Speaker log method, device and equipment for closed-loop learning and storage medium

Abstract

The invention provides a speaker log method and device for closed-loop learning, equipment and a storage medium. The method comprises the following steps: generating an initial recognition result with confidence through a speaker log model; a composite uncertainty sampling strategy is adopted, and audio clips which are most required to be manually checked are intelligently screened out to obtain high-quality feedback; a speaker prototype memory bank is updated in real time by using the feedback, and low-delay registration and online adaptation of a new speaker are realized; meanwhile, a reliable pseudo label is generated from high-confidence prediction based on a dynamically adjusted reliability threshold, and periodic fine tuning is performed on the model in combination with feedback data. According to the method, a closed loop of'active learning-real-time adaptation-self-learning optimization 'is constructed, the labeling cost is remarkably reduced, and the generalization ability and the real-time interaction performance of the model in an actual scene are improved.

Description

Technical Field

[0001] This invention relates to the field of speech recognition technology, and in particular to a speaker log method, apparatus, device, and storage medium for closed-loop learning. Background Technology

[0002] Speaker digitization is a key task in speech processing, aiming to solve the problem of "who speaks and when," and has wide applications in scenarios such as meeting transcription, customer service analysis, and multimedia content indexing. With the development of deep learning, the end-to-end neural network digitization (EEND) model has become the mainstream solution, significantly improving the ability to process overlapping speech by modeling the task as multi-label sequence classification.

[0003] However, existing technologies, especially EEND-based frameworks, still have several prominent shortcomings that limit their widespread application in real-world scenarios. First, model performance heavily relies on large-scale, finely labeled datasets. Such data is expensive and scarce, leading to insufficient model generalization ability and difficulty adapting to diverse accents, environmental noise, and dialogue styles. Second, existing systems lack effective real-time online adaptability. In dynamic scenarios (such as ongoing meetings or calls), traditional systems cannot achieve low-latency registration and recognition when a new speaker appears, often requiring interruptions for batch processing or retraining, severely impacting the interactive experience. Furthermore, semi-supervised learning (such as pseudo-labeling techniques) and active learning strategies introduced to alleviate data dependence issues have efficiency bottlenecks. The inherent noise in pseudo-labels contaminates the training process, while traditional active learning strategies fail to accurately address the core challenge of speaker log tasks—the blurred speaker switching boundaries—resulting in low cost-effectiveness of manual annotation.

[0004] In summary, the shortcomings of the existing technology urgently need to be addressed. Summary of the Invention

[0005] This invention provides a speaker log method, apparatus, device, and storage medium for closed-loop learning, which addresses the shortcomings of existing technologies, reduces the model's dependence on large-scale pre-labeled data, and improves the model's generalization ability and robustness in diverse real-world environments.

[0006] This invention provides a speaker log method for closed-loop learning, comprising: Acquire the audio data to be processed, and perform speech activity detection and acoustic feature extraction on the audio data to obtain front-end features; The front-end features are received through the speaker log model, and at least one initial recognition result is generated by referring to the speaker prototype embedding obtained from the speaker prototype memory. The initial recognition result includes speaker label and confidence score. Based on the initial identification results, uncertainty analysis is performed on the audio data, and audio segments with uncertainty scores higher than a preset threshold are selected from the audio data. Explicit feedback on the audio segment to be queried is collected to obtain the corrected truth label; Based on the truth labels, the speaker prototype embedding is updated, and the updated speaker prototype embedding is used to perform real-time speaker matching and recognition on subsequent audio data.

[0007] According to the speaker log method for closed-loop learning provided by the present invention, after the step of updating the speaker prototype embedding based on the truth label, the method further includes: Based on the audio data that was not selected as the query segment and its initial recognition results, its confidence level is compared with a reliability threshold, and predictions with a confidence level higher than the reliability threshold are generated as pseudo-labels. The model parameters of the core speaker log model are updated using the truth labels, the pseudo labels, and the original labeled data.

[0008] According to the speaker log method for closed-loop learning provided by the present invention, the step of updating the model parameters of the core speaker log model using the truth labels, the pseudo labels, and the original labeled data specifically includes: Construct a total loss function, which is a weighted sum of a first loss calculated based on the original labeled data, a second loss calculated based on the ground truth labels, and a third loss calculated based on the pseudo labels; The parameters of the speaker log model are updated by minimizing the total loss function using an optimization algorithm.

[0009] According to the speaker log method for closed-loop learning provided by the present invention, the step of performing uncertainty analysis on the audio data based on the initial recognition result and filtering out the query audio segments with uncertainty scores higher than a preset threshold from the audio data specifically includes: Calculate the uncertainty score for each audio time frame in the audio data, where the uncertainty score is a weighted sum of classification entropy and minimum boundary distance; The consecutive time frames whose uncertainty scores exceed a preset segment threshold are aggregated into the audio segment to be queried; The classification entropy is used to measure the overall disorder in the model's assignment of speaker labels, and the minimum boundary distance is used to measure the intensity of competition between the two speakers with the highest predicted probabilities.

[0010] According to the speaker log method for closed-loop learning provided by the present invention, the step of collecting explicit feedback on the audio segment to be queried specifically includes: Collect users' tacit confirmation behavior for high-confidence audio segments that were not selected as query segments, as an implicit positive feedback signal; The implicit positive feedback signal is used to assist in the generation of the pseudo-tag or the adjustment of the reliability threshold.

[0011] According to the present invention, a speaker log method based on closed-loop learning is provided, wherein the speaker log model is used to describe the speaker log task as a multi-label sequence classification problem based on an end-to-end neural network architecture; the first loss, the second loss and the third loss are all calculated using a binary cross-entropy loss trained with permutation invariance.

