Data tracing method and device based on cooperative training, equipment and medium

Abstract

The invention relates to the technical field of voice processing, can be applied to business scenes such as financial science and technology, medical health and the like, and discloses a data tracing method, device, equipment and medium based on cooperative training, which comprises the following steps: obtaining a generative model and a discrimination model to output initial intermediate feature representation, generating voice data and inputting the voice data into the discrimination model to generate discrimination loss; updating the generation model based on the discrimination loss to obtain a traceable intermediate feature representation, generating an updated intermediate feature representation by using the updated generation model, generating training voice data, and inputting the training voice data and real data into the discrimination model for training to obtain an updated discrimination model; and when input voice data is received, the updated discrimination model is used for discriminating and outputting a detection result whether the input voice data is generated by the generation model or not. According to the invention, through cooperative training of the generative model and the discrimination model, the generative model outputs voice embedded with traceable features, the rejection capability of the discrimination model to unknown sources is improved, and waveform disturbance reduction and accurate source determination are realized.

Description

Technical Field

[0001] This invention relates to the field of speech processing technology, and in particular to a data tracing method, apparatus, device, and storage medium based on collaborative training. Background Technology

[0002] In recent years, the performance of generative speech synthesis has continued to improve, and high-similarity speech under zero-shot conditions has been rapidly adopted in various scenarios. Accompanying this is the real-world demand for speech source verification and traceability. Existing approaches mainly fall into two categories: output layer watermark embedding and deepfake detection. However, the lack of coordination between the generation and detection ends makes traceable features susceptible to data distribution drift and unstable. The detection end lacks the ability to reject unknown or unregistered sources, and the representation spaces from the generation to the detection end are inconsistent, resulting in a distribution mismatch between training and inference, which leads to a decrease in the reliability of judgments.

[0003] In the watermark embedding path, information is typically injected into the vocoder waveform or encoded features. Additional perturbations can easily affect the auditory perception. Furthermore, the embedding position is strongly coupled with the specific generation pipeline. Once the feature distribution on the generation side changes, the detection end struggles to stably extract and align the feature at the feature layer, leading to signal attenuation or mismatch. This type of method is separated from the training objective of the discriminator, lacking collaborative optimization based on a unified representation. Consequently, source confirmation relies on post-processing rules rather than end-to-end consistent representation learning.

[0004] In deepfake detection, discriminators mainly rely on short-time spectral features and binary classification. The distribution seen during training is offset from the distribution in actual application, and there is a lack of alignment mechanism with the representation of the generator. It is difficult to give a stable positive judgment for inputs from registered generator models. At the same time, it lacks robust rejection threshold decisions and evidence support for inputs from non-registered sources, which easily leads to misjudgments under open set conditions.

[0005] In the healthcare sector, voice is used for question-and-answer interaction, assisted diagnosis and treatment, and rehabilitation guidance. The verifiability of the voice source directly impacts the credibility of medical records and compliance audits. Current practices struggle to maintain a stable alignment between intermediate representations at the generation end and decision-making at the detection end while ensuring usable medical audio quality. This results in insufficient robustness in verifying the source of registered audio and a lack of reliable evidence to quantify the rejection of non-registered sources, hindering medical evidence collection and post-audit.

[0006] In the fintech business, voice is used for intelligent customer service, remote verification, and transaction instruction interaction. Misconceptions about the source of information can amplify business risks and dispute costs. In the existing system, the mismatch in representation caused by the separate training of the generation and discrimination sides makes it difficult to robustly set the confirmation threshold for the registration generation source in complex scenarios. When faced with non-registered sources, there is a lack of a joint decision-making mechanism that combines source similarity clues and discrimination confidence, resulting in insufficient rejection capability and affecting the certainty and interpretability of risk control strategy triggering. Summary of the Invention

[0007] The main objective of this invention is to provide a data tracing method, apparatus, device, and storage medium based on collaborative training, aiming to solve the technical problems of existing technologies where the lack of collaboration between generation and discrimination leads to instability of traceable intermediate features and insufficient robustness in determining whether a data belongs to a registered generation model.

[0008] To achieve the above objectives, the present invention provides a data tracing method based on collaborative training, comprising:

[0009] Obtain the generative model and the discriminative model, and output the initial intermediate feature representation through the generative model;

[0010] Based on the initial intermediate feature representation, speech data is generated through acoustic conversion, and the speech data is input into the discrimination model to generate discrimination loss;

[0011] The generative model is updated based on the discriminative loss so that the generative model outputs a traceable intermediate feature representation, thus obtaining the updated generative model;

[0012] The updated generative model is used to generate an updated intermediate feature representation containing traceable features;

[0013] Based on the updated intermediate feature representation, training speech data is generated through acoustic conversion, and the training speech data and real data are input into the discrimination model for training to obtain the updated discrimination model.

[0014] Upon receiving input speech data, the updated discrimination model is used to discriminate the input speech data, and a detection result is output regarding whether the input speech data was generated by the generation model.

[0015] Furthermore, to achieve the above objectives, the present invention provides a data tracing device based on collaborative training, comprising:

[0016] The generative model building module is used to acquire the generative model and the discriminative model, and to output the initial intermediate feature representation through the generative model;

[0017] An acoustic feature conversion module is used to generate speech data based on the initial intermediate feature representation through acoustic conversion, and input the speech data into the discrimination model to generate discrimination loss;

[0018] The generative model update module is used to update the generative model based on the discriminative loss so that the generative model outputs a traceable intermediate feature representation, thereby obtaining the updated generative model.

[0019] The traceable feature generation module is used to generate an updated intermediate feature representation containing traceable features using the updated generation model;

[0020] The discriminant model update module is used to generate training speech data through acoustic conversion based on the updated intermediate feature representation, and input the training speech data and real data into the discriminant model for training to obtain the updated discriminant model.

[0021] The speech data discrimination module is used to discriminate the input speech data using the updated discrimination model when the input speech data is received, and output a detection result type of whether the input speech data was generated by the generation model.

[0022] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a data tracing program based on collaborative training stored in the memory and executable on the processor, wherein when the data tracing program based on collaborative training is executed by the processor, it implements the steps of the data tracing method based on collaborative training as described above.

[0023] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a data tracing program based on collaborative training, wherein when the data tracing program based on collaborative training is executed by a processor, it implements the steps of the data tracing method based on collaborative training as described above.

[0024] Beneficial Effects: This invention relates to the field of speech processing technology and can be applied to business scenarios such as fintech and healthcare. It discloses a data tracing method, apparatus, device, and medium based on collaborative training, comprising: acquiring a generative model and a discriminative model and outputting initial intermediate feature representations; generating speech data based on the initial intermediate feature representations and inputting it into the discriminative model to generate a discriminative loss; updating the generative model based on the discriminative loss to output traceable intermediate feature representations; generating updated intermediate feature representations containing traceable features using the updated generative model; generating training speech data based on the updated intermediate feature representations and inputting it into the discriminative model with real data to obtain an updated discriminative model; and upon receiving input speech data, using the updated discriminative model to discriminate and outputting a detection result indicating whether the speech data was generated by the generative model. This invention establishes collaborative training between the generative and discriminative models, forming stable traceable features in the intermediate representations and aligning them with the discriminative side representation space. This enables the system to reliably determine whether the input speech originates from a registered generative model and to robustly reject non-registered sources. It also reduces additional perturbations to the waveform without relying on output layer watermarks, thereby minimizing adverse effects on auditory perception. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0026] Figure 1 This is a schematic diagram of an application environment for a data tracing method based on collaborative training according to an embodiment of the present invention;

[0027] Figure 2 This is a flowchart illustrating an embodiment of the data tracing method based on collaborative training according to the present invention.

[0028] Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the data tracing device based on collaborative training of the present invention;

[0029] Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0030] Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0031] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0032] The data tracing method based on collaborative training provided in this invention can be applied to, for example... Figure 1In this application environment, the user terminal communicates with the server via a network. The server can obtain the generative model and the discriminative model from the user terminal and output an initial intermediate feature representation. Based on the initial intermediate feature representation, it generates speech data and inputs it into the discriminative model to generate a discriminative loss. Based on the discriminative loss, it updates the generative model to output a traceable intermediate feature representation. The updated generative model is then used to generate an updated intermediate feature representation containing traceable features. Based on the updated intermediate feature representation, it generates training speech data and inputs it into the discriminative model with real data to obtain an updated discriminative model. When receiving input speech data, the updated discriminative model is used to discriminate and output a detection result indicating whether the speech data was generated by the generative model. This invention establishes collaborative training between the generative model and the discriminative model, forming stable traceable features in the intermediate representation and aligning them with the discriminative side representation space. This enables the system to reliably determine whether the input speech originates from the registered generative model and to robustly reject non-registered sources. It also reduces additional perturbations to the waveform without relying on output layer watermarking, thereby minimizing adverse effects on auditory perception. The user terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0033] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the data tracing method based on collaborative training provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0034] like Figure 2 As shown, the data tracing method based on collaborative training proposed in this invention includes the following steps:

[0035] S10, Obtain the generative model and the discriminative model, and output the initial intermediate feature representation through the generative model;

[0036] In this embodiment, obtaining the generative and discriminative models requires constructing generative and discriminative models with specific architectures in the computing environment. The generative model can be obtained by calling existing pre-trained model weight files and loading them into the computing framework, or by training a deep neural network based on large-scale speech and text pairing data. The generative model generally consists of an encoder, a decoder, and an intermediate feature transformation module. The encoder can employ a Transformer structure with multi-head attention to encode the input text information into a high-dimensional feature sequence. The discriminative model can be obtained by pre-training a classification network for distinguishing between real and fake speech. This network can include a context information capture unit and a multi-scale convolutional feature extraction unit. The context information capture unit is responsible for extracting cross-time step correlation information from the temporal data of the input speech, while the multi-scale convolutional feature extraction unit extracts local and global acoustic features at different time and frequency resolutions.

[0037] After model acquisition, the process of generating the initial intermediate feature representation requires encoding the input text data into vectorized features and forming a high-dimensional feature matrix in the feature generation path of the generative model, which can be used for subsequent acoustic transformation. The text data is first transformed into a dense vector representation through a character or phoneme embedding layer, and then sequentially passed through multiple self-attention modules for context modeling, generating an encoded sequence containing information such as language, intonation, and prosody. This encoded sequence is mapped to an intermediate layer in the acoustic feature space by a vector transformation module. The generated intermediate feature representation retains the semantic and acoustic prior information required for speech generation, while avoiding the direct generation of waveform signals, so as to embed source features in subsequent processes. In implementation, the intermediate feature representation can use Mel-spectral features, cepstral coefficient features, or high-dimensional latent space vectors, and its dimension and resolution can be adjusted according to the target speech quality and model computational performance. To ensure that the output intermediate feature representation is compatible with the subsequent discriminative model interface, a feature normalization module needs to be added at the output end to map the feature value range to a numerical range suitable for the input layer of the discriminative model, preventing inconsistent numerical distributions from causing training instability.

[0038] In different implementation environments, generative and discriminative models can be loaded and built in various ways. When running on a high-performance server environment, a Transformer architecture generative model containing hundreds of millions of parameters can be loaded to improve the naturalness and detail fidelity of the speech. When running on resource-constrained embedded devices or mobile devices, a compressed, lightweight generative model can be selected, reducing computational overhead through weight quantization and structural pruning. The discriminative model can be implemented using a convolutional neural network structure specifically optimized for Mel-spectral features, or a multi-input network structure fusing time-domain waveforms and frequency-domain features, to improve robustness against different types of speech. In the feature representation generation stage, the fine-grained features and computational efficiency can be balanced by adjusting the number of hidden units in the coding layer, the number of attention heads, and the sequence length. Furthermore, the number and frequency range of Mel filter banks can be customized according to the speech sampling rate of the target task to adapt to the speech signal characteristics of different acquisition devices or application scenarios.

[0039] Example Explanation: In the healthcare field, input medical record text information can be converted into spoken audio. The intermediate feature representation output by the generation model supports subsequent addition of medical record traceability information and discriminant analysis, thereby ensuring the security and traceability of medical information dissemination. This feature representation can record the spoken language content and acoustic patterns, enabling rapid identification of the source when liability determination is required.

