spatial distribution information associated with the current environmental sound feature set is obtained to form knowledge information. The spatial distribution information can include not only administrative region or ecological zone names, but also co-occurrence probability and confidence level in a specific time period and specific background.

[0116] The embodiment can accurately extract directional information of sound sources of different sources by deconstructing environmental sound data into specific acoustic events and performing respective adaptive recognition algorithms on biological and non-biological events. Through conversion of the acoustic event set into an environmental sound feature set and further matching with a geographic knowledge graph, structured interpretation of geographic semantics in the recording environment is achieved. This processing process not only improves the interpretation ability of environmental data, but also provides input basis based on geographic knowledge for subsequent position judgment, semantic comparison and multi-modal cross-validation. This mechanism realizes effective mapping to structured semantics while maintaining the non-structured characteristics of environmental data, thereby enhancing the stability and expandability of the system in complex spatial semantic analysis.

[0117] In one embodiment, the above step S40 includes:

[0118] S401, performing frame windowing processing on the pure speech data to generate a frame windowing processed speech signal;

[0119] S402, extracting a mel-frequency cepstral coefficient feature of the frame windowing processed speech signal to generate an acoustic feature vector;

[0120] S403, performing phoneme sequence decoding on the acoustic feature vector by using a pre-trained speech recognition model to generate initial text data;

[0121] S404, performing grammar correction on the initial text data based on a grammar correction model to generate standardized dialogue text data.

[0122] In this embodiment, for the processing of pure speech data, it is first subjected to frame windowing operation. This processing is to divide the continuous speech signal into short time fixed length signal frame, each frame has local stability in time, usually the frame length is set between 20 to 30 milliseconds. In the frame division process, a window function (such as Hamming window or Blackman window) is applied for weighting to reduce the discontinuity of the signal between frames, thereby reducing the spectral leakage phenomenon in the short-time Fourier transform process. The processed speech signal is a plurality of frame structures with time domain envelope constraints, which helps to extract the acoustic features subsequently.

[0123] After obtaining the frame windowing processed speech signal, its acoustic feature vector needs to be extracted. In this process, the Mel frequency cepstral coefficient feature is used as the acoustic representation method. Mel frequency cepstral coefficient (MFCC) has strong robustness in speech recognition field by simulating the nonlinear perception mechanism of human auditory system. The specific implementation includes performing short-time Fourier transform (STFT) on each frame signal, calculating power spectrum diagram, then performing frequency band compression through Mel filter bank, and finally extracting a number of dimensional feature coefficients to form acoustic feature vector through discrete cosine transform (DCT). The feature vector effectively preserves the pronunciation structure information of the speech, which is an important input for the subsequent recognition model.

[0124] Next, the acoustic feature vector is input into the pre-trained speech recognition model for phoneme sequence decoding. The model can adopt an end-to-end structure, such as a deep neural network based on CTC (Connectionist Temporal Classification) or Transducer architecture, which supports mapping of indefinite length feature vector sequence to speech phoneme or sub-word level text sequence. The model can include multiple layers of convolution, recurrent structure, attention mechanism, etc. to extract time correlation and complete context information integration. The output initial text data is the direct transcription result without language rule constraint, which may have spelling, format or sequence errors.

[0125] In order to improve the usability of the recognized text, the initial text data needs to be subjected to grammar correction processing to obtain more structured dialogue content. This process is based on language model, the model structure can be N-gram statistical model, or deep neural language model based on Transformer, such as BERT or GPT variant. Through language model, the initial text is subjected to syntax analysis, part-of-speech repair, context reorganization, etc. processing, and the grammar errors, improper word usage and sequence confusion phenomenon are recognized and corrected, and finally the standardized dialogue text data is output. The text result has the characteristics of strong readability and clear semantics, which is convenient for semantic fusion analysis with environmental information, declaration information and other multi-modal data.