[0012] According to the speaker log method for closed-loop learning provided by the present invention, the step of updating the speaker prototype embedding based on the truth label specifically includes: Extract the speaker embedding vector from the audio segment corresponding to the truth value label; For speakers already existing in the speaker prototype memory, an exponential moving average mechanism is used to iteratively update their corresponding prototype embedding vectors using the newly extracted speaker embedding vectors. For a newly emerging speaker, its speaker embedding vector is stored as a new prototype embedding vector in the speaker prototype memory.

[0013] The present invention also provides a speaker log device for closed-loop learning, comprising: The audio acquisition module is used to acquire audio data to be processed, and to perform speech activity detection and acoustic feature extraction on the audio data to obtain front-end features; An initial recognition module is used to receive the front-end features through a speaker log model and generate at least one initial recognition result by referring to the speaker prototype embedding obtained from the speaker prototype memory bank. The initial recognition result includes speaker labels and confidence scores. An uncertainty analysis module is used to perform uncertainty analysis on the audio data based on the initial identification results, and to filter out the audio segments to be queried from the audio data whose uncertainty scores are higher than a preset threshold. The explicit feedback module is used to collect explicit feedback on the audio segment to be queried and obtain the corrected truth label. The parameter update module is used to update the speaker prototype embedding based on the truth label, and to use the updated speaker prototype embedding to perform real-time speaker matching and recognition on subsequent audio data.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the speaker log method for closed-loop learning as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the speaker log method for closed-loop learning as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a speaker log method for closed-loop learning as described above.

[0017] The speaker log method, apparatus, device, and storage medium for closed-loop learning provided by this invention bring about the following comprehensive technical improvements by constructing a dynamic feedback loop that integrates active learning, online adaptation, and model iterative optimization: First, it significantly improves the efficiency of human-computer collaboration. The system can intelligently select the most informative segments (especially those with ambiguous speaker transitions) based on the confidence information output by the model and request human intervention. This allows valuable human annotation resources to be accurately invested in the key links that can best correct model errors and improve performance, thereby obtaining the greatest model performance benefits with extremely low annotation costs. Second, this invention achieves true low-latency online adaptation capability. When a user provides feedback on the system's prediction or confirms a new speaker, the system can immediately use this feedback to update the corresponding features in the speaker prototype memory bank and immediately apply them to the real-time recognition and matching of subsequent audio streams. This effectively solves the problem of slow response in traditional methods when a new speaker appears or voice features change, greatly enhancing the practicality of the system in real-time interactive scenarios such as meetings and customer service. Finally, this method forms a self-reinforcing learning loop. User feedback is not only used directly for real-time correction, but also indirectly optimizes the foundation for future recognition by updating the prototype library. It also combines unlabeled high-confidence predictions for self-learning. This design enables the system to continuously evolve in use and accumulate domain knowledge, thereby fundamentally reducing the dependence on large-scale pre-labeled data and improving the model's generalization ability and robustness in diverse real-world environments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the speaker log method for closed-loop learning provided by the present invention. Figure 2 This is a schematic diagram of the speaker log device for closed-loop learning provided by the present invention; Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] To address the problems in existing technologies, this invention proposes a closed-loop learning speaker logging method, which reduces the model's dependence on large-scale pre-labeled data and improves its generalization ability and robustness in diverse real-world environments. The closed-loop learning speaker logging method is described below, as follows: Figure 1 As shown, including but not limited to the following steps: Step 110: Obtain the audio data to be processed, and perform speech activity detection and acoustic feature extraction on the audio data to obtain front-end features.

[0022] In this step, the system receives the raw audio stream or audio file as input. First, the audio data is preprocessed, such as through pre-emphasis, framing, and windowing. Then, Voice Activity Detection (VAD) is performed to filter out non-speech segments (such as silence or background noise), retaining only audio segments containing valid speech. VAD can employ traditional methods based on energy and zero crossover rate, or a neural network-based model. Next, acoustic features are extracted from the speech segments segmented by VAD. In a preferred embodiment, the extracted features are Mel-frequency cepstral coefficients (MFCCs), and the extraction process includes calculating a short-time Fourier transform, passing through a Mel filter bank, taking the logarithm, and then performing a discrete cosine transform. Finally, a front-end feature sequence characterizing the acoustic properties of the audio signal is obtained.

[0023] Step 120: Receive the front-end features through the speaker log model, and generate at least one initial recognition result by referring to the speaker prototype embedding obtained from the speaker prototype memory bank. The initial recognition result includes speaker label and confidence level.

[0024] The speaker log model is preferably based on an end-to-end neural network (EEND) architecture, trained to handle multi-label sequence classification tasks. The model receives the front-end feature sequence extracted in step 110 as its primary input. Simultaneously, the system retrieves the prototype embedding vectors of all currently known speakers from the speaker prototype memory. These prototype embedding vectors are points in a high-dimensional feature space, representing abstract features of a speaker's identity. Internally, the model integrates these prototype embeddings as reference information through an attention mechanism or a specific conditional input layer. The model calculates a probability distribution vector for each time frame, where each element corresponds to the predicted probability that a potential speaker (including a "non-speaking" state) is speaking in that time frame. This probability is the confidence score, and the speaker label with the highest probability constitutes the initial recognition result for that frame. For an audio clip, the model outputs a sequence containing a series of timestamped speaker labels and their confidence scores.

[0025] Step 130: Based on the initial recognition results, perform uncertainty analysis on the audio data, and filter out the audio segments to be queried whose uncertainty scores are higher than a preset threshold from the audio data.

[0026] This step aims to achieve efficient active learning. First, a composite uncertainty score is calculated for each time frame in the audio stream. This score is composed of a weighted sum of two parts: Classification entropy: Calculated based on the probability distribution of all speaker labels output by the model for that time frame. The higher the entropy value, the more confused the model's overall judgment of that frame.

[0027] Minimum boundary distance: Calculated based on the probability values ​​of the two speakers with the highest and second highest predicted probabilities in this time frame. The smaller the distance (i.e. the closer the probabilities of the two speakers are), the more ambiguous the model's decision between the two speakers is. It usually corresponds to the speaker switching point.