[0040] In the fintech sector, customer transaction notifications or financial announcements can be input into a generation model to produce intermediate feature representations with specific rhythms and speech rates. Subsequent tracing and verification processes then ensure the authenticity and reliability of the voice information, thereby reducing the risk of voice fraud. In this process, the intermediate feature representation serves as a crucial data structure connecting the generation and verification stages, combining generation accuracy with verification convenience, making it suitable for business scenarios with high security requirements.

[0041] This embodiment, by outputting structured intermediate feature representations in the generative model, can retain rich semantic and acoustic information without directly generating speech waveforms. This provides a stable and controllable feature foundation for subsequent embedding of source identification and discriminant analysis, while improving the interface compatibility between the generative and discriminant models and reducing the risk of cross-model information loss.

[0042] S20, based on the initial intermediate feature representation, speech data is generated through acoustic conversion, and the speech data is input into the discrimination model to generate discrimination loss;

[0043] In this embodiment, the process of generating speech data through acoustic conversion based on the initial intermediate feature representation requires mapping the high-dimensional intermediate feature representation to the acoustic feature domain and further synthesizing it into a playable speech signal. This intermediate feature representation can be a Mel-spectral latent space vector, a cepstral coefficient matrix, or a quantized high-dimensional embedded feature, originating from the output of the generative model encoder and already containing information such as the semantics, prosody, and timbre of the target speech. The acoustic conversion is typically implemented in two stages. The first stage uses a vocoder conversion module to convert the intermediate feature representation into Mel-spectral features or linear spectral features. This process can employ a mapping model based on convolutional neural networks, recurrent neural networks, or autoregressive structures to ensure the accuracy of the mapping from intermediate features to spectral features. The second stage uses a waveform reconstruction module to restore the spectral features to a time-domain waveform. This module can use a neural vocoder based on generative adversarial networks (such as HiFi-GAN, WaveGlow, etc.) or a signal reconstruction method based on inverse Fourier transform, and uses phase reconstruction techniques (such as the Griffin-Lim algorithm) during the reconstruction process to improve sound quality.

[0044] After generating speech data, it needs to undergo signal processing to adapt to the input requirements of the discriminant model. This includes operations such as sampling rate unification, amplitude normalization, and silence clipping to ensure that the data received by the discriminant model during training or inference has stable signal characteristics. These processing operations can be implemented through a digital signal processing module, which can process different batches of speech samples in batches to reduce manual intervention.

[0045] The process of inputting the generated speech data into the discriminative model requires first converting the speech signal into a feature representation that the discriminative model can recognize, such as Mel spectrum, linear spectrum, MFCC, etc., and then feeding it into the context-aware feature extractor of the discriminative model to capture the time-series features and acoustic patterns of the speech. The context-aware feature extractor can consist of one-dimensional convolutional layers, bidirectional long short-term memory networks, Transformer coding layers, etc., to capture short-term and long-term dependencies. Subsequently, the extracted features are input into a multi-layer convolutional classification network, which progressively compresses the feature space through multiple convolutional and pooling layers and performs pattern discrimination, outputting the probability distribution of each category (real or generated).

[0046] In the discriminative loss generation stage, the output probability distribution of the discriminative model needs to be compared with the corresponding ground truth labels. The ground truth labels are derived from the annotation information of the dataset; generated samples are labeled as the generated class, and real recordings are labeled as the ground truth class. The comparison process uses the binary cross-entropy loss function, with the following formula:

[0047] L = -[ylog(p) + (1-y)log(1-p)]

[0048] Where y is the true label and p is the probability of the generated class predicted by the model. This loss value can quantify the model's classification error on the input samples and provide gradient signals for backpropagation of the generative model.

[0049] In high-performance server environments, high-fidelity vocoders based on large-scale neural networks can be used for spectrum-to-waveform conversion to achieve near-realistic recording quality. In mobile or embedded environments, lightweight vocoder models combined with low sampling rate outputs can reduce computational overhead. For the input to the discriminative model, a multi-feature fusion approach can be employed, concatenating Mel-frequency spectra with MFCC features to improve discrimination accuracy. Additional regularization terms, such as label smoothing loss, can be introduced when generating the discriminative loss to enhance the model's generalization ability to unseen synthetic samples.

[0050] In some business scenarios, an environmental simulation module can be introduced during the acoustic conversion stage to add different noises, reverberation, or channel distortion to the generated speech, thereby improving the adaptability of the discrimination model to real-world complex environments. This strategy has a significant advantage when dealing with detection tasks that require handling various recording conditions.

[0051] Example: In the field of healthcare, when generating voice broadcasts from medical record text, this acoustic conversion and discrimination loss generation process can ensure that the synthesized voice not only meets the requirements for broadcast clarity, but can also be accurately distinguished in subsequent verification stages, preventing medical instructions from being tampered with by the voice.

[0052] In the fintech business, when transaction confirmation information is converted into voice prompts, generating voice and producing discriminative loss in the discriminative model can improve anti-counterfeiting detection capabilities and reduce the success rate of voice fraud. Even if attackers use advanced voice synthesis models to generate prompts, the discriminative model can still identify and trace the source.

[0053] This embodiment transforms the initial intermediate feature representation into high-fidelity speech data and inputs it into the discriminative model to generate a discriminative loss. This allows for simultaneous optimization of the capabilities of both the generative and discriminative models during training, ensuring that the generated speech meets the requirements in terms of quality and traceability. At the same time, it keeps the discriminative model sensitive to the differences between the generated speech and the real speech, thereby improving the reliability of source tracing detection.

[0054] S30, update the generative model based on the discriminative loss so that the generative model outputs a traceable intermediate feature representation, and obtain the updated generative model;

[0055] In this embodiment, the process of updating the generative model based on the discriminative loss essentially involves using the information fed back from the discriminative model through a backpropagation mechanism to make targeted adjustments to the parameters of the generative model. This allows it to embed traceable information into the intermediate feature layer while outputting sound quality and semantic information. The discriminative loss is calculated by the discriminative model based on the classification difference between the generated speech and the real speech. This loss value not only reflects the difference in authenticity between the synthesized results of the generative model but also serves as a guiding signal, driving the generative model to increase traceable features while maintaining the naturalness of the speech.

[0056] In implementation, the gradient update process of the generative model is typically divided into three stages. The first stage is gradient calculation, which involves backpropagating the discriminative loss to the feature output layer of the generative model to obtain a gradient tensor. This gradient tensor encompasses the sensitivity of different parameters to the discriminative loss. The second stage is embedding traceability information. This process requires introducing a controllable identifier signal into the feature embedding space of the generative model. This signal can be a high-dimensional vectorized watermark, a hash-based identifier embedding, or a frequency domain feature perturbation pattern. To ensure the stability and anti-interference of the embedded information, the embedding amplitude of the identifier can be controlled by modulating the traceability strength parameter, which can be continuously adjusted between [0,1]. The third stage is weight matrix adjustment, which uses a weighted tensor superposition method to fuse the modified feature embedding space with the existing weight matrix, allowing the traceability features to be naturally integrated into the generated intermediate feature representation in the subsequent speech synthesis process.

[0057] The updated generative model not only optimizes the overall weights but may also involve targeted updates to low-rank parameter matrices. These low-rank matrices are typically located in the bottleneck layer or decoder of the generative model; updating them can significantly alter the traceability of output features without significantly affecting the synthesis quality. To improve training efficiency and model convergence speed, adaptive optimization algorithms (such as Adam, RAdam, or AdamW) can be used to quickly adjust the weights related to traceability embeddings, and gradient clipping can be used to prevent training instability caused by gradient explosion.

[0058] In cloud training environments with high computational resources, a multi-GPU distributed training framework can be used to train the generative and discriminative models in parallel, reducing latency through asynchronous gradient propagation. In this approach, the calculation result of each discriminative loss is quickly fed back to the generative model for adjusting weights and embedding parameters. In edge computing environments requiring low-latency responses, a lightweight gradient update mechanism can be used, updating only the sub-network parameters highly correlated with the traceable feature embeddings, thus reducing computational load.

[0059] The encoding method for traceable information can vary depending on the business scenario. In healthcare scenarios where voice data needs to be stored for a long time and transmitted across systems, information such as medical record numbers and consultation dates can be mapped to embedding vectors into the feature space through hash encoding. In financial transaction scenarios, transaction serial numbers, timestamps, and other unique identifier vectors can be generated and embedded to verify the data source in subsequent transaction audits.

[0060] In terms of optimization strategy, the loss function of the generative model can be designed as a weighted sum of discriminative loss and source embedding loss. The former ensures generation quality, while the latter ensures the detectability of source embedding. The weight ratio can be dynamically adjusted according to the actual task. For example, in environments with extremely high security requirements, the weight of source embedding loss can be set to 0.6, while the weight of generation quality loss can be 0.4, to prioritize ensuring traceability.

[0061] Example: In the healthcare business, when the system generates medical record broadcast voice, the medical information identifier can be embedded in the feature layer of the voice through this update process. Even if the voice is intercepted, compressed or transmitted, the discrimination model can still identify its source through the traceability mechanism to prevent the medical record data from being illegally replaced.

[0062] In the fintech business, when generating voice authentication for customers, the transaction serial number is embedded into the feature layer to ensure that even if the voice is reprocessed or mixed, its source can still be confirmed through detection mechanisms in subsequent audits, reducing the risk of voice synthesis fraud.

[0063] This embodiment achieves stable and difficult-to-remove traceable source information embedding in the intermediate feature layer of synthesized speech through generative model updates based on discriminative loss, enhancing subsequent source tracing capabilities while maintaining the naturalness of the synthesized speech. This training method maintains high source identification accuracy even when facing model substitution attacks and audio transcoding attacks.

[0064] S40, using the updated generative model to generate an updated intermediate feature representation containing traceable features;

[0065] In this embodiment, generating updated intermediate feature representations containing traceable features using the updated generative model is a process of performing inference computation on new input data, assuming the generative model has already been trained using discriminative loss and has the ability to embed traceable information. The core of this process is to use specific feature encoding, identifier matching, and fusion mechanisms to ensure that the generated intermediate features not only retain the semantic and structural information of the input data but also embed specific information that can be used for subsequent traceability.

[0066] In implementation, the input data first needs to be encoded to enter the feature processing flow of the generative model. For text-to-speech generation tasks, the Transformer encoder inside the generative model can first be used to transform the input text into a structured index sequence, which can effectively preserve the semantic relationships, word order information, and contextual dependencies of the text. Subsequently, sampling is performed in the vector embedding space of the generative model to obtain a set of latent vectors representing the generated content. These latent vectors can be divided into content latent vectors representing semantic features and positional latent vectors representing temporal, positional, and other information. At this stage, a dynamic sampling control module can be introduced to adjust the sampling density according to the task's resource budget, real-time requirements, target audio quality standards, etc., to achieve a balance between synthesis quality and computational efficiency.

[0067] Next, traceability information needs to be incorporated into the intermediate feature representation. This process is achieved by querying the updated implicit traceability identifier library of the generative model. This identifier library has been integrated with the traceability strength parameter during the training phase of the generative model and contains various identifier vectors. These identifier vectors can come from various unique identification information, such as medical record IDs in healthcare scenarios and transaction serial numbers in financial transaction scenarios. Based on the hash value or other unique encoding of the input data, identifier vectors associated with the input data can be selected from the identifier library.

[0068] To integrate the identifier information into the final generated intermediate features, a specific fusion method is required, such as weighted tensor superposition, which fuses the content latent vector and the identifier vector according to a preset weight ratio. The weight ratio can be dynamically adjusted to ensure that the strength of the source information embedding meets detection requirements without significantly impacting the naturalness and intelligibility of the generated speech. The fused vector is called the traceable content latent vector. This vector retains the identifier features in subsequent feature decoding, making the generated intermediate feature representation inherently contain traceable information. Finally, the traceable content latent vector is concatenated with the position latent vector and input into the decoding part of the generative model to obtain an updated intermediate feature representation containing traceable features.

[0069] In speech synthesis scenarios requiring high precision, high-dimensional dense vectors can be used as content latent vectors and identifier vectors, and a multi-head weighted fusion mechanism can be employed during fusion. Different weighting heads can adjust the embedding weights for frequency domain features, time domain features, or semantic features respectively. In scenarios with high real-time requirements, low-dimensional sparse vectors can be selected and a single-channel fusion mechanism can be used to reduce computational latency.