[0126] The embodiment can convert the voice signal into a structured acoustic feature vector by frame windowing and Mel frequency cepstral coefficient extraction of pure voice data, effectively retaining the acoustic characteristics of the voice and eliminating time domain interference. With the help of a speech recognition model, the acoustic feature vector can be phoneme decoded to realize high-precision voice transcription. Meanwhile, the initial text is post-processed by combining a grammar correction model to further optimize the recognition quality and language fluency. The processing procedure realizes the conversion from unstructured voice signal to analyzable text data, provides a stable information foundation for subsequent semantic recognition, context association and fact verification, and has strong adaptability in the task of comparing user claims with actual context.

[0127] In one embodiment, the step S50 comprises:

[0128] S501, performing data alignment processing on the environmental sound features, knowledge information, dialogue text data and user statement information to generate a set of spatio-temporally aligned input data;

[0129] S502, encoding the environmental sound features in the set of spatio-temporally aligned input data into environmental feature vectors;

[0130] S503, encoding the knowledge information in the set of spatio-temporally aligned input data into geographical knowledge vectors;

[0131] S504, encoding the dialogue text data in the set of spatio-temporally aligned input data into semantic vectors;

[0132] S505, encoding the user statement information in the set of spatio-temporally aligned input data into statement position vectors;

[0133] S506, fusing the environmental feature vectors, geographical knowledge vectors, semantic vectors and statement position vectors to generate a multi-modal fusion feature vector;

[0134] S507, inputting the multi-modal fusion feature vector into a pre-trained location anomaly analysis model to output a geographical location authenticity analysis result and an analysis reason.

[0135] In this embodiment, when processing multi-source input information for authenticity analysis, it is necessary to first perform unified data alignment processing on environmental sound features, knowledge information, dialogue text data, and user statement information. The purpose of data alignment is to eliminate the differences of different modal data in time scale and spatial coordinate system, so as to make them comparable and fusible. In the time domain, each data item can be sliced, normalized or interpolated according to a unified time window; in the spatial domain, a unified coordinate mapping method (such as geographic grid coding or administrative block standard) can be used to map geographic information to a consistent spatial reference framework. The spatio-temporal aligned input data set obtained after alignment has structural consistency, providing a processable input format for subsequent encoding.

[0136] When encoding the environmental sound features in the input data set as environmental feature vectors, shallow or deep feature encoding methods can be used. Environmental sound features may include recognized species labels, mechanical sound types, and their occurrence frequencies, etc. They are converted into numerical vector form through embedding layers or statistical encoding models. Such vectors usually have fixed dimensions and preserve the similarity structure between categories, for example, birds can establish a semantic space through hierarchical biological classification relationships, and sound types can be modeled through sound source analogy.

[0137] The encoding of knowledge information is based on a geographic knowledge graph, which contains the mapping relationship between known environmental sound labels and geographic areas, such as a certain type of bird song only appearing in a specific region in the south, and certain mechanical sounds only appearing in a specific traffic environment, etc. During encoding, the spatial distribution information and acoustic labels can be jointly represented by combining graph embedding methods to generate geographic knowledge vectors. This vector can express the geographical bias and its relationship with environmental elements.

[0138] Dialogue text data needs to be further encoded into semantic vectors after alignment, which represent the implicit meaning and context information of the speaker's speech content. Text encoding methods can be based on pre-trained language models, such as using BERT-like models or lightweight sentence embedding networks to extract sentence-level semantic representations. The vector output is usually a high-dimensional space point that can capture information such as location description, scene depiction, and temporal semantics in the text, providing language-based evidence for subsequent judgments.

[0139] User statement information also needs to be converted into a vector form that can be operated. Statement content usually includes location claims (such as "I am in Shanghai") and time context (such as "at 10 am today"), which can be parsed into structured geographic locations and time periods through natural language parsing and entity alignment methods. Then use the encoder to represent it as a statement location vector with spatial coordinate features and time range features, making it easy to compare with environmental information.

[0140] The four types of vectors are further fused into a multi-modal fusion feature vector. The fusion process can be achieved by splicing, weighted average or attention mechanism, so that the multi-source data produces complementary effect in a unified vector space. The feature representation formed after fusion contains geographic, semantic, environmental and subjective declaration information, and has cross-modal consistency and conflict detection capability.

