Self-evolution method and system of speech recognition model

Abstract

A self-evolving method and system for a speech recognition model is disclosed. This method extracts high-value text fragments and constructs a training set by combining them with multi-style speech synthesis data. After incorporating general data, parameter fine-tuning techniques are used to optimize the parameters of the temporary model. Simultaneously, the optimized parameters are stored in an experience replay buffer, and differential encoding of these parameters within the buffer generates a lightweight update package. This application achieves continuous model evolution through a closed-loop optimization process, avoids model forgetting by incorporating general data, significantly reduces the number of parameters during evolution through parameter fine-tuning, and significantly reduces training costs during evolution through the experience replay buffer, thus ensuring the stability of model evolution.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and machine learning technology, and in particular relates to a self-evolution method and system for a speech recognition model. Background Technology

[0002] In the field of speech recognition technology, the mainstream approach to improving existing model performance is to periodically collect large amounts of user speech data, manually annotate it, and then train the model offline. This method is time-consuming and costly. Some systems optimize the model by recording user corrections to the recognition results (such as manually correcting transcribed text). However, this method is highly dependent on the feedback of the original speech data, posing privacy and compliance risks, and consuming significant computational resources. Other methods utilize unlabeled data (such as untranscribed speech streams) for self-supervised model learning, or combine a small amount of labeled data to achieve semi-supervised training. However, these methods can only improve the model's robustness to existing knowledge and cannot effectively learn new vocabulary or domain knowledge. Still other methods generate additional training data through text-to-speech synthesis models, such as using publicly available text libraries to synthesize speech to expand the training set. However, existing methods often lack simulation of real-world scenarios, resulting in a lack of diversity in the generated speech and insufficient model generalization ability.

[0003] Therefore, existing technologies for improving speech recognition models cannot quickly respond to new words and scenarios, resulting in a poor user experience. To meet data compliance requirements, enterprises need to invest additional resources in data anonymization or desensitization, increasing the technical burden. Edge devices need to frequently upload data or perform complex calculations, resulting in high energy consumption and response delays, making it difficult to balance performance and cost. In noisy environments or specific dialect scenarios, the recognition accuracy drops significantly, affecting the expansion of application scenarios.

[0004] In summary, existing technologies struggle to achieve efficient, low-cost, and privacy-preserving self-evolving methods for speech recognition models. Summary of the Invention

[0005] The purpose of this application is to provide a self-evolving method and system for speech recognition models that can break free from dependence on raw speech data, quickly respond to new knowledge, ensure privacy and security, and be highly efficient and low-cost, thereby solving the bottlenecks that are difficult to overcome in existing technologies.

[0006] A first aspect of this application provides a self-evolution method for a speech recognition model, the method comprising the following steps:

[0007] The background knowledge document and the user's correction text of the speech recognition model's output are stored in the corpus;

[0008] Extracting high-value text fragments from a corpus;

[0009] Convert high-value text fragments into speech data;

[0010] The parameters of the temporary model of the speech recognition model are fine-tuned by using the training sample dataset to obtain the evolution parameters of the speech recognition model. The training sample dataset includes high-value text segments and corresponding speech data.

[0011] When the key metrics of the temporary model reach a threshold, the user actively triggers a deployment command, or a decline in the performance of the speech recognition model is detected, the evolution parameters will be deployed to the speech recognition model.

[0012] In other embodiments of this application, extracting high-value text fragments from a corpus includes:

[0013] Obtain the number of corrections and the error type of the correction text, and label the correction text with the scene tag;

[0014] Based on the background knowledge document corresponding to the scene tags of the corrected text, obtain the keywords and text topics in the corrected text;

[0015] The value score of the corrected text is determined based on keywords, text topic, number of corrections, error type, and scenario tags.

[0016] Corrected texts in the corpus with value scores higher than the preset value are designated as high-value text fragments.

[0017] Furthermore, high-value text fragments are converted into speech data, including:

[0018] High-value text fragments are input into the text-to-speech engine to generate simulated speech data with different timbres, speech rates, and intonations. The text-to-speech engine supports switching between multiple speaker styles or timbre cloning based on user-provided speech samples.