[0070] In healthcare scenarios, a fixed-length identifier vector, derived from the patient's medical record code through hashing, can be stored in an implicit source identifier library. When generating a patient diagnosis announcement, the corresponding identifier vector is retrieved from this library and embedded into the content implicit vector, allowing the source of the speech to be traced during subsequent transmission, storage, or transcoding.

[0071] In fintech business scenarios, transaction serial numbers and timestamps can be jointly encoded to generate identifier vectors, ensuring that multiple transactions within the same time period each have unique traceability characteristics. When generating transaction broadcast voice, this identifier vector is fused with the content latent vector, so that during subsequent compliance audits or dispute resolution, the embedded features can be detected to verify whether the voice is the original generated result.

[0072] During identifier matching and fusion, the traceability strength parameter can be adjusted according to the security requirements of different application scenarios. For example, in financial scenarios, the parameter can be set higher to enhance the robustness of traceability feature detection, while in medical scenarios, it can be appropriately reduced to ensure the audibility and comfort of the voice during diagnosis and treatment communication.

[0073] This embodiment fuses identifier vectors and content latent vectors during the generative model inference stage. The resulting intermediate feature representations naturally carry traceable source information that corresponds one-to-one with the input data, enabling the subsequent generated speech to be accurately traced to its source. This embedding mechanism does not rely on additional external watermark injection and is not easily corrupted by model replacement or audio transcoding, thereby improving the long-term stability and security of the source information.

[0074] S50, based on the updated intermediate feature representation, training speech data is generated through acoustic conversion, and the training speech data and real data are input into the discrimination model for training to obtain the updated discrimination model;

[0075] In this embodiment, training speech data is generated through acoustic conversion based on the updated intermediate feature representation. The training speech data and real data are then input into the discrimination model for training to obtain the updated discrimination model. After the generation model has output the updated intermediate feature representation with traceable features, the feature is further converted into speech data that can be directly used for training the discrimination model. This data is then jointly input with the real collected speech data to optimize the discrimination model's ability to recognize generated speech.

[0076] First, the updated intermediate feature representation needs to be transformed into a speech-processable feature form. This process can be achieved through an acoustic feature transformation module, which maps the updated intermediate feature representation to the Mel-spectral feature space, making it a high-dimensional spectrogram representation that reflects frequency domain information. During the transformation process, a traceability intensity parameter is injected, ensuring that the generated acoustic features retain the traceability information embedded in the preceding generative model. This injection process can be achieved through additive shift, spectral masking, or frequency-domain phase modulation, ensuring that identifier information is embedded into the spectral structure of the training speech data without significantly compromising the naturalness of the speech.

[0077] Subsequently, the generated acoustic features are normalized to reduce interference from channel conditions, recording equipment differences, or background noise on the training of the discriminant model. Cepstral mean subtraction is a commonly used normalization method. By calculating the mean of the features in each frame and subtracting this mean, the long-term spectral envelope component is eliminated, making the training data acoustically closer to the real data distribution. After normalization, the neural vocoder module of the generative model is used to restore the spectral features to a time-domain waveform, obtaining directly playable training speech data. The neural vocoder can employ structures such as HiFi-GAN or WaveGlow to ensure high fidelity in waveform quality of the synthesized speech.

[0078] In terms of real data preparation, a dynamic sampling module can be used to perform Mel-spectrum transformation and normalization on real speech data, ensuring consistency in feature representation between real and generated data and avoiding bias in the discrimination model caused by differences in feature distribution. Dynamic sampling can also adjust the input data size according to computational resources and training time requirements; for example, when resources are limited, high-confidence samples with strong representativeness can be sampled first.

[0079] Finally, the training speech data and standardized real data are jointly input into the multi-scale convolutional discriminant network of the discriminant model for training. The multi-scale convolutional structure can extract features at different temporal and frequency resolutions, comprehensively modeling audio feature differences from local details to global patterns. During training, spectral consistency loss is used to measure the difference between generated data and real data in the spectral space, while adversarial loss guides the discriminant model to learn the potential distributional differences of the generated data. This dual-loss joint optimization method can simultaneously improve the discriminant model's ability to recognize details and distinguish overall distributions. The updated discriminant model obtained after training has higher generated speech detection accuracy and source identification rate.

[0080] In scenarios requiring high audio quality, higher-dimensional Mel spectra (e.g., 128-dimensional) and STFT parameters with small windows and high overlap can be used during the acoustic feature transformation stage to enhance the fidelity of spectral details. In scenarios prioritizing real-time performance, lower-dimensional spectra (e.g., 64-dimensional) and reduced overlap can be selected to reduce computational overhead.

[0081] In healthcare operations, training voice data can come from simulated doctor-patient dialogues, while real data can be collected from authorized real medical voice recordings. By maintaining the consistency of the two types of data in the training set, the discrimination model can identify whether the medical broadcast is generated by the system.

[0082] In fintech business, training voice data can be used to synthesize transaction announcements by updating intermediate feature representations, while real data can be collected from real transaction voice confirmation processes. By using a discriminative model to learn the subtle differences in acoustic features between the two types of data, the system can accurately distinguish between automatically generated and manually announced announcements in subsequent detection.

[0083] For different training objectives, the traceability intensity parameter and dynamic sampling strategy can be adjusted. For example, to enhance the performance of the discriminative model under low signal-to-noise ratio conditions, mixed noise environment samples can be introduced during training, and the weight ratio of spectral consistency loss can be appropriately increased to strengthen the stability of frequency domain features.

[0084] This embodiment uses training speech data generated by updating intermediate feature representations and real speech data simultaneously for training the discriminative model, which significantly improves the discriminative model's ability to distinguish between generated and real data. At the same time, the embedded traceable features in the generated data are not weakened, thus ensuring that the subsequent detection stage can accurately determine the data source and identify the traceability identifier.

[0085] S60, upon receiving input speech data, the updated discrimination model is used to discriminate the input speech data, and a detection result is output regarding whether the input speech data was generated by the generation model.

[0086] In this embodiment, upon receiving input speech data, the updated discriminant model is used to discriminate the input speech data and output a detection result indicating whether the input speech data was generated by the generative model. This is part of the process of using the discriminant model for actual speech source identification and tracing after training. This process requires the input speech signal to undergo multiple stages, including feature processing, model inference, and source tracing judgment, ultimately generating a detection result with a clear source determination and traceable identifier.

[0087] After receiving the input speech data, preprocessing is performed first. The original waveform signal is input into the dynamic sampling module to perform operations such as noise reduction, segmentation, and sampling rate adjustment. Noise reduction can eliminate background noise using spectral subtraction or deep neural network noise reduction models. Segmentation divides long-term speech into short-time frame sequences suitable for model processing. Sampling rate adjustment ensures the consistency of the input features with those used during the training of the discrimination model.

[0088] The preprocessed speech frames require feature extraction. The time-domain waveform is mapped to a two-dimensional spectral representation using a Mel-spectrum conversion module, reflecting the energy distribution of speech in the time and frequency dimensions. To eliminate the long-term influence of device differences and pronunciation conditions on the features, cepstral mean subtraction normalization is further applied to remove the mean of each frame's features, making the features more concentrated on the dynamic changes of short-term speech, resulting in a detection-normalized acoustic feature vector.

[0089] The normalized acoustic feature vector is input into the updated discriminant model. The multi-scale convolutional discriminant network extracts deep features at different time and frequency scales and generates the original discriminant score vector through forward inference. This score vector is then probability-mapped to obtain a discriminant confidence scalar corresponding to the category generated by the generative model, reflecting the probability that the input speech belongs to the generated data.

[0090] In the source tracing detection stage, the implicit source tracing identifier library maintained in the updated generative model is used to retrieve and detect the identifier vector that is closest to the normalized acoustic feature vector. Euclidean distance, cosine similarity, or other metrics can be used during the matching process. The matched identifier vectors are processed with weighted similarity and combined with multi-dimensional weight coefficients (such as semantic, frequency domain, time domain, and spatial dimensions) to generate a source tracing confidence scalar, which is used to measure the credibility of embedding source tracing identifiers in the speech.

[0091] A joint decision is made by combining the discrimination confidence scalar and the source tracing confidence scalar. If both are higher than a preset threshold, the input speech data is determined to be generated by the generative model; otherwise, it is determined to be real speech data. To form a final traceable detection result, a preset hash function (such as SHA-256) is applied to the matched identifier vector to generate a unique source tracing identifier, ensuring that the original vector data of the identifier is not leaked during the storage and transmission of source tracing information. The detection result includes the data source type and the source tracing identifier, facilitating subsequent tracking and auditing.

[0092] In high-security applications, more stringent joint decision-making strategies can be selected, such as increasing the discrimination and tracing thresholds, or introducing a third-party trusted verification service to encrypt and sign the hash-generated tracing identifier. In scenarios requiring real-time detection, the computational process of dynamic sampling and Mel-frequency spectrum conversion can be optimized, and a lightweight convolutional discriminant model structure can be adopted to reduce inference latency.

[0093] In healthcare operations, the input voice may be the voice of a patient or doctor during a remote consultation. The system detects whether the voice was generated by the voice generation system and can trace it back to the generation record to prevent fake medical record audio or forged medical instructions from entering the system.

[0094] In fintech businesses, voice input may be transaction instructions or account verification announcements. By detecting and tracing its source, we can prevent black market operators from using voice synthesis to impersonate customers and issue transaction instructions. At the same time, the source of the generated information can be restored through traceability identifiers during post-event audits.

[0095] Furthermore, in low-bandwidth environments, the tracing identifier can be output as a short code, and the complete tracing information can be queried on the backend server based on the short code, thereby reducing the amount of output data at the terminal. In multilingual environments, a language recognition module can be added to the front end of the discrimination model, enabling the retrieval of the corresponding language branch when the normalized acoustic feature vector is detected and matched with the identifier library, thus improving the accuracy of cross-language tracing.

[0096] This embodiment combines the discriminative model with an implicit source identifier library during the actual detection phase. This not only determines whether the speech was generated by the generative model but also outputs a unique source identifier for the generated data, improving the reliability and traceability of speech source determination. Joint confidence assessment reduces the risk of misjudgment from a single detection path, while hash processing ensures the security of the source identifier when publicly output.

[0097] This invention relates to the field of speech processing technology and can be applied to business scenarios such as fintech and healthcare. It discloses a data tracing method, apparatus, device, and medium based on collaborative training, comprising: acquiring a generative model and a discriminative model and outputting initial intermediate feature representations; generating speech data based on the initial intermediate feature representations and inputting it into the discriminative model to generate a discriminative loss; updating the generative model based on the discriminative loss to output traceable intermediate feature representations; generating updated intermediate feature representations containing traceable features using the updated generative model; generating training speech data based on the updated intermediate feature representations and inputting it into the discriminative model with real data to obtain an updated discriminative model; and upon receiving input speech data, using the updated discriminative model to discriminate and outputting a detection result indicating whether the speech data was generated by the generative model. This invention establishes collaborative training between the generative and discriminative models, forming stable traceable features in the intermediate representations and aligning them with the discriminative side representation space. This enables the system to reliably determine whether the input speech originates from a registered generative model and to robustly reject non-registered sources. It also reduces additional perturbations to the waveform without relying on output layer watermarking, thereby minimizing adverse effects on auditory perception.

[0098] In one embodiment, step S10 includes:

[0099] S101, construct a deep Transformer backbone network configured with a multi-layer attention mechanism;

[0100] S102, Connect flow-based models on the deep Transformer backbone network to form feature generation paths;

[0101] S103, integrate a text-audio alignment module on the feature generation path;

[0102] S104, Deploy a dynamic sampling module on the text-audio alignment module;

[0103] S105, integrate the deep Transformer backbone network, the flow-based model, the text-audio alignment module, and the dynamic sampling module to form a complete generative model;

[0104] S106, Construct a context-aware feature extractor;

[0105] S107, configured with a lightweight convolutional classifier network and set up a multi-scale feature capture mechanism;

[0106] S108, The context-aware feature extractor is integrated with the lightweight convolutional classification network to form a complete discrimination model;

[0107] S109, The input text data is processed through the complete generation model, and the initial intermediate feature representation is output.

