Self-evolution method and system of speech recognition model

Abstract

According to the self-evolution method and system of the speech recognition model, a training set is constructed by extracting high-value text fragments and combining multi-style speech synthesis data, after general data is combined, parameter optimization is conducted on a temporary storage model through a parameter fine tuning technology, and meanwhile, the optimized parameters are stored in an experience playback buffer area; and generating a lightweight update packet by differentially encoding the optimized parameters in the empirical playback buffer. According to the method, continuous evolution of the model is achieved through a closed-loop optimization process, model forgetting is avoided in combination with general data, the parameter quantity in the evolution process is remarkably reduced through parameter fine adjustment, and the training cost in the evolution process can be remarkably reduced through experience playback buffering; and the stability of model evolution in the evolution process is ensured.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and machine learning, and particularly relates to a self-evolution method and system of a speech recognition model. BACKGROUND

[0002] In the field of speech recognition technology, the mainstream method for improving the performance of existing models is to periodically collect a large amount of user speech data, manually annotate it, and then train the model offline. This method is time-consuming and costly. Some systems record the user's corrections to the recognition results (such as manually correcting the transcription text) to optimize the model. However, this method highly depends on the return of the original speech data, has privacy compliance risks, and consumes a lot of computing resources. There are also methods that use unlabeled data (such as untranscribed speech streams) for self-supervised learning of the model, or combine a small amount of labeled data for semi-supervised training. However, such methods can only improve the robustness of the model to existing knowledge, and cannot effectively learn new vocabulary or domain knowledge. There are also methods that generate additional training data through a text-to-speech synthesis model, such as using public text libraries to synthesize speech to expand the training set. However, existing methods often lack simulation of real-world scenarios, and the generated speech lacks diversity, resulting in insufficient generalization ability of the model.

[0003] Therefore, the existing technology cannot quickly respond to new vocabulary and new scenarios, resulting in poor user experience. To meet data compliance requirements, enterprises need to invest additional resources in data desensitization or anonymization processing, increasing the technical burden. End-side 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 is significantly reduced, affecting the expansion of application scenarios.

[0004] In summary, the existing technology cannot achieve efficient and low-cost self-evolution of speech recognition models while ensuring privacy security. SUMMARY

[0005] The purpose of the present application is to provide a self-evolution method and system of a speech recognition model that can break away from dependence on original speech data, quickly respond to new knowledge, ensure privacy security, and be efficient and low-cost, to solve the bottleneck that the existing technology cannot break through.

[0006] The first aspect of the embodiment of the present application provides a self-evolution method of a speech recognition model, which comprises the following steps:

[0007] storing background knowledge documents and user corrections to the output results of the speech recognition model into a corpus;

[0008] extracting high-value text segments from the corpus;

[0009] Converting the high-value text segment into voice data;

[0010] The evolution parameters of the speech recognition model are obtained by parameter fine-tuning of the temporary model of the speech recognition model through a training sample data set, the training sample data set including the high-value text segment and corresponding voice data;

[0011] The evolution parameters are deployed to the speech recognition model when the key indicators of the temporary model reach the threshold, the user initiatively triggers the deployment instruction, or the performance of the speech recognition model is detected to be degraded.

[0012] In other embodiments of the present application, the high-value text segment is extracted from the corpus, including:

[0013] The correction times of the corrected text and the error types of the corrected text are obtained, and the scenario labels of the corrected text are labeled;

[0014] According to the background knowledge document corresponding to the scenario label of the corrected text, the keywords and text topics in the corrected text are obtained;

[0015] The value score of the corrected text is determined according to the keywords, text topics, correction times, error types, and scenario labels;

[0016] The corrected text with a value score higher than a preset value in the corpus is set as a high-value text segment.