[0141] Finally, the multi-modal fusion feature vector is input into a pre-trained location anomaly analysis model. The internal structure of the model can be a multi-layer perceptron, a graph attention network or a large language model fine-tuning architecture, which has logical reasoning and inconsistency detection capability. The model performs contradiction analysis, correlation reasoning and anomaly labeling on the input vector, outputs the geographic location authenticity analysis result, and generates an analysis reason with explainability. The reason can include conflict labels, reasoning paths or example explanations to enhance the transparency and credibility of the analysis result.

[0142] In this embodiment, by aligning various heterogeneous information including environmental sound features, knowledge graph information, text semantic data and user declarations in unified space-time, and respectively encoding them into unified format vectors, and then using multi-modal fusion strategy to construct a complete information expression structure, systematic judgment of conflict relationships between various types of information can be realized. After inputting the fusion vector into an anomaly analysis model with reasoning ability, spatial deviation, semantic contradiction and temporal inconsistency between information can be further identified, thereby outputting authenticity analysis results with high explainability and accuracy. This structure effectively solves the problem of difficult interaction and comparison of different modal data, and improves the ability to judge the credibility of user statements in complex contexts.

[0143] In one embodiment, the above step S507 includes:

[0144] S5071, by a logical conflict detection module in the location anomaly analysis model, analyzing the spatial correlation of the environmental feature vector and the geographic knowledge vector in the multi-modal fusion feature vector, and generating a geographic contradiction point credibility score;

[0145] S5072, by a time rule analysis module in the location anomaly analysis model, detecting the time distribution consistency of the environmental feature vector and the semantic vector in the multi-modal fusion feature vector, and generating a time window deviation coefficient;

[0146] S5073, generating intermediate analysis data containing spatial conflict heat and time anomaly level according to the geographic contradiction point credibility score and the time window deviation coefficient;

[0147] S5074, based on the intermediate analysis data, determining the anomaly probability of the user declaration information through an anomaly judgment decision tree, and generating a geographic location authenticity analysis result;

[0148] S5075, generating a visual analysis reason text according to the anomaly probability, the geographic contradiction point credibility score, and the time window deviation coefficient.

[0149] In this embodiment, the multi-modal fusion feature vector is taken as input data, integrating encoding information from different sources. The environmental feature vector and the geographic knowledge vector represent acoustic labels extracted from environmental sound and their geographic relationship with the regional knowledge graph, respectively. Both of them constitute a spatial information subset. The logical conflict detection module identifies the deviation of the two types of vectors in the dimensions of geographic range, habitat, and occurrence probability by analyzing the spatial semantic compatibility between them. The model calculates the conflict score based on spatial embedding similarity, label distribution overlap rate, or regional semantic co-occurrence degree, and obtains the geographic contradiction point credibility score for judging the credibility of information. This score measures the contradiction intensity between environmental content and the user's claimed region.

[0150] In the time dimension, the time sequence features in the environmental feature vector and the semantic vector are sent to the time regularity analysis module for comparison. Events in environmental sound usually have specific time distribution patterns, such as specific bird songs appearing only in the morning or certain seasons, and certain mechanical sounds such as traffic noise having statistical differences between weekdays and weekends. The semantic vector contains time information embodied in the dialogue text, such as specific time descriptions, daily routine vocabulary, or event sequence clues. The module uses time window overlap rate, event-time co-occurrence statistics, or time series matching algorithms to calculate the degree of consistency between the two in the time dimension, and outputs the time window deviation coefficient as the quantitative result.

[0151] The geographic contradiction point credibility score and the time window deviation coefficient are jointly input into the structured mapping module to generate intermediate analysis data, which describes the conflict state between information. This intermediate data explicitly indicates two indicators: spatial conflict intensity and time anomaly level. The former represents the degree of inconsistency between the user's declared location and the objective environmental information in space, and the latter reflects significant deviations or misplacements in time usage. Such labeled conflict levels facilitate subsequent classification and judgment under the rule-driven mechanism.