[0108] In this embodiment, a generative model and a discriminative model are acquired, and the generative model outputs an initial intermediate feature representation. The focus is on the controllable mapping from text to acoustic representation and the stability of the data interface for subsequent source tracing training. First, a deep Ransformer backbone network with a multi-layer attention mechanism is constructed. Text data is converted into an index sequence by a word segmenter, establishing trainable word embeddings and learnable positional encodings. At least twelve layers of multi-head self-attention and feedforward sub-layers are stacked, employing a pre-normalized structure and residual connections to ensure deep gradient propagation. The number of attention heads, hidden dimensions, and feedforward expansion factor are set based on the longest duration of the target speech and its semantic complexity. Attention masks are used to block invalid padding positions, and regularization is achieved through Dropout and weight decay. The hidden sequence output by this backbone network maintains a one-to-one semantic load corresponding to the input tokens, providing a conditional sequence for subsequent duration modeling and acoustic conditionalization.

[0109] A flow-based model is connected to a deep Ransformer backbone network to form a feature generation path. The flow model uses reversible coupling layers and learnable one-dimensional convolutions with LU decomposition or 1×1 reversible convolutions to construct an analytically reversible mapping with Jacobian determinant. Affine coupling and ActNorm are introduced to stabilize log-likelihood training. A bidirectional mapping from semantic conditions to continuous latent variables is established with a Gaussian prior as the target distribution. The conditional path linearly projects and repeatedly upsamples the hidden sequences of the backbone network to align the conditional length with the target acoustic frame length, and then injects them into the conditional branches of each coupling layer. Through maximum likelihood training, the latent variables output by the flow model statistically fit the prior, and the inverse graph can serve as a carrier of continuously differentiable acoustic content, constituting the numerical domain and topological structure of the initial intermediate feature representation.

[0110] A text-audio alignment module is integrated into the feature generation path to achieve monotonic duration mapping from token-level to frame-level. Three implementations are possible: a CTC-based aligner, a monotonic attention search, or a learnable duration predictor. The CTC aligner obtains the optimal monotonic path through forward and backward recursion and outputs the frame occupancy of each token; the monotonic attention search uses attention peaks and cumulative thresholds to constrain temporal progression; and the learnable duration predictor regresses discrete frame counts on the backbone hiddens and uses a negative binomial distribution or Gaussian NLL constraints. The alignment results are used to extend the conditional sequence to the frame domain, enabling the streaming model to generate continuous latent variables under frame-level conditions, ensuring that the semantic-to-temporal mapping is traceable and reproducible.

[0111] A dynamic sampling module is deployed on the text-to-audio alignment module to perform frame-level subsampling and segment reweighting to address the computational and bandwidth constraints of large-scale training. The dynamic sampling strategy constructs priority scores based on frame energy, attention entropy, and duration confidence. It generates sampling masks using segmented Reservoir sampling or Top-p frame retention. The threshold and retention ratio are determined through a dual-objective optimization of the validation set's recognition rate and alignment error. The alignment path is fixed before subsampling, and the alignment boundary remains on the sampling index after subsampling to avoid disrupting the monotonicity of text-to-frame transitions. This module exposes the sampling index and reweighting coefficients downstream, ensuring reproducibility and playability while reducing the disturbance of redundant frames to training stability.

[0112] The deep Transformer, a flow-based model, a text-audio alignment module, and a dynamic sampling module are integrated into a directed acyclic computation graph to form a generative model capable of end-to-end inference. The forward pass involves the text index entering the backbone network to obtain the hidden sequence, which is then extended to the frame domain using the alignment results. After dynamically sampling and cropping the frame set, the flow model outputs a continuous latent variable tensor under the constraints of conditional branches, serving as the initial intermediate feature representation. This representation is defined as a real-valued matrix of shape T×C, where T is the sampled frame length and C is the latent variable dimension. Each column maintains statistical constraints of unit variance and zero mean, and each row is strictly aligned with the corresponding text frame. The representation retains spatial sub-channels and scalar placeholders for subsequent source tracing control. The interface exposes the length, mask, alignment index, and reweighting coefficients as key-value pairs, facilitating direct calls for acoustic conversion and collaborative training.

[0113] A context-aware feature extractor for the discriminative side is constructed in parallel. This extractor primarily employs a multi-branch structure consisting of a one-dimensional convolutional pyramid, gated linear units, and lightweight self-attention. The receptive fields of each branch are configured with dilation coefficients to cover both short-term and long-term speech phenomena. Channel attention is adaptively recalibrated between branches using Squeeze-Excitation or gated normalization. Temporal normalization utilizes grouped layer normalization to stabilize long-sequence inference. The extractor outputs a hierarchical multi-scale representation tensor, providing separable time-frequency patterns for the classification head.

[0114] A lightweight convolutional classification network with a multi-scale feature capture mechanism is configured as a deployable discriminant head. The classification network consists of depthwise separable convolutions, switchable dilated convolutions, and cross-scale fusion units, with kernel sizes of 3, 5, and 7 in parallel combinations. The number of channels decreases along the layers to control the number of parameters. Global temporal convergence and affine calibration are used to output binary or multi-class logarithmic values. The classification head and context-aware feature extractor are sequentially concatenated to form the discriminant model, outputting the raw score and interpretable intermediate layer features for reuse in subsequent adversarial training and spectral consistency constraints.

[0115] The actual implementation process of generating a generative model to process input text data and output initial intermediate feature representations is as follows: the text index enters the backbone network to obtain the hidden sequence; the alignment module provides the frame interval corresponding to each token; dynamic sampling extracts representative frames within each interval and assigns weights; the conditional branch broadcasts the hidden sequence in the frame domain and fuses it with the sampling weights; the streaming model generates continuous latent variables forward under this condition; the latent variables, along with the alignment and sampling metadata, are encapsulated into an initial intermediate feature representation and written to a shared buffer for direct reading by acoustic conversion. To ensure numerical stability, a small amount of Gaussian noise is applied at the entrance of the streaming model during training to scatter the channels, and noise is turned off during inference to maintain determinism; to ensure alignment consistency, all random subsampling uses a fixed seed and records the sampling index in the representation for complete reproduction.

[0116] This embodiment enhances the expressive power of text-to-semantic conditions through a deep Transformer. Stream-based invertible mapping enables continuous latent variables to satisfy analyzable density and maintain differentiable conditional generation. Text-audio alignment ensures that the one-to-one mapping from token to frame is recordable and replayable. Dynamic sampling reduces redundant computation and overfitting risks without compromising monotonicity, while retaining reweighted coefficients to maintain statistical consistency. Parallel construction of context-aware extraction and lightweight convolutional classification provides a structured interface and reusable representations for subsequent collaborative training. Based on the above structure, the initial intermediate feature representation is strictly aligned with the text temporally, statistically approximates a standard Gaussian, and includes alignment and sampling metadata in its interface. Therefore, it is easier to be stably decoded by acoustic conversion and more conducive to the subsequent introduction of source tracing channels and adversarial signals, resulting in faster training convergence, lower distortion, and stronger traceability.

[0117] In one embodiment, step S20 above includes:

[0118] S201, Based on the initial intermediate feature representation, Mel spectrum features are generated through the acoustic feature conversion module;

[0119] S202, The waveform reconstruction module is used to convert the Mel spectrum features into a time-domain speech signal to generate speech data;

[0120] S203, Perform signal normalization processing on the voice data to generate optimized voice data;

[0121] S204, The optimized speech data is input into the context-aware feature extractor of the discrimination model to extract multi-scale acoustic features;

[0122] S205, the multi-scale acoustic features are analyzed through the lightweight convolutional classification network of the discrimination model to obtain the probability distribution of speech authenticity;

[0123] S206, Based on the difference between the speech authenticity probability distribution and the real label, determine the binary cross-entropy loss and generate the discriminant loss.

[0124] In this embodiment, speech data is generated based on the initial intermediate feature representation, driving the discriminative model to produce logged supervision signals. A complete computational chain is constructed around the reversible mapping from features to acoustic frames, audible waveform restoration, input normalization, discriminative front-end representation, and probabilistic output. The initial intermediate feature representation is a frame-aligned continuous matrix of shape T×C, containing alignment indices and frame masks. The acoustic feature conversion module reads this matrix and alignment metadata, converting frame-level latent variables into Mel spectral features through conditional mapping. This can be achieved using a conditional flow model or conditional normalized flow, mapping the latent variables to a logarithmic power spectrum, multiplying by a Mel filter bank, and taking the logarithmic order of magnitude to obtain a Mel spectral sequence of dimension K; alternatively, a diffuse acoustic mapping with autoregressive priors can be used, treating the latent variables as conditional vectors for noise injection paths, iteratively denoising and outputting the Mel spectrum. The center frequencies of the Mel filter bank are uniformly placed in the range of 0 to Nyquist frequencies on a geometric scale, the bandwidth is calculated according to Slaney calibration, and the frame shift and frame length use window parameters consistent with the alignment to maintain time synchronization.

[0125] The waveform reconstruction module receives Mel-domain spectral features and reconstructs the time-domain speech signal. This can be achieved using a neural vocoder, such as HiFi-GAN, WaveRNN, or a fully convolutional inverse spectral model. Taking HiFi-GAN as an example, the Mel-domain spectrum is first upsampled via one-dimensional convolution, then sub-band signals are synthesized using multi-branch residual blocks, and finally, the full-band waveform is reconstructed using a polyphase synthesizer. During the training phase, the discriminator can employ a multi-scale structure to stabilize the reconstruction quality. Alternatively, Griffin-Lim inversion can be used to approximately reconstruct the linear spectrum from the Mel-domain, followed by iterative phase estimation to obtain the waveform. However, a phase gradient acceleration strategy can be introduced to reduce distortion. The output waveform is buffered at a sampling rate of 16kHz or 24kHz, in 16-bit integer or floating-point format, to ensure consistency with the subsequent standardization and discrimination front-end processing.

[0126] Signal normalization processes unify the waveform to a stable amplitude and statistical distribution. First, DC removal and amplitude clipping are performed, using short-time mean square energy to estimate peak values ​​and proportionally limit amplitude to eliminate abnormal spikes. Then, mean-variance normalization is performed, calculating the mean μ and standard deviation σ for the entire segment or segments, outputting (x-μ) / (σ+ε) to avoid numerical instability. To reduce channel effects, a cepstral mean normalization process can be superimposed, subtracting the channel mean frame-by-frame in the Mel or logarithmic power spectral domain, inversely transforming back to the waveform, or directly delivering it to the discrimination front-end in the frequency domain. The normalized audio definition optimizes speech data, fixing the sampling rate, bit width, and peak level to ensure consistency across batches and devices.

[0127] The context-aware feature extractor receives optimized speech data and outputs multi-scale acoustic features. This can be achieved using a parallel structure of a multi-branch one-dimensional convolutional pyramid and lightweight attention. The receptive fields of the branches are set to different spans ({5, 9, 17, 33}) through dilated convolutions to cover short-term excitations and long-term prosody. Channel attention is adaptively recalibrated between branches using Squeeze-Excitation or gated normalization. Temporal normalization employs grouped layer normalization to improve the steady-state properties of long sequences. The outputs are aligned in the temporal dimension and concatenated in the channel dimension to form a multi-scale acoustic feature tensor, preserving short-term details and cross-sentence context.

[0128] A lightweight convolutional classification network generates speech realism probability distributions on multi-scale acoustic features. The network consists of depthwise separable convolutions, switchable dilated convolutions, and cross-layer fusion units. A parallel combination of kernel sizes of 3, 5, and 7 enhances the texture separability between real and synthesized speech. The number of channels decreases from top to bottom to control parameter quantity and latency. The terminal outputs a binary log-probability using global temporal convergence and affine calibration, which is then mapped using a sigmoid algorithm to obtain the probability distributions belonging to the real and synthesized categories. To improve robustness to unknown generators, frequency band occlusion and temporal occlusion enhancements can be introduced in the intermediate layers, and temperature scaling is added before the output to calibrate the confidence level.