[0017] Further, the high-value text segment is converted into voice data, including:

[0018] The high-value text segment is input into a text-to-speech engine to generate simulated voice data with different timbres, speech rates, and tones, and the text-to-speech engine supports multi-speaker style switching or timbre cloning based on a user-provided voice sample;

[0019] At least one scene noise is selected from a preset noise database and added to the simulated voice to obtain noisy simulated voice data;

[0020] The signal-to-noise ratio of the noisy simulated voice is adjusted to a preset range to obtain noisy synthesized voice data.

[0021] Further, after converting the high-value text segment into voice data, the method further includes constructing a training sample data set:

[0022] The noisy synthesized voice and the corresponding high-value text segment are paired as training sample data, and the scene type, speaker style, noise parameter, and synthesis timestamp are labeled for the training sample data.

[0023] Further, the training sample data set further includes: general text segments and corresponding voice data, the general text segments being text segments selected from a large-scale general text sample training set.

[0024] Further, after the evolution parameters of the speech recognition model are obtained by parameter fine-tuning of the temporary model of the speech recognition model through the training sample data set, the method further comprises:

[0025] storing the evolution parameters into an experience replay buffer;

[0026] randomly sampling the evolution parameters from the experience replay buffer to obtain experience tuples;

[0027] adjusting the speech recognition model according to the experience tuples.

[0028] Further, the evolution parameters of the speech recognition model are obtained by parameter fine-tuning of the temporary model of the speech recognition model through the training sample data set, comprising:

[0029] constructing a temporary model of the speech recognition model based on the parameters of the speech recognition model,

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

[0031] iteratively training the trainable low-rank matrices using the training sample data set;

[0032] extracting a differential update of the trained trainable low-rank matrices and the fixed basis parameters to form an evolution parameter set;

[0033] differentially encoding or incrementally compressing the evolution parameter set to generate evolution parameters that can be efficiently deployed.

[0034] In other embodiments of the present application, the key indicators of the temporary model reaching the threshold value comprise:

[0035] triggering deployment when at least one of the following conditions is met:

[0036] the performance indicators of the temporary model on the validation set reach a preset threshold;

[0037] the training round reaches a preset upper limit or the resource consumption meets a budget constraint;

[0038] the user actively triggers a deployment instruction or detects that the model performance has significantly decreased.

[0039] Further, the speech recognition model is adjusted according to the experience tuples, comprising:

[0040] deploying the evolution parameters to the speech recognition model, replacing the original parameters of the corresponding layer of the speech recognition model with the trainable low-rank matrices, and performing weighted merging of the differential update of the fixed basis parameters and the original parameters of the speech recognition model.

[0041] The second aspect of the embodiment of the application provides a self-evolution system of a speech recognition model, comprising:

[0042] The knowledge acquisition module is configured to store the background knowledge document and the correction text of the user to the output result of the speech recognition model into a corpus;

[0043] The text extraction module is configured to extract a high-value text segment from the corpus;

[0044] The data synthesis module is configured to convert the high-value text segment into speech data;

[0045] The model fine-tuning module is configured to perform parameter fine-tuning on the temporary model of the speech recognition model through a training sample data set to obtain evolution parameters of the speech recognition model, wherein the training sample data set comprises the high-value text segment and corresponding speech data;

[0046] The automatic deployment module is configured to deploy the evolution parameters to the speech recognition model when a key indicator of the temporary model reaches a threshold value, a user initiatively triggers a deployment instruction, or a performance of the speech recognition model is detected to be degraded.

[0047] Compared with the prior art, the embodiment of the application has the beneficial effects that: the application extracts key knowledge such as new words and entities from user correction feedback and scene-related documents, and fine-tunes the model through text-to-speech synthesis technology to generate diversified training data. On the premise of ensuring user privacy and security, the application can capture and convert new knowledge in user feedback and documents in real time. The self-evolution of the model can be realized without human intervention, and the model can adapt to changing language and scene requirements. At the same time, the recognition error rate in professional scenes will be greatly reduced, and the terms in professional scenes can be accurately recognized. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A flowchart of a self-evolution method of a speech recognition model provided by the application;