[0019] Select at least one scene noise from a preset noise database and add it to the simulated speech to obtain noisy simulated speech data;

[0020] Adjust the signal-to-noise ratio of the noisy analog speech to a preset range to obtain noisy synthesized speech data.

[0021] Furthermore, after converting high-value text fragments into speech data, the method also includes constructing a training sample dataset:

[0022] The noisy synthesized speech was paired with the corresponding high-value text fragments as training sample data, and the training sample data was labeled with scene type, speaker style, noise parameters and synthesis timestamp.

[0023] Furthermore, the training sample dataset also includes: general text fragments and corresponding speech data, where the general text fragments are selected from a large-scale general text sample training set.

[0024] Furthermore, after fine-tuning the parameters of the temporary model of the speech recognition model using the training sample dataset to obtain the evolutionary parameters of the speech recognition model, the process further includes:

[0025] Store the evolutionary parameters in the experience replay buffer;

[0026] Evolutionary parameters are randomly sampled from the experience replay buffer to obtain experience tuples;

[0027] The speech recognition model is adjusted based on the empirical tuple.

[0028] Furthermore, the parameters of the temporary speech recognition model are fine-tuned using the training sample dataset to obtain the evolutionary parameters of the speech recognition model, including:

[0029] Based on the parameters of the speech recognition model, a temporary storage model for the speech recognition model is constructed.

[0030] The parameters of the temporary model are decomposed into fixed basis parameters and trainable low-rank matrices;

[0031] Iterative training of the trainable low-rank matrix is ​​performed using the training sample dataset;

[0032] Extract the difference update between the trained trainable low-rank matrix and the fixed basis parameters to form the evolution parameter set;

[0033] Differential encoding or incremental compression of the evolutionary parameter set can generate evolutionary parameters that can be deployed efficiently.

[0034] In other embodiments of this application, the key metrics of the temporary storage model reach a threshold, including:

[0035] Deployment is triggered when at least one of the following conditions is met:

[0036] The performance metrics of the temporary model on the validation set have reached the preset threshold.

[0037] The number of training rounds reaches the preset limit or resource consumption meets the budget constraint;

[0038] The user actively triggers the deployment command or detects a significant drop in model performance.

[0039] Furthermore, based on the empirical tuples, the speech recognition model is adjusted, including:

[0040] The evolutionary parameters are deployed to the speech recognition model, the trainable low-rank matrix replaces the original parameters of the corresponding layer of the speech recognition model, and the fixed basis parameters updated by difference are weighted and merged with the original parameters of the speech recognition model.

[0041] A second aspect of this application provides a self-evolving system for a speech recognition model, comprising:

[0042] The knowledge acquisition module is used to store background knowledge documents and user correction texts of the speech recognition model's output into a corpus.

[0043] The text extraction module is used to extract high-value text fragments from the corpus;

[0044] The data synthesis module is used to convert high-value text fragments into speech data;

[0045] The model fine-tuning module is used to fine-tune the parameters of the temporary model of the speech recognition model using the training sample dataset to obtain the evolution parameters of the speech recognition model. The training sample dataset includes high-value text segments and corresponding speech data.

[0046] An automated deployment module is used to deploy evolution parameters to the speech recognition model when the key indicators of the temporary model reach a threshold, the user actively triggers a deployment command, or a decline in the performance of the speech recognition model is detected.

[0047] The beneficial effects of this application's embodiments compared to existing technologies are as follows: This application extracts key knowledge such as new vocabulary and entities from user correction feedback and scene-related documents, and fine-tunes the model by generating diverse training data through text-to-speech synthesis technology. While ensuring user privacy and security, this application can capture and transform new knowledge from user feedback and documents in real time. The model can self-evolve without manual intervention, adapting to constantly changing language and scene requirements. Simultaneously, the recognition error rate in professional scenarios is significantly reduced, enabling accurate recognition of terminology in professional scenarios. Attached Figure Description

[0048] Figure 1 A flowchart illustrating a self-evolutionary method for a speech recognition model provided in this application;

[0049] Figure 2 The embodiment provided in this application is a schematic diagram of the process by which a corpus scores stored text fragments;

[0050] Figure 3 A schematic diagram of the speech data synthesis process provided in this application for an embodiment;

[0051] Figure 4 This is a schematic diagram of the structure of a self-evolving speech recognition model provided in an embodiment of this application. Detailed Implementation

[0052] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0053] Figure 1 A flowchart illustrating a self-evolution method for a speech recognition model according to the first embodiment of this application is shown below, in detail:

[0054] S1. Store the background knowledge document and the user's correction text of the speech recognition model's output into the corpus.