[0129] The binary cross-entropy loss generates a discriminative loss based on the difference between the speech authenticity probability distribution and the real label. Let y∈{0,1} represent the real label, p represent the synthetic class probability, and the loss is defined as -[y·log(p)+(1-y)·log(1-p)]. To alleviate class imbalance, a weighting factor or Focal variant can be introduced. The hyperparameter γ and class weight α are determined through a validation set grid search. The loss is accumulated within batches and backpropagated to the discriminative network to update the convolutional kernels and normalization parameters. Simultaneously, in the collaborative training framework, it is fed as a supervisory signal to the generation side, prompting the initial intermediate feature representations to be rearranged in a traceable direction in subsequent iterations. To ensure numerical stability, gradient clipping is limited to within the L2 norm threshold. The optimizer can be AdamW, and the learning rate adopts a cosine annealing or piecewise descent strategy, decoupled from the learning rate scheduling of the generation side to avoid synchronous oscillations.

[0130] This embodiment establishes a strict alignment relationship between continuous latent variables and Mel spectrum, high-fidelity reconstruction of audible waveforms, standardization of amplitude and channel consistency, multi-scale contextual representation extraction, and calibrated probability output. This enables the discriminator to stably distinguish between real and synthetic data under a unified distribution. The supervision signal in the form of cross-entropy is fed back to the training process in a numerically stable manner, making the initial intermediate feature representation separable and traceable in subsequent iterations. This reduces the discrimination error rate and gradient noise, improves the generalization ability to different text content and speaking conditions, and provides a stable and reproducible experimental interface for subsequent source embedding and collaborative updates.

[0131] In one embodiment, step S30 above includes:

[0132] S301, Based on the discriminant loss, the gradient tensor of the generated model is determined through the backpropagation module, and the gradient information of the feature layer is generated;

[0133] S302, Based on the gradient information of the feature layer, modify the traceability strength parameter of the feature embedding space of the generator model, and modulate the weight matrix of the generator model to embed implicit traceability identifiers.

[0134] S303 generates a traceable intermediate feature representation prototype by superimposing and fusing the modified feature embedding space and the modulated weight matrix using weighted tensors.

[0135] S304, Based on the traceable intermediate feature representation prototype, the low-rank parameter matrix of the generated model is updated using an adaptive optimization module to generate an optimized model parameter set;

[0136] S305, Load the optimized model parameter set into the generator model architecture, and generate the updated generator model.

[0137] In this embodiment, the discriminative loss is derived from the binary classification error signal of the discriminative model for speech authenticity. Its numerical domain is non-negative real numbers, and it is commonly represented by binary cross-entropy or its focal variant. This loss is fed into the computation graph on the generation side as external supervision, and is propagated back through the acoustic feature transformation subgraph and text conditional subgraph via the backpropagation module until all trainable nodes are reached. The backpropagation module performs automatic differentiation and gradient accumulation, providing a gradient tensor equivalent to the first derivative of the loss for each trainable tensor. The gradient tensor has the same shape as the forward tensor in both the time and channel dimensions, and uses mixed-precision storage to reduce GPU memory pressure. To suppress numerical explosion, gradient clipping is limited by a global L2 norm threshold, and gradient noise injection is used to smoothly update the curve in low signal-to-noise ratio ranges.

[0138] The gradient information of the feature layers points to the key layers in the generative model that produce intermediate feature representations, such as the embedding layer, the attention projection layer, and the bottleneck layer of the feedforward network. By performing channel reduction and frequency band aggregation on the gradient tensors of these layers, a sensitivity estimate for each semantic channel and each frequency band can be obtained. This sensitivity is used to drive two types of operations: one is the hyperparameter adjustment for the feature embedding space, and the other is the controllable modulation of the weight matrix. The sensitivity estimate can be obtained using the absolute value of the gradient mean, the gradient variance, or an approximation based on Fisher information; the choice of strategy is determined by the stability of the validation set.

[0139] The traceability intensity parameter in the feature embedding space exists in scalar or vector form and is used to control the energy level of the traceability signal injected into the intermediate representation. The scalar form uses an α to uniformly control the embedding intensity of all channels, while the vector form uses α... → Intensity is controlled separately in the channel or frequency band dimension. The initial value of the intensity parameter comes from offline calibration, or it can be adaptively fine-tuned according to the backpropagation gradient during collaborative training. The numerical domain is limited to a closed interval between zero and one to avoid excessive perturbation. To ensure semantic preservation, the intensity parameter is positively correlated with semantic similarity and negatively correlated with the perceptual distortion index. This constraint is applied to the optimization objective through a regularization term.

[0140] Weight matrix modulation is used to embed implicit source identifiers. These implicit source identifiers are given as fixed-dimensional vectors or sparse codes, generated either as pseudo-random codes based on content hashing or as index vectors from a distributed fingerprint database. The modulation targets are the query key-value projection matrix of the attention layer and the bottleneck mapping matrix of the feedforward network. Two mechanisms can be used for implementation. One is weight perturbation, where the identifier vector is expanded using Kronecker or block-broadcast and mapped to a mask of the same shape as the weights. Then, additive or multiplicative perturbation is performed in the weight space using a scaling factor β, while constraining the spectral norm change to not exceed a set threshold. The other is low-rank injection, introducing an outer product patch ΔW of rank r equal to A multiplied by B, where A and B are generated from the identifier vector through two layers of perceptrons. Since r is much smaller than the original dimension, the identifier signal is injected into the controlled subspace. These two mechanisms can be used independently or in combination, distributed across layers.

[0141] Weighted tensor superposition achieves functional fusion of the modified feature embedding space and the modulated weight matrix. Specifically, two adjustments are applied simultaneously at the computation node generating the intermediate feature representation: one adjusts the scale or sparsity of the hidden representation h through intensity parameters, and the other modulates the spectral characteristics of the mapping operator W through weight modulation. A new forward mapping h′ is then constructed, equal to f, modulated by W, and intensity-modulated by h, synthesized using residual gating. The function f represents the nonlinear computation of the corresponding layer, and the residual gating threshold is given by the reduction result of the feature layer gradient information at that layer. The synthesized output maintains the same length as the original representation in the sequence dimension, and is called the traceable intermediate feature representation prototype. To ensure the reversibility of the downstream acoustic mapping, a smoothing regularization is applied to the prototype in the frequency domain, and a total variational constraint is applied in the time domain.

[0142] The adaptive optimization module is responsible for parameter updates, with the target limited to trainable variables related to low-rank parameter matrices and weight perturbations. If the LoRA mechanism is used, the updated variables are two sets of matrices, A and B. The optimizer can be AdamW, and a group learning rate strategy can be introduced, allowing groups more relevant to semantic channels to use smaller step sizes and groups more relevant to source channels to use larger step sizes. If a hybrid architecture of parallel outer product patching and gated residuals is used, the gate coefficients and strength parameters also need to be updated simultaneously, either in a two-stage or joint manner. The two-stage approach first freezes the strength parameters and updates only the low-rank matrix; after convergence, the strength parameters are unfrozen, and the joint objective is fine-tuned. The joint approach uses multi-objective weighted optimization to simultaneously minimize the discrimination loss and perceptual distortion regularity, with the loss weights automatically adjusted through adaptive uncertainty estimation. During the update process, weight decoupling regularization and low-rank constraints are applied to keep the rank of ΔW below r and control the Frobenius norm within a set interval.

[0143] The optimized model parameter set is written into the generative model architecture after consistency verification. Verification items include shape matching, spectral radius constraints, gradient checkpoint recoverability, and inference throughput evaluation. The writing process distinguishes between persistent parameters and session parameters; the former is written to disk for subsequent training and offline evaluation, while the latter resides in memory for immediate inference. After loading, a dry run is performed to verify the extractability and semantic preservation of the traceable intermediate feature representation prototype on representative samples. If the threshold requirement is met, it is considered the updated generative model. To avoid forgetting and drift, an exponential moving average shadow copy can be introduced for the inference channel, managed separately from the training copy.

[0144] This embodiment backpropagates the discriminative loss to the key layer that generates the intermediate representation, and adjusts the hidden representation with an intensity parameter while modulating the mapping weights in a low-rank or perturbative manner. This allows for the steady-state encoding of implicit identifiers into the intermediate representation without significantly altering the semantic output distribution. Weighted tensor superposition and gated residuals decouple the identifier channel from the semantic channel, and optimization only applies to low-rank variables and small-amplitude spectral perturbations, resulting in stable training and controllable resource overhead. This improves the survival rate and detectability of the source signal in the downstream acoustic mapping and resampling chain, while maintaining synthesis quality and content consistency, reducing gradient noise on the discriminative side, and enhancing the convergence speed and cross-domain generalization ability of collaborative training.

[0145] In one embodiment, step S40 above includes:

[0146] S401, Input the input text data into the updated generative model's Transformer encoder to generate a text index string;

[0147] S402, based on the text index string, sample in the vector embedding space of the quantized variational autoencoder of the updated generative model, and apply a dynamic sampling module to control resource consumption, generating content latent vectors and position latent vectors.

[0148] S403, based on the hash value of the input text data, match the identifier vector from the implicit traceability identifier library of the updated generative model, and fuse the content implicit vector and the identifier vector by weighted tensor superposition to generate a traceable content implicit vector.

[0149] S404, concatenate the traceable content latent vector with the location latent vector, input the concatenation result into the updated generative model, and generate an updated intermediate feature representation containing traceable features.

[0150] In this embodiment, the input text data is discretized and contextualized using the Transformer encoder within the updated generative model. Specifically, the token sequence after word segmentation is mapped to a d_model dimensional embedding, superimposed with learnable positional encoding or rotational positional encoding, and then processed through multi-head self-attention and feedforward network stacking to output a hidden sequence of equal length. To form the text index string, dictionary maximum likelihood projection or vector quantization projection is used. Each hidden vector is compared with the vocabulary prototype vector or codebook prototype, and the index with the highest similarity is selected, resulting in an integer index sequence of length T. This process supports mixed-precision computation and cached key-value pairs to reduce latency.

[0151] The text index string drives the vector embedding space of the quantized variational autoencoder for sampling. The implementation involves constructing multiple codebook groups, each containing K d_code-dimensional prototype vectors. Each codebook group represents a subspace such as semantics, prosody, and acoustic position. During sampling, the index string is first routed to the target codebook group via a learnable mapping. Then, nearest neighbor retrieval is performed based on Euclidean distance or cosine similarity, and the nearest prototype vector is used as the discrete embedding. The gradient is then backpropagated via a pass-through estimator. A dynamic sampling module is located between the retrieval and assembly stages, adjusting the sampling density based on computational budget and uncertainty score: when the confidence level is above a threshold, only the top-1 prototypes are retained; when the confidence level is in the middle range, a diversity temperature is activated, and random sampling is performed on the candidate set according to a softmax distribution; when the confidence level is too low, resampling is triggered and the number of candidates is increased. This mechanism controls the number of codebook accesses at the batch level and limits memory consumption. By combining embeddings from the semantic codebook, a content latent vector sequence is obtained; by combining embeddings from the position / rhythm codebook, a position latent vector sequence is obtained. The two types of latent vectors are aligned in the time dimension and configured with d_content and d_pos respectively in the channel dimension, both maintaining the same length as the text index string.

[0152] The hash value of the input text data is used to determine the tracing identifier retrieval target. This is achieved by normalizing the original token sequence or its embedded sequence before calculating the hash. Normalization includes character case regularization, whitespace folding, and language tag injection. The hash algorithm can be SHA-256 or BLAKE3, and a fixed-length prefix is ​​taken from the hash result as the retrieval key. The implicit tracing identifier library is organized in key-value pairs, where the key is the retrieval key and the value is one or more identifier vectors. The identifier vector dimension is consistent with the content implicit vector dimension, and the source can be an offline assigned fingerprint library, fingerprint embeddings obtained during training distillation, or a weighted synthesis of both. To improve robustness, the identifier library stores multiple versions of fingerprints and records the version number and expiration time; the latest valid version is returned first during online retrieval.

[0153] Weighted tensor superposition is used to fuse the content latent vector and the identifier vector. Let the content latent vector be c_t, the identifier vector be s_t, and the fused output be... The fusion weights are given by the traceability strength control variable within the updated generative model, using two weights α and β, where α + β = 1. Initial values ​​are calibrated during training and can be adaptively fine-tuned online based on context entropy or quality score. The fusion calculation is as follows: Where g is a combination of orthogonal projection and bandpass filtering of the identifier vector, ensuring that the identifier is only injected into the subspace orthogonal to the semantics or weakly coupled, reducing disturbance to the content representation. To avoid energy spikes, Layer normalization and range pruning are performed, with the pruning threshold adaptively determined based on channel variance. The fusion process is executed time-by-time along the sequence dimension, maintaining temporal alignment.