[0049] Figure 2 A flowchart of scoring of stored text segments by a corpus provided by the embodiment of the application;

[0050] Figure 3 A speech data synthesis flowchart of the embodiment of the application;

[0051] Figure 4 A structure diagram of a self-evolution system of a speech recognition model provided by the embodiment of the application. DETAILED DESCRIPTION

[0052] In order to make the technical problems, technical solutions and beneficial effects of the present application clearer, the present application will be further described in detail below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0053] Figure 1 A flowchart of a self-evolution method of a speech recognition model according to the first embodiment of the present application is shown, and will be described in detail as follows:

[0054] S1, store the background knowledge documents and the user's correction text of the output result of the speech recognition model into the corpus.

[0055] Specifically, the system monitors and accesses the background knowledge documents and the correction text in real time, wherein the background knowledge documents, i.e. the text materials related to the speech recognition scene, include:

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

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

[0058] Scene-specific data, for example, in the medical scene, access electronic medical record templates, drug instruction manuals, or in the customer service scene, access product documents.

[0059] User correction text, i.e. the text generated by the user's correction behavior on the output result of the model, for example:

[0060] In the meeting scene, the user manually corrects the text transcribed by the speech recognition model.

[0061] After processing the above background knowledge documents and correction text, store them into the special corpus, which can be divided into "background knowledge area" (classified by document type) and "user correction area" (classified by correction type and scene), and real-time append new data, and clean up low-value or outdated information regularly.

[0062] S2, extract high-value text segments from the corpus.

[0063] It should be noted that before extracting high-value text segments from the corpus, the corpus will score the stored text segments, for example, the present application uses a multi-dimensional scoring model, and the scoring indicators include but are not limited to:

[0064] Novelty score, by detecting whether the text contains new words or new entities outside the word list, through the probability distribution of the language model or character-level n-gram statistics calculation, the higher the novelty score of the text segment, the higher the value score;

[0065] Correcting frequency score, by counting the number of times the user corrects the entity corresponding to the text, the more corrections, the higher the value score;

[0066] Scene relevance score, according to the scene of the current model deployment (such as meeting, customer service, medical treatment), the terms strongly related to the scene (such as product name in the meeting, drug name in the medical treatment) are given higher value score;

[0067] Timeliness score, for the newly added text in recent period, give higher timeliness score to ensure that the latest knowledge has higher value score.

[0068] Specifically, see Figure 2 Value score of the corrected text, including the following steps:

[0069] S201, get the number of corrections of the corrected text and the error type of the corrected text, and mark the scene label of the corrected text.

[0070] Maintain a dynamic database to record the number of corrections of each corrected text. For example, every time the user corrects the model output "XX product" to "YY brand", the number of corresponding corrections is increased by 1.

[0071] Automatic identification of error type of the corrected text by natural language processing technology, exemplary, common types include:

[0072] Wrong word (such as "sapphire" → "sapphire");

[0073] Homonym confusion (such as "ten points" → "test point");

[0074] Term missing (such as not recognizing the professional term "photon quantum chip");

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

[0076] Error type labeling provides targeted optimization direction for subsequent fine-tuning.

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

[0078] S202, according to the background knowledge document corresponding to the scene label of the corrected text, get the key words and text theme in the corrected text.

[0079] The system retrieves a set of background knowledge documents corresponding to the scene tag, and extracts keywords from the set of background knowledge documents after multi-document fusion. For example, product names in conference PPTs and technical terms in industry white papers can be extracted simultaneously.

[0080] S203, determining a value score of the corrected text according to the keyword, the text topic, the number of corrections, the error type, and the scene tag.

[0081] The value score (V) is calculated by weighted summation. An example is as follows:

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

[0083] Error type (weight w2): different error types are given different weights (for example, term missing 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 is improved;

[0085] Topic heat (weight w4): if the entity belongs to the high-frequency topic of the current scene (such as “financial data” in the conference) or new knowledge, the score is increased.

[0086] S204, setting the corrected text with a value score higher than a preset value in the corpus as a high-value text segment.