[0055] Specifically, the system monitors and accesses background knowledge documents and correction texts in real time. The background knowledge documents, which are textual materials related to the speech recognition scenario, include:

[0056] Structured documents, such as meeting agendas, speeches, product manuals, technical documents, etc.;

[0057] Unstructured text, such as web page excerpts, industry white papers, terminology dictionaries, etc.;

[0058] Scenario-specific data, such as electronic medical record templates and drug instructions in a medical scenario, or product documents in a customer service scenario.

[0059] User-corrected text refers to the text generated by users' corrections to the model's output, such as:

[0060] In a meeting setting, users manually correct the text transcribed by the speech recognition model.

[0061] After processing the aforementioned background knowledge documents and correction texts, they are stored in a dedicated corpus, which can be divided into a "background knowledge area" (classified by document type) and a "user correction area" (classified by correction type and scenario). New data is added in real time, and low-value or outdated information is cleaned up periodically.

[0062] S2. Extract high-value text fragments from the corpus.

[0063] It should be noted that before extracting high-value text fragments from the corpus, the corpus scores the stored text fragments. For example, this application uses a multi-dimensional scoring model, and the scoring indicators include, but are not limited to:

[0064] Novelty score is calculated by detecting whether the text contains new words or entities outside the vocabulary. The score is calculated using the probability distribution of the language model or character-level n-gram statistics. The higher the novelty score, the higher the value score of the text segment.

[0065] Revise the frequency score. By counting the number of corrections made by users to the corresponding entities in the text, the more corrections are made, the higher the value score will be;

[0066] Scene relevance score. According to the scene where the current model is deployed (such as meeting, customer service, medical), higher value scores are assigned to terms that are strongly related to the scene (such as product names in meetings, drug names in medical).

[0067] Timeliness score. Higher timeliness scores are assigned to newly added texts recently to ensure that the latest knowledge has a higher value score.

[0068] Specifically, see Figure 2 and conduct a value score on the corrected text, including the following steps:

[0069] S201. Obtain the number of corrections and the error types of the corrected text, and label the scene tags of the corrected text.

[0070] Maintain a dynamic database to record the number of corrections for each corrected text. For example, whenever a user corrects "XX product" output by the model to "YY brand", the number of times of the corresponding correction pair is incremented by 1.

[0071] Automatically identify the error types of the corrected text through natural language processing technology. Exemplarily, common types include:

[0072] Misspelled words (such as "sapphire" → "sapphire time");

[0073] Homophone confusion (such as "ten o'clock" → "pilot project");

[0074] Term missing (such as the professional term "photon quantum chip" not recognized); [[ID=3)]]

[0075] Entity boundary error (such as "Huawei Mate40" → "Huawei Mate 40").

[0076] The error type annotation provides a targeted optimization direction for subsequent fine-tuning. [[ID= )]]

[0077] Use context information and semantic analysis to automatically label the scene tags of the text. It should be noted that if the scene is clearly associated with a background knowledge document (such as the corresponding meeting PPT has been uploaded), the scene tag corresponding to this document will be preferentially associated.

[0078] S202. According to the background knowledge document corresponding to the scene tag of the corrected text, obtain the keywords and text themes in the corrected text.

[0079] The system retrieves a collection of background knowledge documents corresponding to scene tags, performs multi-document fusion on the collection, and extracts keywords from the documents. For example, it can simultaneously extract product names from meeting PPTs and technical terms from industry white papers.

[0080] S203. Determine the value score of the corrected text based on keywords, text topic, number of corrections, error type, and scene label.