[0154] The concatenation stage combines the traceable content latent vector and the position latent vector into a unified intermediate representation. This is implemented by performing a join operation on the channel dimension, resulting in... Where p_t is the latent position vector. To improve cross-modal consistency, p_t is first projected onto a linear mapping h. The statistical scales are matched and then concatenated. The concatenated result is input into the subsequent sub-networks of the updated generative model for feature integration. These sub-networks can be U-Net, a diffuse noise predictor, or a gated feedforward integration layer. The z_t is then re-modulated using a cross-attention textual hidden sequence, outputting an updated intermediate feature representation containing traceable features. This output is of equal length to the input in the temporal dimension and includes three parts in the channel dimension: a semantic subspace, a location subspace, and an identifier subspace. To facilitate downstream acoustic mapping and upstream extractability evaluation, optional extractability metrics are also output, such as identifier energy level, subspace orthogonality, and threshold margin.

[0155] This embodiment introduces dynamic sampling between discretized text representation and vector quantization embedding to achieve an adaptive trade-off between computational overhead and uncertainty, stably generating content latent vectors and location latent vectors while maintaining temporal alignment. Using a text hash-driven identifier library retrieval, source-associated identifier fingerprints are mapped to a subspace isomorphic to the content latent vectors and injected using intensity-controlled weighted tensor superposition. Orthogonal projection and energy pruning limit perturbations to the semantic channels. Through channel-level concatenation with the location latent vectors and subsequent cross-attention integration, an updated intermediate feature representation containing traceable source features is formed. Thus, stable, searchable, and weakly perceptible source signals are embedded without altering the temporal structure and semantic consistency, providing an extractable intermediate carrier for subsequent acoustic mapping and discriminative training, while keeping inference latency and memory usage within a controllable range.

[0156] In one embodiment, step S50 above includes:

[0157] S501, through the acoustic feature conversion module of the generated model, the updated intermediate feature representation is mapped to the Mel spectrum feature space, and traceability intensity parameters are injected to generate a modulation acoustic feature vector.

[0158] S502, apply cepstral mean subtraction to the modulated acoustic feature vector to eliminate channel distortion and generate a normalized acoustic feature vector;

[0159] S503, through the neural vocoder module of the generative model, training speech data is generated based on the normalized acoustic feature vector;

[0160] S504 applies a dynamic sampling module to perform Mel spectrum normalization on real speech data to generate standardized real data.

[0161] S505, the training speech data and standardized real data are input into the multi-scale convolutional discriminant network of the discriminant model, and the convolutional kernel weight parameters of the discriminant model are jointly optimized by spectral consistency loss and adversarial loss to generate an updated discriminant model.

[0162] In this embodiment, updating the intermediate feature representation first involves performing time-frequency mapping in the acoustic feature transformation module within the generative model. This module receives the time-aligned hidden sequence z(t) and obtains the acoustic basis h(t) through one-dimensional convolutional downsampling and gated activation. Subsequently, it generates Mel spectral features through Mel filter banks and logarithmic amplitude operations. Where F is the number of Mel bands and T is the number of frames. The center frequencies of the Mel filter bank are distributed according to equidistant Mel scales. Front-end pre-emphasis and Hamming window are used to improve the high-frequency signal-to-noise ratio and reduce leakage. After mapping, traceability intensity parameters are injected into the feature layer by performing weighted injection on each frame vector. Here, u is a fingerprint vector derived from implicit identifier resources. After orthogonal projection and bandpass mask G(·), it occupies only the subband with weak semantic coupling. α+β=1, and β is given by the traceability intensity parameter, which can be adaptively fine-tuned according to the intra-batch distribution to maintain stable average energy. To avoid energy peaks after injection, frame-level normalization and amplitude clipping are performed immediately after injection. The clipping threshold is adaptively set according to the channel variance, thus obtaining the modulated acoustic feature vector.

[0163] The modulated acoustic eigenvectors are then subjected to cepstral mean subtraction, aiming to eliminate channel effects and unify channel statistics. This is implemented by calculating the time-domain mean for each Mel dimension. Execute again When necessary, a moving average is superimposed to handle non-stationary environments, and variance normalization and an ε-stability term are performed on the frame axis to protect numerical safety. The normalized matrix forms a normalized acoustic feature vector, with its dimensions and temporal layout remaining unchanged, facilitating subsequent waveform synthesis.

[0164] The neural vocoder module generates training speech data using normalized acoustic feature vectors as conditional input. Implementation can employ an autoregressive WaveRNN or a HiFi-GAN / BigVGAN structure that is friendly to anti-aliasing upsampling and multi-period discriminators. The conditional upsampling network interpolates features to a step size matching the audio sampling rate, and then synthesizes a time-domain waveform x(n) through multi-scale upsampling blocks and residual blocks. To ensure the fingerprint remains extractable in the waveform domain, the conditional features are injected into the modulation branch of the vocoder through feature gating rather than direct addition, and an upper limit is introduced on the gating coefficients to limit perturbations to the excitation distribution. When concatenating the derived speech segments, a cross-segment overlapping window is used to avoid discontinuous seams. The generated result is the training speech data, with the sampling rate and bit depth aligned with the real data.

[0165] Real speech data enters a standardization channel to construct comparable training pairs. The dynamic sampling module first segments long samples based on duration, signal-to-noise ratio, and information entropy metrics, removing silent intervals. Then, it performs the same Mel transform and cepstral mean subtraction to obtain standardized real data with amplitude and statistical distribution consistent with the training speech data. The dynamic sampling strategy adjusts the number of segments and frame coverage according to the computational budget in each training round, ensuring that real and synthetic samples match in temporal and frequency coverage, reducing discrimination bias.

[0166] Training speech data and standardized real data are fed into a multi-scale convolutional discriminant network of the discriminant model. This network includes parallel receptive field branches, with convolutional kernel sizes combining short-term details and long-term envelopes. Branch outputs are concatenated in the channel dimension and aggregated through a spectral domain attention layer, finally mapping to real-pseudo probabilities. The training objective uses a joint loss. The spectral consistency loss is calculated in the Mel domain using L1, i.e., L_spec = ||M(x_real) - M(x)||_1, used to constrain the generated distribution to approximate the real distribution. The adversarial loss uses binary cross-entropy, with real sample labels as 1 and synthetic sample labels as 0. Let the discriminant output be D(·), then L_adv = -[logD(x_real) + log(1-D(x))]. The total loss L = λ_spec·L_spec + λ_adv·L_adv, with weights λ_spec and λ_adv set based on a validation set stability grid search. Backpropagation updates only the convolutional kernel weights on the discriminant side; gradient clipping restricts the norm to prevent unstable oscillations; batch normalization or spectral normalization is used for convergence control. After training, the updated discriminant model is exported, retaining the forward paths and threshold configurations necessary for inference.

[0167] This embodiment injects a fingerprint with controllable strength into the Mel domain and performs cepstral mean subtraction after injection. The fingerprint energy is limited to a subband that is weakly coupled with semantics, and the overall statistics are consistent with the real distribution. When the vocoder generates training speech data using gated modulation, it can retain extractability while reducing perceptible artifacts. After dynamic sampling and isomorphic normalization, the real data is aligned with the synthetic data in terms of time-frequency coverage and statistics. The multi-scale convolutional discriminant network obtains a robust discrimination boundary under joint spectral consistency and adversarial objectives. The parameter update of the discriminant side depends on stable gradients and regularization constraints, and finally obtains an updated discriminant model that is sensitive to synthetic data and maintains generalization to the real speech distribution. In this way, when facing unknown inputs in the subsequent inference stage, it can output reliable true and false probabilities and provide a feature base with a consistent distribution for source tracing detection, while balancing the computational cost and convergence stability during the training process.

[0168] In one embodiment, step S60 above includes:

[0169] S601, The input speech data is processed by a dynamic sampling module to reduce noise and segment it to generate a preprocessed speech frame;

[0170] S602, perform Mel spectrum conversion and cepstral mean subtraction normalization on the preprocessed speech frame to generate a detection normalized acoustic feature vector;

[0171] S603, the detected normalized acoustic feature vector is input into the multi-scale convolutional discriminant network of the updated discriminant model to perform forward inference, obtain the original discriminant score vector, and obtain the discriminant confidence scalar for the category generated by the generative model;

[0172] S604, in the implicit source identification library of the updated generative model, match the identifier vector that is closest to the detected normalized acoustic feature vector, and generate a source confidence scalar based on the identifier vector through weighted similarity processing;

[0173] S605, if the discrimination confidence scalar is greater than the preset discrimination threshold and the source tracing confidence scalar is greater than the preset source tracing threshold, then it is determined that the input speech data is generated by the generation model; otherwise, it is determined that the input speech data is real data.

[0174] S606, apply a preset hash function to the matched identifier vector to generate a source identifier, and output the detection result containing the data source type of the input voice data and the source identifier.

[0175] In this embodiment, the input speech data first enters the dynamic sampling module for noise reduction and segmentation. This module performs endpoint detection and energy thresholding to remove silent frames in the time domain, uses spectral subtraction or least mean square error estimation to suppress steady-state noise, and employs cepstral smoothing in the frequency domain to suppress reverberation trails. It also adaptively sets the frame shift and frame length based on duration and signal-to-noise ratio (SNR) to ensure that subsequent spectral estimation covers complete semantic units. The output is a preprocessed speech frame with a timestamp and SNR indicator for easy backfilling in subsequent decision-making processes.

[0176] The preprocessed speech frames are then converted to Mel spectra and normalized using cepstral mean subtraction to generate detection-normalized acoustic feature vectors. The Mel transformation employs pre-emphasis, framing, windowing, Fast Fourier Transform, and Mel filter bank projection, outputting an F×T matrix with a logarithmic magnitude scale. Cepstral mean subtraction calculates and centers the time mean in each Mel dimension, adding variance normalization and ε-stabilization terms as needed to ensure statistical consistency across recording devices and channel conditions. The resulting detection-normalized acoustic feature vectors are distributed in alignment with the features from the training phase, reducing domain offset.

[0177] A multi-scale convolutional discriminant network within the updated discriminant model is constructed by inputting the normalized acoustic feature vector. Forward inference is performed to obtain the original discriminant score vector and the discriminant confidence scalar corresponding to the categories generated by the generative model. The network includes parallel receptive field branches and cross-scale channel attention aggregation. Short kernel branches cover microstructure artifacts, while long kernel branches cover resonance and rhythmic patterns. The output layer uses sigmoid or softmax mapping to generate the true / false distribution for each segment or frame. The original discriminant score vector is preserved along the time dimension, and the segment-level discriminant confidence scalar is obtained through temporal pooling and temperature calibration. Temperature is fitted using maximum likelihood estimation on the validation set to improve probability calibration and avoid overconfidence.

[0178] The detected normalized acoustic feature vectors are also used for source retrieval. Nearest neighbor matching is performed in the implicit source identifier library of the updated generative model to obtain the nearest identifier vector. Library retrieval can employ vector quantization and inverted multi-indexing to improve the efficiency of large-scale retrieval. The distance metric can be cosine or Euclidean norm, and the vectors are L2 normalized before retrieval. After obtaining candidates, weighted similarity processing is performed to generate a source confidence scalar. The weighted similarity s_w is formed by a linear combination of multiple similarity components, including frequency domain correlation coefficient, spectral envelope similarity, latent space projection similarity, and necessary temporal elastic matching components. The weight vector is set through training set grid search or Bayesian optimization. To reduce mismatches, a background contrast term is added, and a quantile threshold is applied to the set of non-matching vectors to ensure that s_w is separated from the background distribution.

[0179] Binary joint decision-making is implemented within a threshold space. The discrimination confidence scalar and the source tracing confidence scalar are compared with preset discrimination thresholds and preset source tracing thresholds, respectively. If both conditions are met simultaneously, the output is "Generated by the generative model"; otherwise, "True data" is output. Thresholds can be selected from error points in the development set based on ROC or DET curves, or the optimal trade-off point can be determined based on the cost of missed detections and the cost of false positives from the business side. If the input is a long audio file, segmented decision-making is used, followed by majority voting or weighted voting to summarize the results at the file level. The weights can be the segment length or the average confidence level within each segment.