[0028] A balanced weighting system is set (e.g., 0.7 for entropy, 0.3 for boundary distance), and a composite score is calculated for each time frame. Then, all consecutive time frames with scores exceeding a preset segment threshold are aggregated into a candidate segment. Finally, from all candidate segments, the top K segments with the highest composite uncertainty scores (e.g., K=10) are selected as the audio segments to be queried and submitted to the user for verification. This method ensures that manual annotation efforts are focused on the areas where the model is most uncertain and most prone to errors.

[0029] Step 140: Collect explicit feedback on the audio segment to be queried to obtain the corrected truth label.

[0030] The system presents the audio segments selected in step 130 to the user through a human-computer interaction interface (such as a visualization tool with audio waveforms and predicted labels). The user can listen to the audio and check the initial recognition results automatically generated by the system. The user can provide explicit feedback by: confirming correct labels, correcting incorrect speaker IDs, and adjusting the start and end time boundaries of speaker activities. After the user submits corrections, the system obtains ground truth labels for these segments, with accurate timestamps and speaker identities. In addition, the system can also monitor user behavior; for example, if a user quickly skips or fails to modify high-confidence segments, this can be considered implicit positive feedback, helping to determine the reliability of the model's predictions.

[0031] Step 150: Update the speaker prototype embedding based on the truth value label, and use the updated speaker prototype embedding to perform real-time speaker matching and recognition on subsequent audio data.

[0032] This step enables online adaptation. First, from the user-corrected audio segment, a pre-trained speaker embedding vector is extracted for that segment using a speaker encoder (such as an x-vector or d-vector extraction network). Then, the speaker prototype memory is updated: If the speaker ID corresponding to the segment already exists in the memory bank, the prototype embedding vector is updated using the exponential moving average (EMA) mechanism: new prototype = β * old prototype + (1 - β) * newly extracted embedding, where β is a smoothing factor (e.g., 0.9).

[0033] If it is a completely new speaker, the newly extracted embedding vector is directly stored in the memory bank as its prototype.

[0034] After the update is complete, the prototype embedding vectors in the memory are updated. For subsequent audio data arriving after step 110, in step 120, the speaker log model will refer to these updated prototype embeddings for identification. Simultaneously, during inference, the system can also directly calculate the cosine similarity between subsequent audio features and all prototypes in the memory for rapid speaker matching, thereby achieving low-latency, real-time identification of new speakers or speakers with changing voices.

[0035] In addition, the method also includes a model iterative optimization step (not shown in the figure): the system collects the ground truth labels obtained in step 140 and the model's high confidence predictions of non-query segments (which are used as pseudo-labels after being filtered by dynamic thresholds), and together with the original training data, it periodically fine-tunes the speaker log model in step 120 and updates its model parameters, thereby achieving continuous improvement in long-term performance and forming a complete "interaction-adaptation-optimization" closed loop.

[0036] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

[0037] As a further optional embodiment, after the step of updating the speaker prototype embedding according to the truth value label, the method further includes: Based on the audio data that was not selected as the query segment and its initial recognition results, its confidence level is compared with a reliability threshold, and predictions with a confidence level higher than the reliability threshold are generated as pseudo-labels. The model parameters of the core speaker log model are updated using the truth labels, the pseudo labels, and the original labeled data.

[0038] In this embodiment, in step 130, the system uses uncertainty analysis to filter out high-value segments for user annotation, while the remaining large number of audio segments (denoted as the unannotated set U) are not selected. For these segments, the model has already output initial recognition results (including speaker labels and confidence scores). To utilize this unannotated data, the system performs a self-learning process.

[0039] First, the system needs to determine which predictions from the model on the unlabeled set U are sufficiently reliable to be used as pseudo-labels for training. To this end, the system sets a reliability threshold τ. For each prediction in set U at a given time frame, its confidence level (i.e., the probability the model assigns to that label) is compared to the threshold τ: if the confidence level is greater than τ, the prediction is accepted as a pseudo-label; otherwise, it is discarded.

[0040] It is worth noting that the reliability threshold τ is dynamically adjustable, and its adjustment is directly based on user feedback. Specifically, the system analyzes the segments corrected by the user in step 140, classifies them according to their acoustic or environmental characteristics (such as background noise level and overlapping speech ratio), and calculates the actual error rate of the model under each type of feature condition. Under feature conditions with a high model error rate (e.g., high-noise environments), the system will correspondingly increase the threshold τ under that condition to make the pseudo-label screening more stringent and avoid introducing low-quality predictions; under conditions where the model performs robustly, the threshold τ can be decreased to utilize more pseudo-labels. This dynamic threshold adjustment mechanism based on error patterns effectively improves the overall quality of pseudo-labels, thereby ensuring the effectiveness of self-learning.

[0041] After obtaining high-quality ground truth labels (from user feedback) and reliable pseudo labels, the system combines them with the original labeled data (such as the large-scale labeled dataset used for initial training) to periodically fine-tune the speaker log model.

[0042] The specific implementation process is as follows: Construct a hybrid training set: combine the original labeled data, the set of ground truth labels from user feedback, and the set of pseudo labels into a single comprehensive training dataset.

[0043] Loss Function Definition: The speaker log model is based on an end-to-end neural network architecture, refining the speaker log task as a multi-label sequence classification problem. The model uses a binary cross-entropy loss combined with permutation invariant training (PIT) as the basic loss function. For each training sample, the model outputs the activity probability of each speaker in each time frame, solving the label permutation problem by traversing all possible speaker label permutations and selecting the permutation with the minimum loss.

[0044] Combined multi-source loss: The total loss function consists of a weighted sum of three parts: First loss: calculated based on the original labeled data, used to preserve the original knowledge of the model; The second loss is calculated based on the truth label from user feedback and used to incorporate high-quality human correction signals. The third loss is calculated based on the filtered, reliable pseudo-labels, and is used to leverage the potential information in a large amount of unlabeled data.