[0087] All corrected texts are sorted by score, and the top N% of texts are extracted.

[0088] S3, converting the high-value text segment into voice data.

[0089] It should be noted that in the present application, the text can be segmented, and specific natural language tools such as jieba and spaCy can be used to identify and label special entities; for multi-pronunciation words, the correct pronunciation is selected through context semantics or scene knowledge base.

[0090] Then, insert prosodic symbols (such as pauses and stress marks) in the text to optimize the naturalness of synthesized speech. User-defined prosodic templates can be supported, for example, a more formal tone is used in the conference scenario, and a friendly voice rhythm is used in the customer service scenario.

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

[0092] The conference scenario can use a slower speaking speed and clear pronunciation to highlight key terms;

[0093] Customer service scenarios can use a warm tone and natural intonation to avoid a mechanical feel;

[0094] Medical scenarios can emphasize the accurate pronunciation of professional terms to avoid ambiguity.

[0095] Specifically, referring to Figure 3 Convert high-value text segments into speech data, which includes the following steps:

[0096] S301, input high-value text segments into a text-to-speech engine to generate simulated speech data with different tones, speech rates, and intonations. The text-to-speech engine supports multiple speaker style switching or tone cloning based on user-provided voice samples.

[0097] Input high-value text segments into a text-to-speech engine and randomly configure the following parameters:

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

[0099] Speech rate adjustment configuration: customize the speech rate coefficient (such as 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, raise the intonation in important notifications to enhance the tone.

[0101] Generate multiple variants for the same text segment: different tones x different speech rates x different intonations, for example, generate "formal male voice-normal speech rate", "warm female voice-slow" and other versions; Random parameters can be disturbed, such as fine-tuning the intonation curve or inserting random pauses to increase data diversity.

[0102] Optionally, with the user's permission, the system can collect their voice samples (such as 10 minutes or more of recordings) to achieve tone cloning through the following steps:

[0103] Extract the acoustic features (such as spectrum, fundamental frequency, formant) of the user's voice;

[0104] Use a vocoder combined with a pre-trained model to generate a text-to-speech model consistent with the user's tone.

[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] Pre-constructed scene-based noise database, covering common environmental sounds (such as conference room, street, cafe, hospital, etc.), users can also upload noise samples, for example, upload the recording of a specific conference room as a dedicated noise source.

[0107] According to the scene label of the text, at least one noise type is selected from the noise library corresponding to the scene;

[0108] Noise can be mixed, for example, adding air conditioner noise and keyboard clicking sound at the same time to better simulate the real conference room scene.

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

[0110] S303, adjust the signal-to-noise ratio of the noisy simulated 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, it can be improved by the following method:

[0113] Proportionally attenuate the noise signal;

[0114] Or, through an audio amplification algorithm to increase the intensity of the speech part.

[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 meets the scene reality.

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

[0117] Original text source (such as user correction text ID, background document path);

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

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

[0120] S4, parameter fine-tuning of the temporary model of the speech recognition model through the training sample data set to obtain the evolution parameters of the speech recognition model. The training sample data set includes high-value text fragments and corresponding speech data.

[0121] The temporary model refers to an intermediate model in data processing and machine learning, which has the same main parameters as the formal model and is mainly used for iterative optimization of the formal model in a safe environment.

[0122] Specifically, the high-value text segments extracted in step S2 are accurately paired with the noisy synthesized speech data generated in step S3 through meta information; the text and the speech are aligned through timestamps or labels, ensuring that each text segment corresponds to the time interval of its speech waveform, forming a training sample dataset.

[0123] In a preferred embodiment, the training sample dataset further includes: general text segments and corresponding speech data, the general text segments being text segments 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, a portion of general text segments and corresponding speech data are mixed into the training sample dataset to avoid the original speech recognition capability being significantly forgotten after the high-value text and corresponding speech data adjust the parameters of the staging model during the training of the staging model.