[0152] The anomaly determination decision tree receives the intermediate analysis data as input and makes reasoning judgments based on pre-set branch conditions. The decision tree may include branch paths such as high spatial conflict intensity and severe time anomaly level outputting a high-risk label. Each path is designed based on actual data verification and has interpretability and stability. The reasoning result is the anomaly probability, which represents the degree of non-compliance of the current data combination in the normal declaration behavior, and the geographic location authenticity analysis result is output as the overall judgment conclusion.

[0153] In combination with the above reasoning information, the model further generates visual analysis reason text, which is constructed based on a structured data interpretation template by combining elements such as anomaly probability, spatial conflict level, time deviation degree, and cited environmental feature labels into natural language expressions. The output text can include specific conflict point descriptions, data inconsistency reasons, reference evidence, or predicted logic paths, etc., which helps the reviewer or user understand the judgment basis, and improves the transparency and trustworthiness of the system.

[0154] Example: In the financial technology business, to prevent remote identity forgery and geographic location fraud, when a bank customer applies for high-risk financial services (such as online large loan activation) remotely, the system automatically collects background voice data stating that the customer is from a certain place in Guangdong. The voice data is obtained through an intelligent customer service access module, which contains the applicant's statement and the surrounding environmental sound. The collected voice data is first sent to the environmental sound enhancement module, which selectively enhances the gain of the non-speech background part to make the bird song, traffic sound, and other environmental features clearer and more distinguishable in subsequent processing. The audio is converted to a standardized format and recorded in the database along with the collection time and access device ID, forming a traceable data file. Subsequently, the audio data is sent to the speech separation model, which extracts the time-frequency domain voiceprint and background noise energy distribution to construct a frequency mask matrix and separate the speech and non-speech signals. The system generates two independent audio: one contains street noise, crowd dialogue, and dialect intermingling, and the other contains the user's pure speech data. The environmental sound is sent to the acoustic event detection system, which is classified and found to contain suspected bird calls and specific non-natural sound. The system further identifies the species of the biological acoustic event and identifies the characteristic voiceprint of the bird species unique to the south in the audio; the non-biological features identify the Cantonese tone pattern and the local ambulance alarm frequency band with high confidence matches. The multi-source labels are summarized as an environmental sound feature set and searched against the built-in geographic knowledge graph, and the system returns the distribution information that the voiceprint group mainly appears in the coastal cities of Guangdong. At the same time, the user's speech content is processed by frame, windowing, and feature extraction, and the speech recognition model is called to generate the preliminary text, which is corrected by the language model to form a structured dialogue text containing semantic clues such as "I am now at the seaside" and "Many people were on the square this morning." The above environmental sound feature set, geographic knowledge, dialogue text data, and user declared location information are sent to the analysis model. The system first unifies the time and geographic information dimensions, then vectorizes each type of information, and performs multi-modal fusion. Subsequently, the fusion features are sent to the location anomaly analysis model, and the system finds that the bird and dialect voiceprints strongly support the declared location through the logical conflict module, with a low spatial conflict score, but the time regularity analysis module finds that the environmental sound and the user's statement are not in the same time period, showing a clear inconsistency. The intermediate analysis data shows that the spatial conflict is low in intensity, and the time anomaly level is moderate, with an anomaly probability of 23%. The model finally outputs "the authenticity is credible" with the analysis reason: the environmental sound and the declared location have a high correlation, but the dialogue semantic timeline and the background sound have a certain time difference, and it is recommended to verify again. Through this mechanism, financial institutions can automatically verify the geographical authenticity of high-risk remote behavior without interrupting the business process.