[0045] The weighting coefficients can be set according to the reliability and quantity of the data sources.

[0046] Perform model optimization: Use an optimization algorithm (such as the Adam optimizer) to minimize the total loss function described above, and update all trainable parameters of the speaker log model. The updated model will have stronger generalization ability and more accurate recognition performance.

[0047] Periodic execution strategy: The above-mentioned model iterative optimization steps are usually not performed in real time, but rather in batches at predetermined intervals (e.g., after a certain number of new labels are collected) or during system idle periods (e.g., at night) to achieve efficient offline learning. The optimized model parameters will be deployed to the online system for subsequent real-time recognition tasks.

[0048] By introducing this self-learning and iterative optimization step, the system can not only achieve real-time online adaptation through step 150, but also continuously improve model performance in the long term by utilizing user feedback and a large amount of unlabeled data, thus forming a complete "interaction-adaptation-optimization" closed loop, enabling the system to have continuous evolution capabilities.

[0049] As a further optional embodiment, the step of updating the model parameters of the core speaker log model using the truth labels, the pseudo labels, and the original labeled data specifically includes: Construct a total loss function, which is a weighted sum of a first loss calculated based on the original labeled data, a second loss calculated based on the ground truth labels, and a third loss calculated based on the pseudo labels; The parameters of the speaker log model are updated by minimizing the total loss function using an optimization algorithm.

[0050] This embodiment is the core of the model's long-term, stable performance evolution, achieved through a periodic fine-tuning mechanism based on multi-source data fusion. Its core idea is to organically combine high-quality but limited user feedback, a large amount of filtered unlabeled data predictions, and the original labeled knowledge base to jointly guide the iterative optimization of the model. This allows the model to continuously incorporate new knowledge and correct errors while maintaining its original capabilities.

[0051] The specific implementation includes the following two key steps: 1. Construct the total loss function for multi-source fusion. The total loss function L_total is the optimization objective that guides the update of the model parameters. It is constructed as a weighted sum of three independent loss components, each corresponding to a type of training data source: The first loss (L_original) is calculated based on the original labeled data. This dataset is typically a large-scale, general-purpose labeled dataset (such as publicly available datasets like AMI and DIHARD) used during the initial training phase of the model. Continuing to use this data to calculate the loss during the fine-tuning phase is to prevent the model from experiencing "catastrophic forgetting" as it adapts to new scenarios or users, ensuring that its general foundation and robustness are maintained.

[0052] The second loss (L_feedback) is calculated based on the truth labels of explicit user feedback obtained from step 140. This data comes from segments actively learned and directly corrected by users during actual system deployment. It is characterized by extremely high quality and strong relevance, directly reflecting the most specific error patterns and weaknesses of the model in the current real-world application environment (e.g., difficulty in recognizing specific user accents or conference room echoes). This loss term aims to powerfully and accurately correct these discovered errors.

[0053] The third loss (L_pseudo) is calculated based on the reliable pseudo-labels generated in step 160. This data comes from the system's high-confidence predictions of massive amounts of unlabeled audio, and has undergone rigorous filtering using a dynamic reliability threshold. It is characterized by its large data volume, wide coverage of scenarios, and rich acoustic variations and dialogue patterns not reflected in the original labeled data. This loss term aims to leverage these "self-generated" supervisory signals to significantly broaden and deepen the model's generalization ability through semi-supervised learning.

[0054] The formula for calculating the total loss function is: *L_total = w1 · L_original + w2 · L_feedback + w3 · L_pseudo* Here, w1, w2, and w3 are preset weight coefficients used to balance the influence of different data sources on this parameter update. The weight setting strategy can be flexibly adjusted: for example, when there is little user feedback in the early stages of deployment, the weights of w1 and w3 can be appropriately increased; after accumulating high-quality feedback, the weight of w2 can be increased for more targeted model correction. In a typical configuration, *w1=0.3, w2=0.5, w3=0.2* can be used to highlight the dominant role of high-quality human feedback.

[0055] 2. Execution optimization and parameter updates During each predetermined model update cycle (e.g., after collecting N new user feedback, or executed daily at a set time), the system performs the following operations: Data preparation: Load a batch of the original labeled dataset, the accumulated set of ground truth user labels, and the latest generated set of reliable pseudo-labels from the corresponding storage.

[0056] Forward computation and loss aggregation: The three data sets are input into the current model, and L_original, L_feedback, and L_pseudo are calculated based on their corresponding labels. As mentioned earlier, each loss is calculated using a binary cross-entropy loss combined with permutation invariant training (PIT) to properly handle permutation ambiguity of speaker labels. Then, L_total is obtained by weighted summation according to the above formula.

[0057] Backpropagation and parameter optimization: An optimization algorithm (such as the Adam optimizer, with a learning rate that can be set to 1e-5) is used to minimize L_total. The gradient of the total loss with respect to all trainable parameters of the model is calculated using the backpropagation algorithm, and then the optimizer is used to update the model parameters based on these gradients. This process drives the model to evolve towards three objectives simultaneously: consolidating old knowledge, correcting known errors, and absorbing new experience.

[0058] Model Deployment and Closed Loop: The updated model parameters are saved and synchronized to the online inference system, replacing the old model and completing one performance iteration. The system then enters the next closed loop cycle of "inference → active selection and feedback → online adaptation and pseudo-label generation → periodic fine-tuning".

[0059] Through the aforementioned mechanism, the model is freed from dependence on static datasets, forming a lifecycle capable of continuous learning and self-improvement from real-world use. This is key to achieving the long-term high accuracy and strong adaptability of this invention. This periodic update process is typically performed offline when system resources are idle (e.g., at night) to ensure the stability of online services.