[0124] Creating a lightweight staging model based on the original speech recognition model:

[0125] The lightweight staging model replicates the structure of the original speech recognition model but only initializes part of the parameters (such as key layer weights), and other parameters are shared with the original speech recognition model.

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

[0127] Low-rank approximation is performed on the key parameter matrices (such as attention matrices and fully connected layer weights) of the staging model:

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

[0129] For example, a weight matrix W ∈ R m×n is decomposed 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] The input speech data (after feature extraction such as Mel spectrum) and text labels are calculated to obtain the error between the model output and the labels; connection time classification loss or cross-entropy loss is used, combined with scene weighting (for example, higher penalty coefficients are given to errors in medical scenarios); the gradient is calculated through backpropagation, and only the parameters of the staging model are updated.

[0131] In each iteration, the low-rank structure is maintained by:

[0132] Projecting the updated parameters back into the low-rank space to prevent rank increase;

[0133] According to the training progress, the rank value is adaptively increased (for example, gradually increased from r=10 to r=20) to balance the training accuracy and efficiency.

[0134] The word error rate, sentence error rate, scene recognition accuracy of the staging model are evaluated on the validation set (containing unseen scene data), and the correction rate of specific error types is calculated.

[0135] Specifically, the difference parameters ΔU and ΔV of 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 . Wherein U 初始 and V 初始 are initial low-rank matrices.

[0137] By saving only the difference parameters instead of the complete matrices, the update package volume is greatly reduced.

[0138] The following are packaged as evolution parameters:

[0139] Difference low-rank matrices (ΔU, ΔV) and corresponding rank r;

[0140] Version control information (including timestamp, scene label, performance index).

[0141] After completing the low-rank adaptive training of the staging model and obtaining the evolution parameters, this step optimizes the model parameters through the experience replay mechanism, avoids the sequence dependence of the training data, and improves the robustness of the model. The specific implementation process is as follows:

[0142] S401: Store evolution parameters in experience replay buffer

[0143] Store the evolution parameters obtained in each iteration, i.e., the difference updates ΔU, ΔV of the trained low-rank matrices and the corresponding metadata, in the experience replay buffer;

[0144] It should be noted that the metadata can include: training scene label, sample ID, training round, performance index, etc.

[0145] Advanced first-in, first-out (FIFO) or priority queue strategy is used to manage data, and the maximum capacity limit can be set; and low-value samples are periodically cleaned up, for example, parameter versions with performance improvement below a threshold value can be preferentially removed.

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

[0147] Randomly sample K experience tuples (K can be dynamically adjusted according to computing resources) from the buffer;

[0148] High-performance promotion samples in target scenarios (such as medical care and law) can be selected preferentially.

[0149] Parameter fusion and training:

[0150] The differential parameters in the experience tuples obtained by sampling are fused with the current staging model:

[0151]

[0152] where W 融合 is the fused weight, W 当前 is the current weight value, ai is the coefficient of the ith weight, K represents the number of weights, ΔU i is the update amount of the ith weight, is the transposed update amount of the ith weight.

[0153] It should be noted that the staging model is retrained using the fused parameters, only the low-rank matrix part is fine-tuned, and the base parameters remain unchanged.

[0154] When the preset sampling round is reached or the buffer data is exhausted, or the model performance converges on the validation set or the resource limit triggers termination.

[0155] S403, deployment of optimized parameters

[0156] The updated low-rank matrix difference is extracted from the reinforcement training staging model to generate a final evolution parameter set;

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

[0158] Before deployment, the experience replay optimized model can be subjected to more stringent 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 staging model reach the threshold, the user actively triggers the deployment instruction, or the performance of the speech recognition model is detected to decline, the evolution parameters are deployed to the speech recognition model.

[0160] Wherein, the key indicators of the staging model reaching the threshold include the following conditions:

[0161] Evaluate whether the key indicators of the staging model reach the threshold on the validation set:

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

[0163] Scenario adaptation improvement: such as recognition accuracy improvement in target scenarios (such as medical care and finance) ≥ Y% (Y is a preset value, such as 5%).