[0155] In the medical health service scenario, a remote chronic disease patient communicates with the hospital health management center through the voice platform for daily follow-up, during which the system automatically collects the call environment sound and dialogue data for location consistency monitoring and health behavior authenticity evaluation. In a voice recording, the patient claims to be "walking around the community now", and the system detects weak signal state containing background broadcast sound and short human voice in the original voice signal. In order to improve the data quality, the system performs gain processing on the audio, enhances the environmental noise part, and marks the collection time and saves the encoded terminal device after using. Using a pre-trained voice separation model to process the audio, the system constructs a frequency band mask according to the time-frequency structure, and outputs two independent data streams. The environmental sound part includes the local news tone in the community broadcast, the conversation of the elderly group, and the occasional motor vehicle horn; while the voice part only retains the patient's statement voice. After acoustic event detection, the system identifies the "Shinan Community Health Station" related announcement in the broadcast segment, and the non-biological voiceprint part contains a voice fragment with Shandong dialect characteristics. The biological acoustic characteristics do not appear obviously, but the mechanical acoustic characteristics are highly consistent with the prompt sound in the community parking lot sound system. After atlas retrieval, the system classifies the above characteristics as the activity range of Shinan District, which is consistent with the patient's claimed location. The dialogue voice is converted into structured text after speech recognition, including "just finished exercising, there is a broadcast for free clinic downstairs today", etc. The system extracts semantic markers related to health activities and timelines. The environmental sound features, spatial knowledge tags, semantic dialogue content, and user registration address information are vectorized and fused into an analysis model, which performs spatial and semantic logic verification to determine whether there is a geographical or temporal conflict between dimensions. The final system analysis result is: "the statement content is credible", and the analysis text shows: "the broadcast content corresponds to the location, the voice semantics is consistent with the time distribution, supporting the current geographical statement." This process enhances the service authenticity verification in the medical interaction scenario without affecting the patient's interactive experience, improves the reliability of remote health data, and assists in achieving closed-loop control of chronic disease follow-up and health behavior management.

[0156] In this embodiment, by structuring the environmental feature vector, geographical knowledge vector, semantic vector, and statement information, etc. multi-modal data, relying on logical conflict detection and time rule analysis to realize the internal consistency verification of information, and introducing structured mapping and interpretable reasoning mechanism, the semantic deviation and regional inconsistency between multi-source information can be systematically identified. Through the generation of credibility score and deviation quantization coefficient, the standardized expression of spatial and temporal conflicts is realized, and the authenticity is probabilistically output under the decision tree mechanism. The finally generated visual analysis reason not only provides the judgment result, but also has the ability of data level traceability and logical path explanation, effectively improving the intelligent analysis level of the authenticity of user claims.

[0157] In an embodiment, a device for authenticity analysis based on environmental sound features is provided, which corresponds to the method for authenticity analysis based on environmental sound features in the above-mentioned embodiments. Referring to Figure 3 , Figure 3 A functional module schematic diagram of a preferred embodiment of the device for authenticity analysis based on environmental sound features of the present application. The voice collection module 10, the voice separation module 20, the environmental perception analysis module 30, the voice recognition module 40 and the authenticity analysis module 50. The detailed description of each functional module is as follows:

[0158] The voice collection module 10 is used to obtain raw voice data containing environmental sound;

[0159] The voice separation module 20 is used to perform voice separation processing on the raw voice data to generate environmental sound data and clean voice data;

[0160] The environmental perception analysis module 30 is used to analyze the environmental sound data, extract environmental sound features, and retrieve knowledge information associated with the environmental sound features;

[0161] The voice recognition module 40 is used to perform voice content recognition on the clean voice data to generate dialogue text data;

[0162] The authenticity analysis module 50 is used to input the environmental sound features, knowledge information, dialogue text data and user declaration information into an analysis model to output authenticity analysis results.

[0163] In an embodiment, the voice collection module 10 is specifically used for:

[0164] Collecting raw voice signals containing environmental sound through the microphone module of the terminal device;

[0165] Performing environmental sound frequency band selective gain processing on the raw voice signals to generate raw voice data containing enhanced environmental sound;

[0166] Converting the raw voice data into a standardized audio format file;

[0167] Storing the standardized audio format file into a storage area and recording the collection time and collection device identifier of the raw voice data.

[0168] In an embodiment, the voice separation module 20 is specifically used for:

[0169] Loading a pre-trained voice separation model based on a deep neural network;

[0170] Extracting time-frequency domain voiceprint features and background noise energy distribution features of the raw voice data through the voice separation model;

[0171] generating a frequency band mask matrix according to a frequency band energy ratio of the time-frequency domain voiceprint feature and the background noise energy distribution feature, and taking the frequency band mask matrix as a sound source separation threshold;

[0172] performing frequency band mask processing on a frequency domain signal of the original voice data based on the sound source separation threshold through the voice separation model;

[0173] performing inverse short-time Fourier transform on the frequency domain signal after the mask processing to generate a time domain reconstruction signal;

[0174] separating the time domain reconstruction signal to generate environmental sound data and clean voice data.