[0060] As a further optional embodiment, the step of performing uncertainty analysis on the audio data based on the initial recognition result, and filtering out the query audio segments with uncertainty scores higher than a preset threshold from the audio data, specifically includes: Calculate the uncertainty score for each audio time frame in the audio data, where the uncertainty score is a weighted sum of classification entropy and minimum boundary distance; The consecutive time frames whose uncertainty scores exceed a preset segment threshold are aggregated into the audio segment to be queried; The classification entropy is used to measure the overall disorder in the model's assignment of speaker labels, and the minimum boundary distance is used to measure the intensity of competition between the two speakers with the highest predicted probabilities.

[0061] The core of this step lies in implementing an efficient, proactive learning sampling strategy tailored to the characteristics of speaker log tasks. The goal is to automatically identify the audio segments where model predictions are least reliable and most likely to err, and prioritize submitting these segments to human annotation and correction, thereby obtaining the most informative training data for model optimization with minimal human effort.

[0062] The specific implementation process includes the following two key sub-steps: Step 1: Calculate the uncertainty score for each audio time frame. For the input audio stream, the model has already output a probability distribution vector P_t for each time frame *t* in step 120. Assuming the system can recognize at most N speakers, P_t is an N-dimensional vector, where each element p_t(i) represents the probability that the model predicts the *i*th speaker is speaking at time *t*.

[0063] Calculate the classification entropy: The classification entropy H_t is calculated based on the entire probability distribution P_t and is used to measure the overall uncertainty or disorder in the model's label assignment for that frame. Its calculation formula is as follows: *H_t = - Σ_{i=1}^{N} [p_t(i) * log(p_t(i))]* The higher the entropy value H_t, the more uncertain the model is about "who is speaking," and the probability quality is spread across multiple speakers.

[0064] Calculate the minimum margin distance (MMD): The minimum boundary distance M_t is specifically designed to address the ambiguity of speaker turns, a common challenge in speaker logging tasks. It focuses only on the competition between the two speakers with the highest probabilities. Its calculation formula is as follows: M_t = p_t(second) / p_t(first) Here, p_t(first) and p_t(second) are the probabilities of the maximum and second largest values ​​in the probability distribution P_t, respectively. The closer M_t is to 1, the more likely the model considers the two most probable speakers to be evenly matched, and the decision boundary is very vague. This position is likely to be the region where speakers switch or speech overlaps.

[0065] Calculate the composite uncertainty score: Finally, for each time frame *t*, a composite uncertainty score U_t is calculated, which is a weighted sum of the two metrics mentioned above: *U_t = α * H_t + (1 - α) * M_t* Here, α is a preset weighting parameter, ranging from 0 to 1, used to balance the impact of overall uncertainty and boundary ambiguity on the selection strategy. For example, if more attention is paid to the regions of overall model confusion, α can be set to 0.7; if more attention is paid to speaker switching points, α can be set to 0.3. H_t and M_t can be normalized before calculation to eliminate the influence of dimensions.

[0066] Step 2: Aggregate high-resolution time frames and filter the segments to be queried. Set a segment threshold: Set a preset segment threshold T_seg. This threshold can be determined by analyzing the distribution of uncertainty scores on a representative validation set, for example, by taking the top X percentile (e.g., the top 10%) of all frame scores.

[0067] Preliminary screening and aggregation: Traverse the entire audio stream and mark all time frames with a composite uncertainty score U_t higher than the threshold T_seg as "high uncertainty frames". Then, aggregate these consecutive time frames marked as high uncertainty to form a candidate "high uncertainty audio segment". Each segment is defined by its start and end times.

[0068] Segment sorting and final selection: Calculate the overall uncertainty score for each candidate segment (e.g., take the average or maximum score of all frames within the segment). Then, sort all candidate segments in descending order of their overall scores. Finally, based on the acceptable frequency of human intervention or budget in the actual application scenario, select the top K segments (K is a preset value, such as 5 or 10) as the final audio segments to be queried, and output them to the human-computer interaction interface in step 140.

[0069] Through the above mechanism, the system can accurately locate and prioritize the submission of audio segments that are most difficult for the model to judge and have the highest risk of error (especially those with ambiguous speaker identities or frequent speaker switching), thereby ensuring that each manual annotation can correct the weak links in the model's cognition to the greatest extent and maximize the efficiency of active learning.

[0070] As a further optional embodiment, the step of collecting explicit feedback on the audio segment to be queried specifically includes: Collect users' tacit confirmation behavior for high-confidence audio segments that were not selected as query segments, as an implicit positive feedback signal; The implicit positive feedback signal is used to assist in the generation of the pseudo-tag or the adjustment of the reliability threshold.

[0071] The core of this step is to obtain users' direct or indirect evaluations of the system's predictions. Besides explicit corrections, the behavioral data generated by users during their interactions with the system also contains valuable feedback information. By monitoring user interaction logs, the system can capture and utilize these implicit feedback signals as an important supplement to explicit feedback.

[0072] The specific implementation method is as follows: Explicit Feedback Acquisition: The system presents the audio segments selected in step 130 to the user through a human-computer interaction interface (such as a web application or desktop tool). The interface also displays the initial recognition results generated by the model (e.g., different speakers identified by timeline bars of different colors). Users can provide explicit feedback by performing the following actions: Confirmation: I agree to the system's annotation; no modifications are needed.

[0073] Correction: Change the speaker labels for a specific time period.

[0074] Refinement: Adjust the start and end times of the speaker's activities.

[0075] Add / Delete: Add an unrecognized speaker or delete an incorrect speaker.

[0076] After the user completes the editing and submits, the system records the modified result as the truth label for that fragment.

[0077] Acquisition and identification of implicit positive feedback signals: For all audio clips presented on the interface, the system synchronously records the user's interaction behavior. In particular, for high-confidence audio clips that were not selected as query clips by the active learning module (i.e., clips with high model prediction confidence and not marked as "requires review"), the system will focus on analyzing the following user behavior patterns: Quick skip or autoplay: The user does not pause, rewind, or perform any editing on the segment during browsing or listening, but allows it to play naturally.