[0164] Resource consumption does not exceed the maximum available memory: such as model inference delay increases by no more than Z% (Z is a preset value, such as 10%), and the memory occupation is acceptable.

[0165] As a preferred embodiment, the preset condition can further include: the verification result is stable for N consecutive iteration periods (such as N = 5) without performance fluctuations; and through the adversarial sample test, it is ensured that the parameter update does not introduce security vulnerabilities.

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

[0167] The following operations are performed on the original model in the production environment:

[0168] Locate the fixed base parameters of the target layer (such as the attention layer, the decoder);

[0169] Reconstruct the complete weight using the differential parameters.

[0170] As a preferred embodiment, after the training is completed, the parameter layer with significant performance improvement in the staging model is extracted, and the evolution parameters are weighted and fused with the corresponding layer of the main model:

[0171] New parameters = a x main model parameters + (1-a) x evolution parameters, where a is a fusion coefficient.

[0172] The fusion coefficient is dynamically adjusted through the validation set performance, for example, the Bayesian optimization or grid search is used to determine the optimal value. The evolution parameters are gradually applied through the fusion coefficient a to reduce the risk of catastrophic forgetting.

[0173] The model after this deployment can be used as a new benchmark model to start the next round of iteration process of steps S2 to S5, and dynamically adjust the training sample and scene weight according to user feedback or new data.

[0174] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0175] Referring to Figure 4 , the present application also provides a self-evolution system of a speech recognition model, comprising:

[0176] A knowledge acquisition module is used to store background knowledge documents and user correction texts of output results of the speech recognition model into a corpus;

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

[0178] a data synthesis module configured to convert the high-value text segment into voice data;

[0179] a model fine-tuning module configured to fine-tune a temporary model of the speech recognition model by using a training sample data set to obtain evolution parameters of the speech recognition model, the training sample data set including the high-value text segment and corresponding voice data;

[0180] an automated deployment module configured to deploy the evolution parameters to the speech recognition model when a preset condition is met.

[0181] Based on the above embodiments, it can be seen that the present application realizes technical breakthroughs and practical values in the following aspects:

[0182] 1. Efficiently constructing a scenario-based speech recognition model to improve recognition accuracy and generalization ability:

[0183] The present application enhances the recognition ability of the model for specific scenarios (such as medical, financial, customer service, etc.) by extracting high-value text and synthesizing multi-style voice data. The parameter fine-tuning technology effectively solves the problem of model overfitting in small sample scenarios, enabling the model to still perform excellently on unseen data.

[0184] 2. Significantly reducing model iteration cost to realize lightweight updating and rapid deployment:

[0185] The present application uses parameter decomposition and differential coding technology to compress the evolution parameters to 10%-30% of the traditional method, thereby significantly reducing storage and transmission overhead. The automated deployment process combined with the gradual fusion strategy supports minute-level deployment and zero-interruption switching, suitable for edge devices or resource-constrained scenarios.

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

[0187] The present application uses a closed-loop process of "data generation → low-rank training → differential update → deployment verification", and the model performance can be continuously improved with data accumulation. Through preset condition judgment, shadow testing, AB experiment and rollback mechanism, the safety and effect of each iteration can be traced back, avoiding the risk of performance degradation and catastrophic forgetting.

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

[0189] The present application supports scenario tag-driven dynamic weight adjustment, which can quickly optimize 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 the model complexity according to actual needs, balancing resource consumption and recognition accuracy.

[0190] The application realizes the "hot update" of the production environment model by the lightweight temporary storage model and the incremental update technology, and absorbs the new data value without rebuilding the entire model. The method significantly shortens the model iteration cycle, reduces the operation and maintenance cost, and provides a feasible solution for large-scale commercial application.