[0175] In an embodiment, the environment perception analysis module 30 is specifically configured to:

[0176] performing acoustic event detection on the environmental sound data to generate an acoustic event set;

[0177] performing species classification and identification on a biological acoustic event in the acoustic event set to generate a biological acoustic feature set;

[0178] performing voiceprint pattern matching on a non-biological acoustic event in the acoustic event set to generate a mechanical acoustic feature set;

[0179] merging the biological acoustic feature set and the mechanical acoustic feature set into an environmental sound feature set;

[0180] taking spatial distribution information associated with the environmental sound feature set as the knowledge information through geographical knowledge graph retrieval.

[0181] In an embodiment, the voice recognition module 40 is specifically configured to:

[0182] performing frame windowing processing on the clean voice data to generate a voice signal after frame windowing processing;

[0183] extracting a mel-frequency cepstral coefficient feature of the voice signal after frame windowing processing to generate an acoustic feature vector;

[0184] performing phoneme sequence decoding on the acoustic feature vector through a pre-trained voice recognition model to generate initial text data;

[0185] performing grammar correction on the initial text data based on a grammar correction model to generate standardized dialogue text data.

[0186] In an embodiment, the authenticity analysis module 50 is specifically configured to:

[0187] aligning the environment sound feature, the knowledge information, the dialogue text data and the user declaration information to generate a set of spatio-temporal aligned input data;

[0188] encoding the environment sound feature in the set of spatio-temporal aligned input data as an environment feature vector;

[0189] encoding the knowledge information in the set of spatio-temporal aligned input data as a geographical knowledge vector;

[0190] encoding the dialogue text data in the set of spatio-temporal aligned input data as a semantic vector;

[0191] encoding the user declaration information in the set of spatio-temporal aligned input data as a declaration position vector;

[0192] fusing the environment feature vector, the geographical knowledge vector, the semantic vector and the declaration position vector to generate a multi-modal fusion feature vector;

[0193] inputting the multi-modal fusion feature vector into a pre-trained location anomaly analysis model to output a geographical location authenticity analysis result and an analysis reason.

[0194] In an embodiment, the authenticity analysis module 50 is specifically configured to:

[0195] analyzing the spatial correlation between the environment feature vector and the geographical knowledge vector in the multi-modal fusion feature vector through a logical conflict detection module in the location anomaly analysis model to generate a geographical contradiction point credibility score;

[0196] detecting the time distribution consistency between the environment feature vector and the semantic vector in the multi-modal fusion feature vector through a time rule analysis module in the location anomaly analysis model to generate a time window deviation coefficient;

[0197] generating intermediate analysis data containing spatial conflict heat and time anomaly level according to the geographical contradiction point credibility score and the time window deviation coefficient;

[0198] determining an anomaly probability of the user declaration information through an anomaly judgment decision tree based on the intermediate analysis data to generate a geographical location authenticity analysis result;

[0199] generating a visual analysis reason text according to the anomaly probability, the geographical contradiction point credibility score and the time window deviation coefficient.

[0200] In an embodiment, a computer device is provided, which can be a server, and an internal structure diagram of the computer device can be as shown in Figure 4As shown in the figure. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide determination and control capabilities. The memory of the computer device includes non-volatile and / or volatile storage media, internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with the external user terminal through the network connection. The computer program is executed by the processor to realize the functions or steps of the server side of the environmental sound feature-based authenticity analysis method.

[0201] In one embodiment, a computer device is provided, which can be a user terminal, and its internal structure diagram can be as shown in the figure. Figure 5 As shown in the figure. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide determination and control capabilities. The memory of the computer device includes non-volatile storage media, internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with the external server through the network connection. The computer program is executed by the processor to realize the functions or steps of the user terminal side of the environmental sound feature-based authenticity analysis method

[0202] In one embodiment, a computer device is provided, including a memory, a processor and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to realize the following steps:

[0203] Obtain raw speech data containing environmental sound;

[0204] Perform speech separation processing on the raw speech data to generate environmental sound data and clean speech data;

[0205] Analyze the environmental sound data, extract environmental sound features, and retrieve knowledge information associated with the environmental sound features;

[0206] Perform speech content recognition on the clean speech data to generate dialogue text data;

[0207] Input the environmental sound features, knowledge information, dialogue text data and user declaration information into an analysis model, and output an authenticity analysis result.