[0081] The value score (V) is calculated by weighted summation, for example:

[0082] V = w1 × number of corrections + w2 × error type weight + w3 × scene relevance + w4 × topic popularity corrections (weight w1): directly use the statistical number of corrections, and high-frequency corrections have higher priority;

[0083] Error type (weight w2): Different error types are assigned different weights (e.g., missing term weight > homonym confusion);

[0084] Scene relevance (weight w3): If the entity in the corrected text appears in the background document of the current scene, the relevance score will improve;

[0085] Topic popularity (weight w4): If an entity belongs to a high-frequency topic in the current scenario (such as "financial data" in a meeting) or new knowledge, the score will increase.

[0086] S204. Set the corrected texts in the corpus with a value score higher than the preset value as high-value text fragments.

[0087] Sort all corrected texts by score and extract the top N percent of texts.

[0088] S3. Convert high-value text fragments into speech data.

[0089] It should be noted that in this application, the text can be segmented into words. Specifically, natural language tools such as jieba and spaCy can be used to identify and label special entities. For words with multiple pronunciations, the correct pronunciation can be selected through contextual semantics or scene knowledge base.

[0090] Next, prosodic symbols (such as pauses and accent marks) are inserted into the text to optimize the naturalness of the synthesized speech. Users can customize prosodic templates, for example, using a more formal tone in meeting scenarios and a friendly speech rhythm in customer service scenarios.

[0091] The text-to-speech engine parameters can be adjusted according to the speech characteristics of different scenarios, for example:

[0092] In meeting settings, a slower speaking speed, clear pronunciation, and highlighting of key terms are appropriate.

[0093] In customer service scenarios, a friendly tone and natural intonation can be used to avoid a mechanical feel;

[0094] In medical settings, it's important to emphasize the accurate pronunciation of medical terminology to avoid ambiguity.

[0095] For details, see Figure 3 Converting high-value text fragments into speech data involves the following steps:

[0096] S301. Input high-value text fragments into the text-to-speech engine to generate simulated speech data with different timbres, speech rates, and intonations. The text-to-speech engine supports multiple speaker style switching or timbre cloning based on user-provided speech samples.

[0097] Input high-value text fragments into the text-to-speech engine and dynamically configure the following parameters:

[0098] Voice parameter configuration offers multiple speaker style options (such as male / female, different ages, different regional accents), or voice cloning based on user-provided voice samples;

[0099] Speech rate adjustment configuration: Customize the speech rate coefficient (e.g., 0.8x-1.2x normal speech rate) to generate speech with different rhythms;

[0100] Intonation control configuration: Adjust emotional expression through prosodic markers (such as emphasis, rising and falling intonation), for example, raising the tone in important announcements to enhance the message.

[0101] Multiple variants can be generated for the same text segment: different timbre × different speech rate × different intonation combinations, such as generating versions like "formal male voice - normal speech rate" and "friendly female voice - slow speed"; random parameters can be perturbed, such as fine-tuning the intonation curve or inserting random pauses to increase the diversity of the data.

[0102] Optionally, with the user's permission, the system can collect their voice samples (such as recordings longer than 10 minutes) and perform voice cloning through the following steps:

[0103] Extract acoustic features of user speech (such as spectrum, fundamental frequency, formants);

[0104] A text-to-speech model that matches the user's voice is generated using a vocoder combined with a pre-trained model.

[0105] S302. Select at least one scene noise from the preset noise database and add it to the simulated speech to obtain noisy simulated speech data.

[0106] A pre-built scenario-based noise database covers common environmental sounds (such as conference rooms, streets, cafes, hospitals, etc.). Users can also upload noise samples, such as uploading a recording of a specific conference room as a dedicated noise source.

[0107] Based on the scene tags of the text, select at least one noise type from the noise library corresponding to the scene;

[0108] Noise can be mixed, for example, by adding air conditioning noise and keyboard typing sounds at the same time, to better simulate a real conference room scene.

[0109] The generated simulated speech is superimposed with selected noise samples to generate noisy simulated speech data. Specifically, noise of different intensities can be inserted into different time periods of the speech to increase the randomness of the data.