[0180] To create an auditable identifier, the matched identifier vector is mapped to a source identifier using a predefined hash function. The hash function can be SHA-256 or BLAKE2. The input is a byte sequence resulting from the identifier vector after fixed-point quantization and salt concatenation. The output is encoded as a hexadecimal or Base32 string. The quantization bit width and salt length are determined through collision probability analysis to ensure discriminability between different sessions or different generation batches. The final detection result includes the data source type and source identifier, and can also include timestamps, model version, and confidence level metadata for auditing purposes.

[0181] Example Description: In the healthcare field, when building a voice source tracing and authenticity determination system for a remote consultation platform, the system first deploys a deep Transformer backbone network with multi-layer attention mechanisms. A flow-based generative model is connected at its top layer to generate initial intermediate feature representations from input medical record text, electronic health records, and doctors' verbal diagnostic instructions. A text-audio alignment module is integrated into the generation path to ensure accurate temporal alignment of medical terminology. A dynamic sampling mechanism is deployed on this module to reduce GPU computation peaks and avoid delays when generating lengthy verbal medical case descriptions. Simultaneously, a context-aware feature extractor is built to capture detailed descriptions of symptoms in medical speech. This, combined with a lightweight convolutional classification network and a multi-scale feature capture mechanism, forms the front end of the discrimination model for subsequent authenticity determination and source tracing verification.

[0182] Next, using the initial intermediate feature representations output by the generative model, Mel-spectral features are generated through the acoustic feature conversion module, and then converted into a time-domain speech signal through the waveform reconstruction module, resulting in simulated speech data containing medical terminology and descriptions of symptoms. To ensure the discriminative stability of the medical data, signal normalization is performed on the speech to eliminate amplitude shifts caused by differences in recording equipment. The optimized speech data is input into the context-aware feature extractor of the discriminative model to extract multi-scale acoustic features, and then a lightweight convolutional classification network outputs a ground truth probability distribution, which is compared with the true labels to calculate the binary cross-entropy loss, thereby generating the discriminative loss.

[0183] Based on discriminative loss, the gradient tensor of each feature layer of the generative model is determined through a backpropagation module. Gradient information is extracted to adjust the traceability strength parameters of the feature embedding space, and the weight matrix is ​​modulated to embed implicit traceability identifiers. For example, in the scenario of generating a doctor's medical record, the unique code of the medical institution can be used as part of the implicit identifier. The modified embedding space is fused with the modulated weight matrix using weighted tensor superposition to form a prototype of the traceable intermediate feature representation. The adaptive optimization module updates the low-rank parameter matrix of the generative model accordingly, obtaining the optimized parameter set, which is then loaded back into the generative model to form an updated version of the generative engine.

[0184] In the subsequent generation of new traceable intermediate feature representations, the medical record text input by the doctor is fed into the updated Transformer encoder. The output text index string drives the quantized variational autoencoder to sample in the embedding space, generating content latent vectors and location latent vectors. The dynamic sampling module controls resource consumption to avoid system overload during peak consultation periods. Based on the hash value of the text data, the corresponding identifier vector is retrieved from the implicit traceability identifier library, and it is weighted and fused with the content latent vector to generate a traceable content latent vector. This content latent vector is then concatenated with the location latent vector to generate an intermediate representation containing traceable features.

[0185] The intermediate representation is input into the acoustic feature transformation module of the generation model and mapped to the Mel spectrum. A traceability intensity parameter is then injected to generate a modulated acoustic feature vector. After applying cepstral mean subtraction to eliminate vocal tract distortion, a normalized acoustic feature vector is obtained, which is then used by a neural vocoder to generate medical speech data for training. Real medical speech data (e.g., actual recordings of conversations between patients and doctors) is normalized using a dynamic sampling module and the Mel spectrum to obtain standardized real data. Both are then input into the multi-scale convolutional discriminant network of the discriminant model. The convolutional kernel weights are jointly optimized using spectral consistency loss and adversarial loss, resulting in an updated discriminant model with traceability detection capabilities in the medical field.

[0186] In real-time consultations, upon receiving input voice data (which could be diagnostic voice from a doctor or potentially tampered voice), the dynamic sampling module first performs noise reduction and segmentation to ensure the integrity of medical orders and prescriptions is not obscured by noise. The preprocessed voice frames are normalized using Mel-frequency conversion and cepstral mean subtraction to generate a detection-normalized acoustic feature vector. This vector is then fed into the multi-scale convolutional discriminant network of the discriminant model for forward inference, yielding the original discriminant score vector and a discriminant confidence scalar for the category generated by the generative model. The system matches the closest identifier vector in the implicit source identifier library and generates a source confidence scalar through weighted similarity calculation. If both the discriminant confidence and source confidence are higher than preset thresholds, the voice is considered to have been generated by the generative model. A preset hash function is used to map the identifier vector to an irreversible source identifier, outputting a detection result containing the data source type and the source identifier. This result is used to generate a security label or archive record for the consultation report.

[0187] In the fintech business, it is deployed in voice interaction scenarios such as intelligent customer service, online transaction risk control, and compliance auditing to enable the identification of the authenticity of voice data and trace its source, thereby providing security in anti-fraud, transaction confirmation, and voice contract execution.

[0188] In the intelligent customer service and transaction confirmation system, a deep Transformer backbone network with a multi-layer attention mechanism is first constructed, and a stream-based generative model is connected on top of it to convert the input business text (such as transfer instructions, transaction confirmation scripts, compliance risk warnings, etc.) into initial intermediate feature representations. A text-audio alignment module is integrated into the generation path to ensure the precise temporal and acoustic rhythmic correspondence of amounts, account information, and conditions specific to financial transactions. Dynamic sampling is deployed in this module to adapt to high-concurrency customer service environments and balance real-time response with computational resource consumption. Simultaneously, a context-aware feature extractor is constructed to capture key information points in financial speech, such as amount, transaction time, and contract number, and is combined with a lightweight convolutional classification network to form a discriminative model that can be used for subsequent authenticity detection using a multi-scale feature capture mechanism.

[0189] During the model training phase, the initial intermediate feature representations generated from the model output are input into the acoustic feature conversion module to generate Mel-spectral features. These features are then processed by the waveform reconstruction module to obtain the time-domain speech signal, forming synthetic speech data containing transaction confirmation content or customer service scripts. To ensure the accuracy of subsequent discrimination results, the speech is normalized to eliminate amplitude shifts introduced by different recording devices or transmission paths. The optimized speech data is then input into the context-aware feature extractor of the discrimination model to extract multi-scale acoustic features. A lightweight convolutional classification network outputs a ground truth probability distribution, and the binary cross-entropy loss is calculated based on the difference between the distribution and the ground truth labels to obtain the discrimination loss.

[0190] Based on discriminative loss, the gradient tensor of the feature layer of the generative model is determined through backpropagation. The gradient information is then used to adjust the traceability strength parameters of the feature embedding space, modulating the weight matrix to embed implicit traceability identifiers. In financial applications, these identifiers can include information such as internal bank transaction serial numbers and transaction channel codes for subsequent traceability comparison. The modified feature embedding space is fused with the modulated weight matrix using weighted tensor superposition to generate a prototype traceable intermediate feature representation. An adaptive optimization module then updates the low-rank parameter matrix, forming an optimized model parameter set, which is loaded back into the generative model to create an updated version.

[0191] When generating an updated intermediate feature representation with traceable characteristics, the input transaction text instruction is fed into the updated Transformer encoder to generate a text index string and drive the quantized variational autoencoder to sample in the vector embedding space, outputting content latent vectors and position latent vectors. The dynamic sampling module controls the computational load during periods of high transaction volume to ensure manageable response latency. Based on the hash value of the input text, matching identifier vectors are retrieved from the implicit traceability identifier library and weighted and fused with the content latent vector to generate a traceable content latent vector. This content latent vector is then concatenated with the position latent vector to form an updated intermediate feature representation containing traceable characteristics.

[0192] The intermediate layer represents the input acoustic feature transformation module mapped to the Mel spectrum and injected with traceability intensity parameters to generate a modulated acoustic feature vector. After eliminating vocal tract distortion through cepstral mean subtraction, a normalized acoustic feature vector is obtained, which is then used by the neural vocoder module to generate training speech data. Simultaneously, collected real financial business speech (such as genuine customer confirmation recordings) undergoes dynamic sampling and Mel spectrum normalization to obtain standardized real data. Both types of data are input into the multi-scale convolutional discriminant network of the discriminant model. The convolutional kernel weights are jointly optimized through spectral consistency loss and adversarial loss to form an updated discriminant model capable of identifying the authenticity and tracing the source of financial speech.

[0193] In actual operation, when the system receives input voice data (which may come from customer transaction confirmation or be potentially fraudulent voice), the dynamic sampling module first performs noise reduction and segmentation processing to ensure the acoustic representation of key fields such as amount and date is complete. The processed voice frame undergoes Mel-spectrum transformation and cepstral mean subtraction normalization to obtain a detection-normalized acoustic feature vector, which is then input into the discriminant model for forward inference, outputting the original discriminant score vector and a discriminant confidence scalar for the generative model category. Subsequently, the closest identifier vector is matched in the implicit source identification library, and a source identification confidence scalar is obtained through weighted similarity calculation. If both the discriminant confidence and source identification confidence exceed preset thresholds, the voice is considered to have been generated by the generative model, and the identifier vector is mapped to a source identification using a preset hash function. Finally, the detection result, containing the data source type and source identification, is output as a security credential for transaction confirmation or subsequent auditing.

[0194] This embodiment uses standardized dynamic sampling and Mel domain normalization to ensure that the normalized acoustic feature vector is consistent with the training distribution, reducing discrimination degradation caused by domain offset. The multi-scale convolutional discriminant network simultaneously covers short-term artifacts and long-term resonance structures in forward inference, and the original discrimination score vector and the calibrated discrimination confidence scalar provide stable true and false probabilities. The nearest neighbor matching and weighted similarity processing of the implicit source tracing identifier library generate the source tracing confidence scalar, which robustly verifies the embedded fingerprint in multiple channels in the frequency domain, latent space, and time series. The dual threshold joint judgment simultaneously constrains the true / false discrimination and source tracing consistency, thereby reducing false positives and false negatives. The hash mapping of the matching identifier vector is performed to obtain the source tracing identifier, ensuring that the output is verifiable, irreversible, and traceable.

[0195] In one embodiment, a data tracing device based on collaborative training is provided, which corresponds one-to-one with the data tracing method based on collaborative training described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the data tracing device based on collaborative training of the present invention. The modules include: Context Generation Module 10, Multimodal Content Parsing Module 20, Feature Extraction Module 30, Source Attribute Matching Module 40, Processing Instruction Generation Module 50, and Feedback Update Module 60. Detailed descriptions of each functional module are as follows:

[0196] The attachment context generation module 10 is used to receive the original attachment data, extract the sender attribute, transmission path attribute and carrier attribute from the transmission carrier, generate attachment context data, and associate the attachment context data with the original attachment data.

[0197] The multimodal content parsing module 20 is used to perform multimodal content parsing on the original attachment data through a content parsing model to generate attachment parsing data;

[0198] Feature extraction module 30 is used to extract feature elements based on the parsed attachment data and generate an attachment feature set;

[0199] The source attribute matching module 40 is used to generate a source attribute based on the attachment context data, and to input the source attribute and the attachment feature set into the anomaly tag library for matching to generate anomaly level data.

[0200] The disposal instruction generation module 50 is used to generate disposal instructions based on the anomaly level data.

[0201] The feedback update module 60 is used to execute the processing instructions, collect feedback data after execution, and update the abnormal tag library and content parsing model.

[0202] In one embodiment, the attachment context generation module 10 is specifically used for:

[0203] Receive raw attachment data containing attachment file data and transmission medium data through the transmission interface;

[0204] The sender identification information, transmission path identification information, and carrier format information are extracted from the transmission carrier data;

[0205] Based on the sender identification information, the transmission path identification information, and the carrier format information, sender attributes, transmission path attributes, and carrier attributes are generated respectively.