[0208] In one embodiment, a computer readable storage medium is provided, having stored thereon a computer program, which when executed by a processor implements the following steps:

[0209] obtaining original speech data containing environmental sound;

[0210] performing speech separation processing on the original speech data to generate environmental sound data and clean speech data;

[0211] analyzing the environmental sound data, extracting environmental sound features, and retrieving knowledge information associated with the environmental sound features;

[0212] performing speech content recognition on the clean speech data to generate dialogue text data;

[0213] inputting the environmental sound features, knowledge information, dialogue text data, and user declaration information into an analysis model, and outputting authenticity analysis results.

[0214] It should be noted that the functions or steps that the above computer readable storage medium or computer device can implement can correspond to the related descriptions of the server side and the user side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0215] Those of ordinary skill in the art can understand that all or part of the processes in the above embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above embodiments. In the embodiments provided in the present application, any reference to memory, storage, database or other medium can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), 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), etc.

[0216] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is taken as an example, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0217] It should be noted that if non-company software tools or components appear in the embodiments of the present application, they are only used for example introduction and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method of authenticity analysis based on environmental sound features, characterized by, The method comprises the following steps: obtaining original speech data containing environmental sound; performing speech separation processing on the original speech data to generate environmental sound data and clean speech data; analyzing the environmental sound data, extracting environmental sound features, and retrieving knowledge information associated with the environmental sound features; the analysis of the environmental sound data, the extraction of the environmental sound features, and the retrieval of the knowledge information associated with the environmental sound features comprise: performing acoustic event detection on the environmental sound data to generate a set of acoustic events; performing species classification and identification on biological acoustic events in the set of acoustic events to generate a set of biological acoustic features; performing voiceprint pattern matching on non-biological acoustic events in the set of acoustic events to generate a set of mechanical acoustic features; combining the set of biological acoustic features and the set of mechanical acoustic features into a set of environmental sound features; retrieving spatial distribution information associated with the set of environmental sound features from a geographic knowledge graph as the knowledge information; performing speech content recognition on the clean speech data to generate dialogue text data; inputting the environmental sound features, knowledge information, dialogue text data, and user declaration information into an analysis model to output a authenticity analysis result.

2. The authenticity analysis method based on environmental sound features according to claim 1, characterized by, Obtaining original speech data containing environmental sound comprises: collecting original speech signals containing environmental sound through the microphone module of the terminal device; performing environmental sound frequency band selective gain processing on the original speech signals to generate original speech data containing enhanced environmental sound; converting the original speech data into a standardized audio format file; storing the standardized audio format file in a storage area and recording the collection time and collection device identifier of the original speech data.

3. The authenticity analysis method based on environmental sound features according to claim 1, characterized by, The speech separation processing on the original speech data to generate environmental sound data and clean speech data comprises: loading a pre-trained deep neural network-based speech separation model; extracting time-frequency domain voiceprint features and background noise energy distribution features of the original speech data through the speech separation model; generating a frequency band mask matrix according to the frequency band energy ratio of the time-frequency domain voiceprint features and the background noise energy distribution features, and taking the frequency band mask matrix as a sound source separation threshold; performing frequency band mask processing on the frequency domain signals of the original speech data based on the sound source separation threshold through the speech separation model; performing inverse short-time Fourier transform on the frequency domain signals after mask processing to generate time domain reconstruction signals; separating the time domain reconstruction signals to generate environmental sound data and clean speech data.