[0110] S303. Adjust the signal-to-noise ratio of the noisy analog speech to a preset range to obtain noisy synthesized speech data.

[0111] Use audio analysis tools to calculate the current signal-to-noise ratio of the noisy speech;

[0112] If the current signal-to-noise ratio is lower than the preset range, improve it using the following methods:

[0113] Attenuate noise signals proportionally;

[0114] Alternatively, the intensity of the speech can be increased using audio amplification algorithms.

[0115] If the signal-to-noise ratio is higher than the preset range, the operation is reversed to generate noisy synthesized speech data to ensure that the data matches the scene's realism.

[0116] Add metadata to each piece of noisy synthesized speech data, including:

[0117] The original text source (e.g., user-corrected text ID, background document path);

[0118] Scene label, error type, priority score;

[0119] Synthesis parameters (timbre, speech rate, noise type)

[0120] S4. Fine-tune the parameters of the temporary model of the speech recognition model using the training sample dataset to obtain the evolution parameters of the speech recognition model. The training sample dataset includes high-value text segments and corresponding speech data.

[0121] A temporary model is an intermediate model used in data processing and machine learning. Its main parameters are the same as those of the formal model, and it is mainly used to iteratively optimize the formal model in a safe environment.

[0122] Specifically, the high-value text fragments extracted in step S2 are precisely paired with the corresponding noisy synthetic speech data generated in step S3 through meta-information; the text and speech are aligned by timestamps or tags to ensure that each text fragment corresponds to the time interval of its speech waveform, forming a training sample data set.

[0123] In a preferred embodiment, the training sample data set further includes: general text fragments and corresponding speech data, and the general text fragments are text fragments selected from a large-scale general text sample training set. The large-scale general text sample training set is a pre-existing, large-scale sample training set containing general speech recognition knowledge. In this embodiment, in the training sample data set, a part of general text fragments and corresponding speech data are mixed in to avoid significantly forgetting the original speech recognition ability after the high-value text and corresponding speech data adjust the parameters of the temporary model during the training of the temporary model.

[0124] Create a lightweight temporary model based on the original speech recognition model:

[0125] The lightweight temporary model copies the structure of the original speech recognition model but only initializes some parameters (such as the weights of key layers), and other parameters share the original speech recognition model;

[0126] The temporary model is used for rapid iterative training to avoid directly modifying the main model in the production environment.

[0127] Perform low-rank approximation on the key parameter matrices (such as the attention matrix and the weights of the fully connected layer) of the temporary model:

[0128] Use singular value decomposition or low-rank neural networks to decompose the high-dimensional parameter matrix into low-rank factor matrices;

[0129] For example, decompose the weight matrix W ∈ R m×n into a low-rank matrix U ∈ R m×r and a low-rank matrix V ∈ R r×n , where r << min(m, n), to reduce the number of parameters.

[0130] Input the speech data (after feature extraction such as Mel spectrum) and text labels, calculate the error between the model output and the labels; adopt the connectionist temporal classification loss or cross-entropy loss, combined with scene weighting (for example, assigning a higher penalty coefficient to errors in the medical scene); calculate the gradient through backpropagation and only update the parameters of the temporary model.

[0131] In each iteration, maintain the low-rank structure in the following way:

[0132] Project the updated parameters back to the low-rank space to prevent the rank from increasing;

[0133] The rank value is adaptively increased based on the training progress (e.g., gradually increasing from r=10 to r=20) to balance training accuracy and efficiency.

[0134] Evaluate the word error rate, sentence error rate, and scene recognition accuracy of the temporary model on the validation set (which contains unseen scene data), and calculate the correction rate for specific error types.

[0135] Specifically, the difference parameters ΔU and ΔV between the trained low-rank matrices U′ and V′ and the initial fixed basis parameters can be calculated:

[0136] ΔU=U′-U 初始 ,ΔV=V′-V 初始 U 初始 and V 初始 This is the initial low-rank matrix.

[0137] By saving only the difference parameters instead of the complete matrix, the update package size is significantly reduced.