[0078] Extremely short dwell time without modification: Although the user briefly lingers at the fragment, the dwell time is far shorter than the average time required to perform an edit operation, and no edit event is generated.

[0079] When the system detects the above behavioral pattern, it can infer that the user tacitly approves or agrees with the prediction result of the segment, and record this as an implicit positive feedback signal. This signal is associated with the specific audio segment and its corresponding model prediction.

[0080] Applications of implicit feedback signals: The acquired implicit positive feedback signals mainly play a supporting role in two subsequent stages: Assisted pseudo-label generation (for step 160): Implicit confirmation enhances the reliability prior of the prediction results for the corresponding segments. Therefore, in step 160, when the system generates pseudo-labels for these segments, a more aggressive strategy can be adopted. For example, the reliability threshold for these segments can be temporarily lowered to make their predictions more likely to be adopted as pseudo-labels; or, at the same confidence level, pseudo-labels from implicitly confirmed segments can be given higher weight.

[0081] Assisted dynamic threshold adjustment (used in step 160): When dynamically adjusting the reliability threshold, implicit positive feedback and explicit error correction feedback together constitute the evaluation of model performance. The system can analyze in which acoustic environments or speaker characteristics the model not only has a low explicit error rate but also obtains more implicit confirmations. For these "robust and accepted" scenarios, the system can appropriately lower the reliability threshold to encourage the model to utilize more unlabeled data in that scenario (generating more pseudo-labels), thereby further consolidating its performance in that scenario.

[0082] By introducing the collection and application of implicit feedback, the system can more subtly understand and utilize all signals generated in human-machine collaboration, forming a complementary learning paradigm of "explicitly correcting errors and implicitly confirming correctness." This further optimizes the efficiency of active learning and, without increasing the user's explicit annotation burden, provides a richer flow of supervisory information for the model's continuous self-optimization, enhancing the data utilization efficiency and intelligence level of the entire closed-loop learning system.

[0083] As a further optional embodiment, the speaker log model is used to formulate the speaker log task as a multi-label sequence classification problem based on an end-to-end neural network architecture; the first loss, the second loss, and the third loss are all calculated using a binary cross-entropy loss trained with permutation invariance.

[0084] In this embodiment, the speaker log model is based on an end-to-end neural network (EEND) architecture. Unlike traditional pipeline systems that require step-by-step speech activity detection, feature extraction, and clustering, this model directly describes the entire speaker log task as a multi-label sequence classification problem. Specifically, for an audio segment of length T, the model aims to simultaneously output the binary activity state (0 indicates no speech, 1 indicates speech) of all possible speakers at each time frame. The model is typically composed of multiple Transformer encoder layers or recurrent neural network layers stacked together. It directly receives acoustic feature sequences (such as the MFCC extracted in step 110) and outputs a matrix of shape T × S, where S is the preset maximum number of possible speakers, and each element in the matrix represents the activity probability of the corresponding speaker at the corresponding time frame.

[0085] The design and calculation of the loss function are the core of this model's training. Because the speaker log task involves a "label permutation" problem—that is, the speaker order predicted by the model may not match the order of the actual labels, even though the content may be identical—directly using the standard binary cross-entropy (BCE) loss would result in an ineffective penalty. Therefore, this embodiment uses a binary cross-entropy loss combined with permutation invariant training (PIT) as the basic loss function.

[0086] The specific calculation process is as follows: Basic loss (for a single sample): For an input sequence X, the model outputs a predicted sequence Y_pred (a T × S probability matrix). The corresponding true label sequence is Y_true (a T × S 0 / 1 matrix).

[0087] First, define the set Π (of size S!) of all possible speaker label permutations.

[0088] Then, for each permutation π ∈ Π, the true label Y_true is rearranged according to that permutation π to obtain Y_true^π.

[0089] Next, calculate the binary cross-entropy loss L_BCE(Y_pred, Y_true^π) between the predicted Y_pred and the true value Y_true^π after permutation.

[0090] Finally, the PIT loss is defined as the minimum loss among all possible permutations: L_PIT = min_{π ∈ Π} [ L_BCE(Y_pred, Y_true^π) ] Applied to multi-source data (specification of step 170): In periodic fine-tuning, the total loss function L_total is composed of a weighted sum of three PIT losses: *L_total = λ1 * L_original + λ2 * L_feedback + λ3 * L_pseudo* in: L_original: The first loss calculated based on the original labeled data. This data forms the basis for model pre-training, ensuring that existing knowledge is not forgotten during fine-tuning.

[0091] L_feedback: The second loss is calculated based on the truth labels of the explicit user feedback obtained from step 140. This part of the data is small in quantity but extremely high in quality, directly reflecting the latest error patterns of the model in the actual deployment environment, and is a key signal driving targeted improvements to the model.

[0092] L_pseudo: A third loss calculated based on reliable pseudo-labels. This part has a large amount of data, and its overall reliability is ensured through a dynamic threshold mechanism. It is used to mine information from massive amounts of unlabeled data and enhance the model's generalization ability.

[0093] λ1, λ2, and λ3 are preset weighting coefficients used to balance the contributions of different data sources to model updates. In an exemplary setting, λ2 > λ1 ≈ λ3 can be set to emphasize the guiding role of high-quality human feedback.

[0094] Model optimization: In each fine-tuning batch or cycle, the optimizer (such as Adam) aims to minimize the total loss L_total. The gradient of the loss with respect to all model parameters is calculated using backpropagation, and the parameters are updated accordingly. This training method allows the model to simultaneously absorb information from a stable knowledge base (raw data), precise calibration signals (user feedback), and extensive empirical data (pseudo-labels), thus achieving continuous and robust performance evolution.

[0095] By adopting this PIT-based end-to-end architecture and multi-source fusion loss calculation method, this embodiment not only fundamentally solves the problem of label replacement, but also provides a solid and flexible mathematical framework for achieving efficient closed-loop learning—that is, continuous self-optimization using mixed-quality data streams.

