Method and device for testing voice analysis system based on large model

Abstract

The invention discloses a method and device for testing a voice analysis system based on a large model, and relates to the technical field of software testing, and the method comprises the steps: generating a voice for testing through a testing model, the training module is obtained by training a machine learning model according to the voice data and text data, language data, speech speed data, accent data and noise data corresponding to the voice data; inputting the voice for testing into the voice analysis system to obtain an analysis text; and obtaining a test result according to the analysis text. According to the embodiment of the invention, various test voices with various characteristics are generated by using the test model, so that the coverage of the test voices is wide, the acquisition difficulty of the voice data set is reduced, the voice analysis system can be comprehensively evaluated, and the accuracy, reliability and automation degree of the test are improved.

Description

Technical Field

[0001] This invention relates to the field of software testing technology, and in particular to a testing method and apparatus for a speech parsing system based on a large model. Background Technology

[0002] This section is intended to provide background or context for the embodiments of the invention set forth in the claims. The description herein is not an admission that it is prior art simply because it is included in this section.

[0003] Speech-to-text (STT) is a speech recognition program that converts spoken language into written language. Currently, speech parsing models based on machine learning techniques are becoming increasingly large and complex. Testing these models often requires massive amounts of speech samples with diverse features, as well as substantial computational resources. However, speech samples and computational resources are often hard to obtain, making it easy for speech parsing models to face problems such as long testing times and inaccurate results during testing. Summary of the Invention

[0004] This invention provides a testing method for a speech parsing system based on a large model, which at least solves the problems of difficulty in obtaining speech test samples, low testing efficiency, and poor accuracy. The method includes: Test speech is generated through a test model, wherein the test model is obtained by training a machine learning model based on speech data and corresponding text data, language data, speech rate data, accent data and noise data. The test speech is input into the speech parsing system to obtain the parsed text; The test results are obtained based on the parsed text.

[0005] Advantageously, after obtaining the test results based on the parsed text, the process includes: Output an optimization strategy for the speech parsing system based on the test results; or Optimize the speech parsing system based on the test results; or The speech parsing system is optimized based on the test results until the test results meet the preset expectations.

[0006] Advantageously, test speech is generated through the test model, including: Based on the preset test text, the test model generates test speech; or Test speech is generated directly from the test model, and corresponding test text is generated based on the test speech; or Based on preset test text, language settings, speech rate settings, accent settings, and noise settings, the test model generates test speech.

[0007] Advantageously, the test results are obtained from the parsed text, including: By comparing the parsed text and the test text, a first test result is obtained; and / or The second test result is obtained based on the time taken to obtain the parsed text.

[0008] Advantageously, the test model includes an input control layer, a text processing module, a conditional and control coding layer, an acoustic model, and a vocoder; based on preset test text, language settings, speech rate settings, accent settings, and noise settings, the test model generates test speech, including: The input control layer receives the test text, language settings, speech rate settings, accent settings, and noise settings. Based on the test text, a standard phoneme sequence is generated by the text processing module; Based on the language setting, speech rate setting, accent setting, and noise setting, a corresponding conditional control embedding vector is generated through the conditional and control coding layer. Based on the standard sonic sequence and the conditional control embedding vector, a Mel spectrogram is generated through the acoustic model, wherein the acoustic model includes a text encoder, a variant adapter, and a Mel spectrogram generator. The standard phoneme sequence is encoded by the text encoder, and the duration, fundamental frequency, and energy characteristics of the elements in the standard phoneme sequence are adjusted according to the conditional control embedding vector to obtain an adapted phoneme sequence. Based on the adapted phoneme sequence, the Mel spectrogram is generated through the Mel spectrogram generator. The test speech is obtained through the vocoder based on the Mel spectrogram.

[0009] Advantageously, the test model further includes a post-processing module; based on preset test text, language settings, speech rate settings, accent settings, and noise settings, the test model generates test speech, and further includes: The post-processing module mixes noise with the test speech obtained from the vocoder to obtain a variant speech of the test speech.