[0138] Package the following content as evolution parameters:

[0139] The difference low-rank matrix (ΔU, ΔV) and its corresponding rank r;

[0140] Version control information (including timestamps, scene tags, and performance metrics).

[0141] After completing the low-rank adaptive training of the temporary model and obtaining the evolutionary parameters, this step uses an experience replay mechanism to enhance and optimize the model parameters, avoiding training data order dependencies and improving model robustness. The specific implementation process is as follows:

[0142] S401: Evolutionary parameters are stored in the experience replay buffer.

[0143] The evolutionary parameters obtained in each iteration (i.e., the low-rank matrix difference update ΔU, ΔV and corresponding metadata after training) are stored in the experience replay buffer.

[0144] It should be noted that metadata may include: training scenario labels, sample IDs, training rounds, performance metrics, etc.

[0145] Data can be managed using First-In-First-Out (FIFO) or priority queue strategies, with the option to set maximum capacity limits and periodically clean up low-value samples. For example, parameter versions with performance improvements below a threshold can be prioritized for removal.

[0146] S402: Randomly sample empirical tuples and adjust the model.

[0147] Randomly sample K empirical tuples from the buffer (K ​​can be dynamically adjusted based on computing resources);

[0148] High-performance enhancement samples can be prioritized for target scenarios (such as medical and legal).

[0149] Parameter fusion and training:

[0150] The difference parameters in the sampled empirical tuples are fused with the parameters of the current temporary model:

[0151]

[0152] Among them, W 融合 For the weights after fusion, W 当前 Let αi be the current weight value, αi be the coefficient of the i-th weight, K represent the number of weights, and ΔU be the weight value. i Let i be the update amount for the i-th weight. Let be the transpose update amount of the i-th weight.

[0153] It should be noted that the temporary model is retrained using the fused parameters, with only the low-rank matrix part being fine-tuned, while the fixed basis parameters remain unchanged.

[0154] The system will terminate when the preset number of sampling rounds is reached, the buffer data is exhausted, or the model performance converges on the validation set or resource limitations are triggered.

[0155] S403, Preparation for Deploying Optimized Parameters

[0156] The updated low-rank matrix difference is extracted from the temporary model after reinforcement training to generate the final set of evolutionary parameters.

[0157] Merge high-performance parameter versions from multiple iterations to form a more robust update package.

[0158] Before deployment, the optimized model can be subjected to more rigorous verification (such as adversarial testing and multi-scenario cross-validation) to ensure that the performance improvement meets the preset threshold and has no significant side effects.

[0159] S5. When the key indicators of the temporary model reach the threshold, the user actively triggers the deployment command, or the performance of the speech recognition model is detected to decline, the evolution parameters will be deployed to the speech recognition model.

[0160] Among them, the key indicators of the temporary storage model reach the threshold under the following conditions:

[0161] Evaluate whether the key metrics of the temporary model reach the threshold on the validation set:

[0162] Improved recognition accuracy: such as a decrease in word error rate of ≥X% (X is a preset value, such as 5%);

[0163] Improved scene adaptability: such as an increase in recognition accuracy of ≥Y% in target scenes (e.g., medical, financial) (Y is a preset value, such as 5%).

[0164] Resource consumption does not exceed the maximum available memory: for example, the increase in model inference latency does not exceed Z% (Z is a preset value, such as 10%), and memory usage is acceptable.

[0165] As a preferred implementation, the preset conditions may also include: the verification results are stable and there is no performance fluctuation for N consecutive iteration cycles (e.g., N=5); and the parameter updates are ensured to not introduce security vulnerabilities through adversarial sample testing.

[0166] Furthermore, differential evolution parameters (including ΔU, ΔV, and rank r information) can be extracted from the evolution parameters.

[0167] Perform the following operations on the original model in the production environment:

[0168] Fixed basis parameters for locating the target layer (such as the attention layer or decoder);

[0169] Reconstruct the complete weights using the difference parameters.

[0170] As a preferred implementation, after training is complete, the parameter layers with significant performance improvements in the temporary model are extracted, and the evolution parameters are weighted and fused with the corresponding layers of the main model:

[0171] New parameter = α × main model parameter + (1-α) × evolution parameter, where α∈[0,1] is the fusion coefficient.