[0096] As a further optional embodiment, the step of updating the speaker prototype embedding based on the truth value label specifically includes: Extract the speaker embedding vector from the audio segment corresponding to the truth value label; For speakers already existing in the speaker prototype memory, an exponential moving average mechanism is used to iteratively update their corresponding prototype embedding vectors using the newly extracted speaker embedding vectors. For a newly emerging speaker, its speaker embedding vector is stored as a new prototype embedding vector in the speaker prototype memory.

[0097] This step is central to achieving the system's low-latency online adaptability. Its purpose is to leverage user feedback to instantly optimize the system's understanding of specific speakers, thereby improving the accuracy of subsequent recognition. The specific implementation includes the following three key steps: 1. Speaker embedding vector extraction Once a user provides feedback (confirmation or correction) on a query audio segment, that segment and its corresponding final speaker label (truth label) are determined to be reliable supervisory data. The system first extracts the speaker's acoustic feature representation from the audio segment corresponding to the truth label, which is called the speaker embedding vector.

[0098] Specifically, the system uses a pre-trained speaker encoder (e.g., a TDNN-based x-vector model or an LSTM-based d-vector model) independent of the core log model to process the audio segment. This encoder converts the speech signal in the segment into a fixed-length, high-dimensional vector representation. This vector aims to capture the speaker's unique, text-independent acoustic identity features (such as timbre and pitch) and eliminate the influence of channel and background noise as much as possible. The extracted embedding vector is denoted as e_new.

[0099] 2. Updating existing speaker prototypes The system maintains a speaker prototype memory, in which a prototype embedding vector, denoted as p_old, is stored for each known speaker (identified by a unique ID). This vector is the system's best current estimate of the speaker's vocal features.

[0100] When the speaker ID corresponding to the newly extracted embedding vector e_new already exists in the memory, the system uses an Exponential Moving Average (EMA) mechanism to update the speaker's prototype embedding. This is a smooth iterative update strategy that allows the prototype to both keep up with the latest changes in the speaker's voice and maintain a certain degree of historical stability, avoiding drastic fluctuations caused by the uncertainty of a single input. The update formula is as follows: p_new = β * p_old + (1 - β) * e_new in: p_new is the updated prototype embedding vector.

[0101] p_old is the original embedding vector before the update.

[0102] e_new is the new embedding vector extracted from the current feedback fragment.

[0103] β is the smoothing factor (or forgetting factor), a preset hyperparameter that typically ranges from 0.5 to 0.95. The closer β is to 1, the more conservative the prototype update, with historical information dominating; the closer it is to 0, the more sensitive it is to new information. For example, a value of 0.85 means that new information contributes approximately 15% to each update.

[0104] 3. Registration of new speaker prototypes If the speaker ID corresponding to the newly extracted embedding vector e_new does not exist in the speaker memory (i.e., this is a speaker that the system is recognizing for the first time), the system performs an online registration process: directly using e_new as the initial prototype embedding vector for the new speaker, and storing it along with its ID in the speaker prototype memory. This gives the system the ability to dynamically expand the set of known speakers.

[0105] Technical Effects: Through the above mechanism, the system achieves "learn-as-you-go" functionality. Once user feedback is provided, the prototype of the relevant speaker can be updated or created within milliseconds. When subsequent audio streams arrive, the core speaker log model or real-time matching module uses the updated memory for inference, immediately demonstrating improved recognition of corrected errors or new speakers. This significantly reduces speaker confusion errors and fundamentally solves the bottleneck of traditional offline systems' inability to adapt to new speakers or voice drift, making it particularly suitable for dynamic scenarios such as continuous meetings and dialogues. This update process is real-time and lightweight, causing no significant delay to the system's main audio processing pipeline.

[0106] The speaker log device for closed-loop learning provided by this invention is described below, such as... Figure 2 As shown, the speaker log device for closed-loop learning described below and the speaker log method for closed-loop learning described above can be referred to in correspondence.

[0107] A speaker log device for closed-loop learning, comprising: The audio acquisition module 210 is used to acquire audio data to be processed, and to perform speech activity detection and acoustic feature extraction on the audio data to obtain front-end features; The initial recognition module 220 is used to receive the front-end features through the speaker log model and generate at least one initial recognition result by referring to the speaker prototype embedding obtained from the speaker prototype memory bank. The initial recognition result includes speaker label and confidence score. Uncertainty analysis module 230 is used to perform uncertainty analysis on the audio data based on the initial identification result, and to filter out the audio segments to be queried from the audio data whose uncertainty scores are higher than a preset threshold. The explicit feedback module 240 is used to collect explicit feedback on the audio segment to be queried and obtain the corrected truth label. The parameter update module 250 is used to update the speaker prototype embedding according to the truth label, and use the updated speaker prototype embedding to perform real-time speaker matching and recognition on subsequent audio data.

[0108] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a closed-loop learning speaker logging method, which includes: Acquire the audio data to be processed, and perform speech activity detection and acoustic feature extraction on the audio data to obtain front-end features; The front-end features are received through the speaker log model, and at least one initial recognition result is generated by referring to the speaker prototype embedding obtained from the speaker prototype memory. The initial recognition result includes speaker label and confidence score. Based on the initial identification results, uncertainty analysis is performed on the audio data, and audio segments with uncertainty scores higher than a preset threshold are selected from the audio data. Explicit feedback on the audio segment to be queried is collected to obtain the corrected truth label; Based on the truth labels, the speaker prototype embedding is updated, and the updated speaker prototype embedding is used to perform real-time speaker matching and recognition on subsequent audio data.