[0206] Aggregate the sender attribute, the transmission path attribute, and the carrier attribute to generate attachment context data;

[0207] Establish a binding relationship between the attachment context data and the attachment original data.

[0208] In one embodiment, the multimodal content parsing module 20 is specifically used for:

[0209] The content parsing model is used to identify the type of the original data in the attachments, distinguishing between text content, image and video content, and compressed or executable content.

[0210] Perform optical character recognition (OCR) on the text content to extract the text content;

[0211] Extract keyframes from the image and video content, and identify visual elements in the keyframes;

[0212] The compressed or executable content is subjected to sandbox environment behavior simulation operations to detect behavioral signs;

[0213] The text content, visual elements, and behavioral indicators are integrated to generate attachment parsing data.

[0214] In one embodiment, the feature extraction module 30 is specifically used for:

[0215] Extract text elements, including semantically anomalous entities and keywords, from the parsed data of the attachments;

[0216] Visual elements, including image text and graphic features, are extracted from the parsed data of the attachments;

[0217] Extract behavioral indicators, including network requests and file system operations, from the parsed attachment data;

[0218] Extract metadata elements, including file attributes and structural information, from the parsed attachment data;

[0219] The text elements, visual elements, behavioral indicators, and metadata elements are aggregated to generate an attachment feature set.

[0220] In one embodiment, the source attribute matching module 40 is specifically used for:

[0221] Extract the sender identification information and transmission path identification information from the attachment context data;

[0222] A sender reputation score is generated based on the sender identification information;

[0223] A transmission path anomaly level is generated based on the transmission path identification information;

[0224] The source attribute is generated by combining the sender's reputation score and the transmission path anomaly level.

[0225] Retrieve personalized and global anomaly tags associated with the attachment feature set from the anomaly tag library;

[0226] Perform multi-level matching operations between the source attribute and the personalized anomaly tag and the global anomaly tag;

[0227] Based on the matching results, high, medium, or low anomaly levels are generated as anomaly level data.

[0228] In one embodiment, the disposal instruction generation module 50 is specifically used for:

[0229] Determine whether the abnormality level data belongs to a high abnormality level, a medium abnormality level, or a low abnormality level.

[0230] When the anomaly level data is high anomaly level, an isolation command and a verification command are generated;

[0231] When the anomaly level data is of medium anomaly level, a labeling instruction is generated;

[0232] When the anomaly level data is low anomaly level, a normal processing instruction is generated;

[0233] A disposal instruction is generated based on at least one of the isolation instruction, the verification instruction, the labeling instruction, and the routine processing instruction.

[0234] In one embodiment, the feedback update module 60 is specifically used for:

[0235] Execute the isolation instructions, verification instructions, labeling instructions, or routine processing instructions included in the disposal instructions to obtain the disposal result;

[0236] Feedback data on the processing results is received through an interactive interface;

[0237] The tag weight values ​​in the abnormal tag library are dynamically adjusted based on the feedback data.

[0238] The classification parameters of the content parsing model are optimized based on the feedback data;

[0239] When the feedback data contains a misjudgment identifier, the corresponding attachment is transferred to the review terminal, and the abnormal features supplemented by the review terminal are received. The training data of the content parsing model is optimized based on the abnormal features supplemented by the review terminal.

[0240] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external user terminals via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a data tracing method based on collaborative training on the server side.

[0241] In one embodiment, a computer device is provided, which may be a user terminal, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a data tracing method based on collaborative training on the user side.

[0242] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0243] Receive the original attachment data, extract the sender attribute, transmission path attribute and carrier attribute from the transmission carrier, generate attachment context data, and associate the attachment context data with the original attachment data;

[0244] The original attachment data is parsed using a content parsing model to generate parsed attachment data.

[0245] Based on the parsed data of the attachments, feature elements are extracted to generate an attachment feature set;

[0246] Source attributes are generated based on the attachment context data, and the source attributes are matched with the attachment feature set and input into the anomaly tag library to generate anomaly level data;

[0247] A handling instruction is generated based on the anomaly level data;

[0248] Execute the processing instructions, collect feedback data after execution, and update the anomaly tag library and content parsing model.

[0249] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0250] Receive the original attachment data, extract the sender attribute, transmission path attribute and carrier attribute from the transmission carrier, generate attachment context data, and associate the attachment context data with the original attachment data;

[0251] The original attachment data is parsed using a content parsing model to generate parsed attachment data.

[0252] Based on the parsed data of the attachments, feature elements are extracted to generate an attachment feature set;

[0253] Source attributes are generated based on the attachment context data, and the source attributes are matched with the attachment feature set and input into the anomaly tag library to generate anomaly level data;

[0254] A handling instruction is generated based on the anomaly level data;

[0255] Execute the processing instructions, collect feedback data after execution, and update the anomaly tag library and content parsing model.

[0256] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and user side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0257] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application 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. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0258] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0259] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The embodiments described above are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A data tracing method based on collaborative training, characterized in that, Includes the following steps: Obtain the generative model and the discriminative model, and output the initial intermediate feature representation through the generative model; Based on the initial intermediate feature representation, speech data is generated through acoustic conversion, and the speech data is input into the discrimination model to generate discrimination loss; The generative model is updated based on the discriminative loss so that the generative model outputs a traceable intermediate feature representation, thus obtaining the updated generative model; The updated generative model is used to generate an updated intermediate feature representation containing traceable features; Based on the updated intermediate feature representation, training speech data is generated through acoustic conversion, and the training speech data and real data are input into the discrimination model for training to obtain the updated discrimination model. Upon receiving input speech data, the updated discrimination model is used to discriminate the input speech data, and a detection result is output regarding whether the input speech data was generated by the generation model.

2. The data tracing method based on collaborative training as described in claim 1, characterized in that, Obtaining a generative model and a discriminative model, and outputting an initial intermediate feature representation through the generative model, including: Construct a deep Transformer backbone network configured with a multi-layer attention mechanism; Flow-based models are connected on the deep Transformer backbone network to form feature generation paths; Integrate a text-audio alignment module into the feature generation path; Deploy a dynamic sampling module on the text-audio alignment module; The deep Transformer backbone network, the flow-based model, the text-audio alignment module, and the dynamic sampling module are integrated to form a complete generative model; Construct a context-aware feature extractor; Configure a lightweight convolutional classifier network and set up a multi-scale feature capture mechanism; The context-aware feature extractor is integrated with a lightweight convolutional classification network to form a complete discriminative model. The complete generative model processes the input text data and outputs an initial intermediate feature representation.

3. The data tracing method based on collaborative training as described in claim 1, characterized in that, Based on the initial intermediate feature representation, speech data is generated through acoustic conversion, and the speech data is input into the discrimination model to generate a discrimination loss, including: Based on the initial intermediate feature representation, Mel spectrum features are generated through the acoustic feature conversion module; The waveform reconstruction module converts the Mel spectrum features into a time-domain speech signal to generate speech data. The speech data is subjected to signal normalization processing to generate optimized speech data; The optimized speech data is input into the context-aware feature extractor of the discrimination model to extract multi-scale acoustic features; The multi-scale acoustic features are analyzed using a lightweight convolutional classification network of the discriminant model to obtain the probability distribution of speech authenticity. Based on the difference between the probability distribution of speech authenticity and the real labels, the binary cross-entropy loss is determined, and the discriminant loss is generated.

4. The data tracing method based on collaborative training as described in claim 1, characterized in that, The generative model is updated based on the discriminative loss so that it outputs a traceable intermediate feature representation, resulting in the updated generative model, including: Based on the discriminative loss, the gradient tensor of the generated model is determined through the backpropagation module, and the gradient information of the feature layer is generated. Based on the gradient information of the feature layer, the traceability strength parameter of the feature embedding space of the generative model is modified, and the weight matrix of the generative model is modulated to embed implicit traceability identifiers. By superimposing and fusing the modified feature embedding space with the modulated weight matrix using weighted tensors, a traceable intermediate feature representation prototype is generated. Based on the traceable intermediate feature representation prototype, the low-rank parameter matrix of the generated model is updated using an adaptive optimization module to generate an optimized model parameter set. The optimized model parameter set is loaded into the generated model architecture to generate an updated generated model.

5. The data tracing method based on collaborative training as described in claim 1, characterized in that, The updated generative model is used to generate an updated intermediate feature representation containing traceable features, including: The input text data is fed into the updated generative model's Transformer encoder to generate a text index string; Based on the text index string, sampling is performed in the vector embedding space of the quantized variational autoencoder of the updated generative model, and a dynamic sampling module is applied to control resource consumption to generate content latent vectors and position latent vectors. Based on the hash value of the input text data, the identifier vector is matched from the implicit traceability identifier library of the updated generative model, and the content latent vector and the identifier vector are fused by weighted tensor superposition to generate a traceable content latent vector. The traceable content latent vector and the location latent vector are concatenated, and the concatenation result is input into the updated generative model to generate an updated intermediate feature representation containing traceable features.

6. The data tracing method based on collaborative training as described in claim 1, characterized in that, Based on the updated intermediate feature representation, training speech data is generated through acoustic conversion, and the training speech data and real data are input into the discrimination model for training to obtain the updated discrimination model, including: The acoustic feature transformation module of the generative model maps the updated intermediate feature representation to the Mel spectral feature space and injects traceability intensity parameters to generate a modulation acoustic feature vector. Applying cepstral mean subtraction to the modulated acoustic feature vector eliminates channel distortion and generates a normalized acoustic feature vector; Training speech data is generated based on the normalized acoustic feature vector through the neural vocoder module of the generative model. The real speech data is processed by applying a dynamic sampling module to perform Mel spectrum normalization to generate standardized real data. The training speech data and standardized real data are input into the multi-scale convolutional discriminant network of the discriminant model. The convolutional kernel weight parameters of the discriminant model are jointly optimized by spectral consistency loss and adversarial loss to generate an updated discriminant model.

7. The data tracing method based on collaborative training as described in claim 1, characterized in that, Upon receiving input speech data, the updated discrimination model is used to discriminate the input speech data, and a detection result is output regarding whether the input speech data was generated by the generation model, including: The input speech data is subjected to noise reduction and segmentation processing using a dynamic sampling module to generate preprocessed speech frames; The preprocessed speech frame is subjected to Mel-spectrum transformation and cepstral mean subtraction normalization to generate a detection normalized acoustic feature vector; The detected normalized acoustic feature vector is input into the multi-scale convolutional discriminant network of the updated discriminant model for forward inference to obtain the original discriminant score vector and the discriminant confidence scalar for the category generated by the generative model. In the implicit source identification library of the updated generative model, the identifier vector that is closest to the detected normalized acoustic feature vector is matched, and a source confidence scalar is generated based on the identifier vector through weighted similarity processing. If the discrimination confidence scalar is greater than the preset discrimination threshold and the source tracing confidence scalar is greater than the preset source tracing threshold, then the input speech data is determined to be generated by the generation model; otherwise, the input speech data is determined to be real data. A preset hash function is applied to the matched identifier vector to generate a source identifier, and the detection result containing the data source type of the input voice data and the source identifier is output.

8. A data tracing device based on collaborative training, characterized in that, The data tracing device based on collaborative training includes: The generative model building module is used to acquire the generative model and the discriminative model, and to output the initial intermediate feature representation through the generative model; An acoustic feature conversion module is used to generate speech data based on the initial intermediate feature representation through acoustic conversion, and input the speech data into the discrimination model to generate discrimination loss; The generative model update module is used to update the generative model based on the discriminative loss so that the generative model outputs a traceable intermediate feature representation, thereby obtaining the updated generative model. The traceable feature generation module is used to generate an updated intermediate feature representation containing traceable features using the updated generation model; The discriminant model update module is used to generate training speech data through acoustic conversion based on the updated intermediate feature representation, and input the training speech data and real data into the discriminant model for training to obtain the updated discriminant model. The speech data discrimination module is used to discriminate the input speech data using the updated discrimination model when the input speech data is received, and output a detection result type of whether the input speech data was generated by the generation model.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a collaborative training-based data tracing program stored in the memory and executable on the processor. When executed by the processor, the collaborative training-based data tracing program implements the steps of the collaborative training-based data tracing method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a data tracing program based on collaborative training, which, when executed by a processor, implements the steps of the data tracing method based on collaborative training as described in any one of claims 1-7.

Images