[0172] The fusion coefficient is dynamically adjusted based on the performance of the validation set, for example, by using Bayesian optimization or grid search to determine the optimal value. Evolutionary parameters are applied incrementally through the fusion coefficient α to reduce the risk of catastrophic forgetting.

[0173] The deployed model can be used as a new baseline model to start the next round of iterations from steps S2 to S5, and the training samples and scene weights can be dynamically adjusted based on user feedback or new data.

[0174] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0175] See Figure 4 This application also provides a self-evolving system for a speech recognition model, comprising:

[0176] The knowledge acquisition module is used to store background knowledge documents and user correction texts of the speech recognition model's output into a corpus.

[0177] The text extraction module is used to extract high-value text fragments from the corpus;

[0178] The data synthesis module is used to convert high-value text fragments into speech data;

[0179] The model fine-tuning module is used to fine-tune the parameters of the temporary model of the speech recognition model using the training sample dataset to obtain the evolution parameters of the speech recognition model. The training sample dataset includes high-value text segments and corresponding speech data.

[0180] The automated deployment module is used to deploy evolutionary parameters to the speech recognition model when preset conditions are met.

[0181] Based on the above embodiments, it can be seen that this application achieves technical breakthroughs and practical value through the following aspects:

[0182] 1. Efficiently build scenario-based speech recognition models to improve recognition accuracy and generalization ability:

[0183] This application enhances the model's ability to recognize specialized terminology, complex sentence structures, and noisy environments in specific scenarios (such as medical, financial, and customer service) through high-value text extraction and multi-style speech data synthesis. Parameter fine-tuning techniques effectively address the overfitting problem in small sample scenarios, enabling the model to maintain excellent performance even on unseen data.

[0184] 2. Significantly reduce model iteration costs, enabling lightweight updates and rapid deployment:

[0185] This application employs parameter decomposition and differential coding techniques to compress evolutionary parameters to 10%-30% of those using traditional methods, thereby significantly reducing storage and transmission overhead. The automated deployment process, combined with a progressive convergence strategy, supports minute-level deployment and zero-disruption switching, making it suitable for edge devices or resource-constrained scenarios.

[0186] 3. Construct a closed-loop optimization system to ensure the continuous evolution and stability of the model:

[0187] This application employs a closed-loop process of "data generation → low-rank training → differential update → deployment and verification," enabling continuous improvement in model performance as data accumulates. Pre-defined conditions, shadow testing, A / B testing, and rollback mechanisms ensure the safety and traceability of each iteration, avoiding the risks of performance degradation and catastrophic obsolescence.

[0188] 4. Flexible adaptation to various scenario requirements, balancing accuracy and efficiency:

[0189] This application supports scene-label-driven dynamic weight adjustment, enabling rapid optimization for single scenarios (such as operating room voice command recognition) or mixed scenarios (multi-industry customer service). The dynamic rank adjustment mechanism of the low-rank matrix allows users to flexibly configure model complexity according to actual needs, balancing resource consumption and recognition accuracy.

[0190] This application achieves "hot updates" of production environment models through a lightweight temporary storage model and incremental update technology, allowing the absorption of new data value without rebuilding the entire model. This method significantly shortens the model iteration cycle, reduces operation and maintenance costs, and provides a feasible solution for large-scale commercial applications.

[0191] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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. Such 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 this application, and should all be included within the protection scope of this application.

Claims

1. A self-evolutionary method for a speech recognition model, characterized in that, The method includes the following steps: The background knowledge document and the user's correction text of the speech recognition model's output are stored in the corpus; High-value text fragments were extracted from the corpus, including: Obtain the number of corrections and the error type of the corrected text, and label the scene of the corrected text; Based on the background knowledge document corresponding to the scene tag of the corrected text, obtain the keywords and text topic in the corrected text; The value score of the corrected text is determined based on the keywords, the text topic, the number of corrections, the error type, and the scene tag. Corrected texts in the corpus with value scores higher than the preset value are designated as high-value text fragments; Convert the high-value text fragments into speech data; The parameters of the temporary model of the speech recognition model are fine-tuned by using the training sample dataset to obtain the evolution parameters of the speech recognition model. The training sample dataset includes the high-value text segments and the corresponding speech data. When the key indicators of the temporary model reach a threshold, the user actively triggers a deployment command, or the performance of the speech recognition model is detected to decline, the evolution parameters are deployed to the speech recognition model.