[0109] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0110] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is capable of executing the closed-loop learning speaker log method provided by the above methods, the method comprising: Acquire the audio data to be processed, and perform speech activity detection and acoustic feature extraction on the audio data to obtain front-end features; The front-end features are received through the speaker log model, and at least one initial recognition result is generated by referring to the speaker prototype embedding obtained from the speaker prototype memory. The initial recognition result includes speaker label and confidence score. Based on the initial identification results, uncertainty analysis is performed on the audio data, and audio segments with uncertainty scores higher than a preset threshold are selected from the audio data. Explicit feedback on the audio segment to be queried is collected to obtain the corrected truth label; Based on the truth labels, the speaker prototype embedding is updated, and the updated speaker prototype embedding is used to perform real-time speaker matching and recognition on subsequent audio data.

[0111] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a speaker log method for closed-loop learning provided by the methods described above, the method comprising: Acquire the audio data to be processed, and perform speech activity detection and acoustic feature extraction on the audio data to obtain front-end features; The front-end features are received through the speaker log model, and at least one initial recognition result is generated by referring to the speaker prototype embedding obtained from the speaker prototype memory. The initial recognition result includes speaker label and confidence score. Based on the initial identification results, uncertainty analysis is performed on the audio data, and audio segments with uncertainty scores higher than a preset threshold are selected from the audio data. Explicit feedback on the audio segment to be queried is collected to obtain the corrected truth label; Based on the truth labels, the speaker prototype embedding is updated, and the updated speaker prototype embedding is used to perform real-time speaker matching and recognition on subsequent audio data.

[0112] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0113] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0114] Finally, it should be noted that the above embodiments 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.

Claims

1. A speaker log method for closed-loop learning, characterized in that, include: Acquire the audio data to be processed, and perform speech activity detection and acoustic feature extraction on the audio data to obtain front-end features; The front-end features are received through the speaker log model, and at least one initial recognition result is generated by referring to the speaker prototype embedding obtained from the speaker prototype memory. The initial recognition result includes speaker label and confidence score. Based on the initial identification results, uncertainty analysis is performed on the audio data, and audio segments with uncertainty scores higher than a preset threshold are selected from the audio data. Explicit feedback on the audio segment to be queried is collected to obtain the corrected truth label; Based on the truth labels, the speaker prototype embedding is updated, and the updated speaker prototype embedding is used to perform real-time speaker matching and recognition on subsequent audio data.

2. The speaker log method for closed-loop learning according to claim 1, characterized in that, Following the step of updating the speaker prototype embedding based on the truth value label, the method further includes: Based on the audio data that was not selected as the query segment and its initial recognition results, its confidence level is compared with a reliability threshold, and predictions with a confidence level higher than the reliability threshold are generated as pseudo-labels. The model parameters of the core speaker log model are updated using the truth labels, the pseudo labels, and the original labeled data.

3. The speaker log method for closed-loop learning according to claim 2, characterized in that, The step of updating the model parameters of the core speaker log model using the truth labels, the pseudo labels, and the original labeled data specifically includes: Construct a total loss function, which is a weighted sum of a first loss calculated based on the original labeled data, a second loss calculated based on the ground truth labels, and a third loss calculated based on the pseudo labels; The parameters of the speaker log model are updated by minimizing the total loss function using an optimization algorithm.

4. The speaker log method for closed-loop learning according to claim 3, characterized in that, The step of performing uncertainty analysis on the audio data based on the initial recognition result, and filtering out the query audio segments with uncertainty scores higher than a preset threshold from the audio data, specifically includes: Calculate the uncertainty score for each audio time frame in the audio data, where the uncertainty score is a weighted sum of classification entropy and minimum boundary distance; The consecutive time frames whose uncertainty scores exceed a preset segment threshold are aggregated into the audio segment to be queried; The classification entropy is used to measure the overall disorder in the model's assignment of speaker labels, and the minimum boundary distance is used to measure the intensity of competition between the two speakers with the highest predicted probabilities.

5. The speaker log method for closed-loop learning according to claim 4, characterized in that, The step of collecting explicit feedback on the audio segment to be queried specifically includes: Collect users' tacit confirmation behavior for high-confidence audio segments that were not selected as query segments, as an implicit positive feedback signal; The implicit positive feedback signal is used to assist in the generation of the pseudo-tag or the adjustment of the reliability threshold.

6. The speaker log method for closed-loop learning according to claim 5, characterized in that, The speaker log model is used to describe the speaker log task as a multi-label sequence classification problem based on an end-to-end neural network architecture; the first loss, the second loss, and the third loss are all calculated using a binary cross-entropy loss trained with permutation invariance.

7. The speaker log method for closed-loop learning according to claim 1, characterized in that, The step of updating the speaker prototype embedding based on the truth value label specifically includes: Extract the speaker embedding vector from the audio segment corresponding to the truth value label; For speakers already existing in the speaker prototype memory, an exponential moving average mechanism is used to iteratively update their corresponding prototype embedding vectors using the newly extracted speaker embedding vectors. For a newly emerging speaker, its speaker embedding vector is stored as a new prototype embedding vector in the speaker prototype memory.

8. A speaker log device for closed-loop learning, characterized in that, include: The audio acquisition module is used to acquire audio data to be processed, and to perform speech activity detection and acoustic feature extraction on the audio data to obtain front-end features; An initial recognition module is used to receive the front-end features through a speaker log model and generate at least one initial recognition result by referring to the speaker prototype embedding obtained from the speaker prototype memory bank. The initial recognition result includes speaker labels and confidence scores. An uncertainty analysis module is used to perform uncertainty analysis on the audio data based on the initial identification results, and to filter out the audio segments to be queried from the audio data whose uncertainty scores are higher than a preset threshold. The explicit feedback module is used to collect explicit feedback on the audio segment to be queried and obtain the corrected truth label. The parameter update module is used to update the speaker prototype embedding based on the truth label, and to use the updated speaker prototype embedding to perform real-time speaker matching and recognition on subsequent audio data.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the speaker logging method for closed-loop learning as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the speaker log method for closed-loop learning as described in any one of claims 1 to 7.

Images