[0191] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the foregoing embodiments of the present application are described in detail, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A self-evolution method for a speech recognition model, characterized in that: The method comprises the following steps, Storing background knowledge documents and user corrections to the output of the speech recognition model into a corpus; extracting high-value text snippets from the corpus; Converting the high-value text segments into voice data; Fine-tuning parameters of a temporary model of the speech recognition model using a training sample data set to obtain evolutionary parameters of the speech recognition model, wherein the training sample data set includes the high-value text segment and the corresponding speech data; When the key indicators of the temporary model reach a threshold, the user actively triggers a deployment instruction, or the performance of the speech recognition model is detected to have degraded, the evolution parameters are deployed to the speech recognition model.

2. The method according to claim 1, wherein The extracting high-value text segments from the corpus includes: Obtaining the number of corrections of the corrected text and the error type of the corrected text, and marking a scene label of the corrected text; Obtaining keywords and text topics in the corrected text according to the background knowledge document corresponding to the scene label of the corrected text; Determining a value score of the corrected text based on the keywords, the text theme, the number of corrections, the error type, and the scene label; Corrected texts in the corpus with a value score higher than a preset value are set as high-value text segments.

3. The method according to claim 1, wherein The converting the high-value text segment into voice data comprises: Inputting the high-value text fragment into a text-to-speech engine to generate simulated voice data with different timbres, speaking speeds, and intonations. The text-to-speech engine supports multi-speaker style switching or voice cloning based on user-provided voice samples; Selecting at least one scene noise from a preset noise database and adding it to the simulated speech to obtain noisy simulated speech data; The signal-to-noise ratio of the noisy simulated speech is adjusted to a preset range to obtain noisy synthesized speech data.

4. The method according to claim 3, wherein After converting the high-value text segment into voice data, the method further includes: The noisy synthesized speech is paired with the corresponding high-value text segment as training sample data, and the training sample data is annotated with scene type, speaker style, noise parameters and synthesis timestamp, and a training sample dataset is constructed based on the annotated training sample data.

5. The method according to claim 4, wherein the training sample data set further comprises: A general text segment and corresponding speech data, wherein the general text segment is a text segment selected from a large-scale general text sample training set.

6. The method according to claim 1, wherein After fine-tuning the parameters of the temporary model of the speech recognition model using the training sample data set to obtain the evolution parameters of the speech recognition model, the method further includes: Storing the evolution parameters in an experience replay buffer; Randomly sampling the evolution parameters from the experience replay buffer to obtain an experience tuple; The speech recognition model is adjusted according to the experience tuple.

7. The method according to claim 6, wherein Fine-tuning the parameters of the temporary model of the speech recognition model using the training sample data set to obtain the evolutionary 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, Decomposing the parameters of the temporary model into fixed basis parameters and a trainable low-rank matrix; Iteratively training the trainable low-rank matrix using the training sample data set; Extracting the differential update of the trained trainable low-rank matrix and the fixed basis parameters to form an evolutionary parameter set; Perform differential encoding or incremental compression on the evolution parameter set to generate deployable evolution parameters.

8. The method according to claim 7, wherein The key indicators of the temporary model reach the threshold, including: The performance index of the temporary model on the validation set reaches a preset threshold; The number of training rounds reaches the preset upper limit or resource consumption meets the budget constraint.

9. The method according to claim 8, wherein The step of adjusting the speech recognition model according to the experience tuple includes: The evolved 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 the differential are weighted and merged with the original parameters of the speech recognition model.

10. A self-evolution system for a speech recognition model, characterized in that: include: The knowledge acquisition module is used to store background knowledge documents and user correction text of the output results of the speech recognition model into the corpus; A text extraction module, configured to extract high-value text fragments from the corpus; A data synthesis module, configured to convert the high-value text segments into voice data; A model fine-tuning module, configured to fine-tune the parameters of the temporary model of the speech recognition model using a training sample data set to obtain evolutionary parameters of the speech recognition model, wherein the training sample data set includes the high-value text segment and corresponding speech data; An automatic 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 instruction, or the performance of the speech recognition model is detected to have degraded.

Images