2. The method as described in claim 1, characterized in that, The process of converting the high-value text fragment into speech data includes: The high-value text fragments are input into a text-to-speech engine to generate simulated speech data with different timbres, speech rates, and intonations. The text-to-speech engine supports multiple speaker style switching or timbre cloning based on user-provided speech samples. At least one scene noise is selected from a preset noise database and added to the simulated speech to obtain noisy simulated speech data; Adjust the signal-to-noise ratio of the noisy simulated speech to a preset range to obtain noisy synthesized speech data.

3. The method as described in claim 2, characterized in that, After converting the high-value text fragment into speech data, the method further includes: The noisy synthesized speech is paired with the corresponding high-value text fragments as training sample data, and the training sample data is labeled with scene type, speaker style, noise parameters and synthesis timestamp. A training sample dataset is constructed based on the labeled training sample data.

4. The method of claim 3, wherein the training sample dataset further comprises: General text fragments and corresponding speech data, wherein the general text fragments are text fragments selected from a large-scale general text sample training set.

5. The method as described in claim 1, characterized in that, After fine-tuning the parameters of the temporary model of the speech recognition model using the training sample dataset to obtain the evolution parameters of the speech recognition model, the method further includes: The evolutionary parameters are stored in the experience replay buffer; The evolutionary parameters are randomly sampled from the experience replay buffer to obtain experience tuples; The speech recognition model is adjusted based on the empirical tuple.

6. The method as described in claim 5, characterized in that, The step of fine-tuning the parameters of the temporary model of the speech recognition model using the training sample dataset to obtain the evolution parameters of the speech recognition model includes: Based on the parameters of the speech recognition model, a temporary model of the speech recognition model is constructed. The parameters of the temporary model are decomposed into fixed basis parameters and trainable low-rank matrices; The trainable low-rank matrix is ​​iteratively trained using the training sample dataset. Extract the difference update between the trained low-rank matrix and the fixed basis parameters after training to form an evolution parameter set; Differential encoding or incremental compression of the evolutionary parameter set generates deployable evolutionary parameters.

7. The method as described in claim 6, characterized in that, The key metrics of the temporary storage model reaching the threshold include: The performance metrics of the temporary model on the validation set reach a preset threshold. The number of training rounds reaches the preset limit or resource consumption meets the budget constraint.

8. The method as described in claim 7, characterized in that, The step of adjusting the speech recognition model based on empirical tuples includes: The evolutionary parameters are deployed to the speech recognition model, the trainable low-rank matrix replaces the original parameters of the corresponding layer of the speech recognition model, and the fixed basis parameters updated by differential are weighted and merged with the original parameters of the speech recognition model.

9. A self-evolving system for a speech recognition model, characterized in that, include: The knowledge acquisition module is used to store background knowledge documents and user correction texts of the speech recognition model's output into a corpus. The text extraction module is used to extract high-value text fragments from the corpus, including: Obtain the number of corrections and the error type of the corrected text, and label the scene of the corrected text; Based on the background knowledge document corresponding to the scene tag of the corrected text, obtain the keywords and text topic in the corrected text; The value score of the corrected text is determined based on the keywords, the text topic, the number of corrections, the error type, and the scene tag. Corrected texts in the corpus with value scores higher than the preset value are designated as high-value text fragments; The data synthesis module is used to convert the high-value text fragments into speech data; The model fine-tuning module is used to fine-tune the parameters of the temporary model of the speech recognition model using the training sample dataset to obtain the evolution parameters of the speech recognition model. The training sample dataset includes the high-value text segment and the corresponding speech data. An automated deployment module is used to deploy the evolution parameters to the speech recognition model when the key indicators of the temporary model reach a threshold, the user actively triggers a deployment command, or the performance of the speech recognition model is detected to decline.

Images