4. The authenticity analysis method based on environmental sound features according to claim 1, characterized by, The speech content recognition on the clean speech data to generate dialogue text data comprises: performing frame windowing processing on the clean speech data to generate frame windowing processed speech signals; extracting mel-frequency cepstral coefficient features of the frame windowing processed speech signals to generate acoustic feature vectors; performing phoneme sequence decoding on the acoustic feature vectors through a pre-trained speech recognition model to generate initial text data; performing grammar correction on the initial text data based on a grammar correction model to generate standardized dialogue text data.

5. The authenticity analysis method based on environmental sound features according to claim 1, characterized by, The environmental sound features, knowledge information, dialogue text data, and user declaration information are input into an analysis model to output authenticity analysis results, including: The environmental sound features, knowledge information, dialogue text data, and user declaration information are subjected to data alignment processing to generate a spatio-temporally aligned input data set; The environmental sound features in the spatio-temporally aligned input data set are encoded into an environmental feature vector; The knowledge information in the spatio-temporally aligned input data set is encoded into a geographic knowledge vector; The dialogue text data in the spatio-temporally aligned input data set is encoded into a semantic vector; The user declaration information in the spatio-temporally aligned input data set is encoded into a declaration position vector; The environmental feature vector, geographic knowledge vector, semantic vector, and declaration position vector are fused to generate a multi-modal fusion feature vector; The multi-modal fusion feature vector is input into a pre-trained location anomaly analysis model to output a geographic location authenticity analysis result and an analysis reason.

6. The authenticity analysis method based on environmental sound features according to claim 5, characterized by, The multi-modal fusion feature vector is input into a pre-trained location anomaly analysis model to output a geographic location authenticity analysis result and an analysis reason, including: The spatial correlation between the environmental feature vector and the geographic knowledge vector in the multi-modal fusion feature vector is analyzed by a logical conflict detection module in the location anomaly analysis model to generate a geographic conflict point credibility score; The temporal distribution consistency between the environmental feature vector and the semantic vector in the multi-modal fusion feature vector is detected by a temporal regularity analysis module in the location anomaly analysis model to generate a time window deviation coefficient; Based on the geographic conflict point credibility score and the time window deviation coefficient, intermediate analysis data containing spatial conflict heat and temporal anomaly level are generated; Based on the intermediate analysis data, the anomaly probability of the user declaration information is determined by an anomaly determination decision tree to generate a geographic location authenticity analysis result; Based on the anomaly probability, geographic conflict point credibility score, and time window deviation coefficient, a visual analysis reason text is generated.

7. An authenticity analysis device based on environmental sound features, characterized by, The authenticity analysis device based on environmental sound features includes: A voice collection module for obtaining original voice data containing environmental sound; A voice separation module for performing voice separation processing on the original voice data to generate environmental sound data and clean voice data; An environmental perception analysis module for analyzing the environmental sound data, extracting environmental sound features, and retrieving knowledge information associated with the environmental sound features; The environmental perception analysis module includes: An acoustic event detection unit for performing acoustic event detection on the environmental sound data to generate an acoustic event set; A biological acoustic classification unit for performing species classification and identification on biological acoustic events in the acoustic event set to generate a biological acoustic feature set; A mechanical acoustic matching unit for performing voiceprint pattern matching on non-biological acoustic events in the acoustic event set to generate a mechanical acoustic feature set; A feature merging unit for merging the biological acoustic feature set and the mechanical acoustic feature set into an environmental sound feature set; The knowledge searching unit is configured to query a geographical knowledge graph based on the set of environmental sound features, and retrieve associated spatial distribution information as the knowledge information. The speech recognition module is configured to perform speech content recognition on the pure speech data, and generate dialogue text data. The authenticity analysis module is configured to input the environmental sound features, the knowledge information, the dialogue text data, and user declaration information into an analysis model, and output an authenticity analysis result.

8. A computer device, comprising: The computer device comprises a memory, a processor, and an environmental sound feature-based authenticity analysis program stored in the memory and executable on the processor, and the environmental sound feature-based authenticity analysis program, when executed by the processor, implements the steps of the environmental sound feature-based authenticity analysis method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores an environmental sound feature-based authenticity analysis program, and the environmental sound feature-based authenticity analysis program, when executed by the processor, implements the steps of the environmental sound feature-based authenticity analysis method according to any one of claims 1-6.

Images