[0010] Advantageously, the test task is divided into at least two test subtasks before generating test speech through the test model; and Each of the test subtasks is allocated corresponding computing resources according to preset rules; and / or The test subtasks are executed in parallel.

[0011] This invention also provides a testing device for a speech parsing system based on a large model, which at least solves the problems of difficulty in obtaining speech test samples, low testing efficiency, and poor accuracy. The device includes: The generation module is used to generate test speech through a test model, wherein the test model is obtained by training a machine learning model based on the speech data and the corresponding text data, language data, speech rate data, accent data and noise data. An input module is used to input the test speech into the speech parsing system to obtain parsed text; and The acquisition module is used to obtain test results based on the parsed text.

[0012] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described test method for the large-model-based speech parsing system.

[0013] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described test method for a large-model-based speech parsing system.

[0014] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described testing method for a large-model-based speech parsing system.

[0015] In this embodiment of the invention, test speech is generated through a test model, which is obtained by training a machine learning model based on speech data and corresponding text data, language data, speech rate data, accent data, and noise data. The test speech is then input into the speech parsing system to obtain parsed text. Test results are obtained based on the parsed text. This invention generates diverse test speech with various features using a machine learning model, thus broadening the coverage of test speech, reducing the difficulty of obtaining speech datasets, comprehensively evaluating the speech parsing system, and improving the accuracy, reliability, and automation of the test. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a schematic flowchart illustrating an example of a testing method for a large-model-based speech parsing system according to an embodiment of the present invention. Figure 2 A schematic flowchart illustrating an example of allocating test tasks according to an embodiment of the present invention; Figure 3 A schematic flowchart illustrating yet another example of allocating test tasks according to an embodiment of the present invention; Figure 4 A schematic diagram illustrating an example of generating test speech for a test model according to an embodiment of the present invention; Figure 5 A schematic structural diagram of an example test apparatus for a large-model-based speech parsing system according to an embodiment of the present invention; Figure 6 This is a schematic structural diagram of an example of a computer device according to an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0018] Figure 1 This is a schematic flowchart illustrating an example of a testing method for a large-model-based speech parsing system according to an embodiment of the present invention. Figure 1 As shown, the testing method for a large-model-based speech parsing system according to an embodiment of the present invention includes: Step 102: Generate test speech using a test model, wherein the test model is obtained by training a machine learning model based on the speech data and the corresponding text data, language data, speech rate data, accent data and noise data. Step 104: Input the test speech into the speech parsing system to obtain the parsed text; Step 106: Obtain the test results based on the parsed text.

[0019] According to the testing method of the large-model-based speech parsing system of the present invention, by using the test model obtained by training a machine learning model based on speech data (raw data) and corresponding text data, language data, speech rate data, accent data and noise data (labels), a variety of test speech with various features can be generated. This can make the coverage of test speech broad, reduce the difficulty of obtaining speech datasets, comprehensively evaluate the speech parsing system, and improve the accuracy, reliability and automation of the test.

[0020] To make efficient use of computing resources and improve testing efficiency, according to embodiments of the present invention, before generating test speech through the test model, test tasks can be allocated, for example, by dividing the test tasks into test sub-tasks and then executing the test sub-tasks. Figure 2 This is a schematic flowchart illustrating an example of assigning test tasks according to an embodiment of the present invention. Figure 2 As shown, before generating test speech using the test model, the following can be included: Step 10: Divide the test task into at least two test subtasks; Step 12: Start and execute the test subtasks in parallel.

[0021] By dividing test tasks into multiple test subtasks and executing these subtasks in parallel, the time spent on testing can be reduced. To achieve parallel testing, multiple test subtasks can be distributed across multiple processors, computers, physical servers, virtual machines, or containers. Each processor, computer, physical server, virtual machine, or container can execute the test subtask independently and generate test results for that subtask. This allows problems to be pinpointed to specific test subtasks, improving the efficiency of error finding. Through a parallel testing architecture, a large number of test subtasks can be executed simultaneously on multiple platforms, significantly reducing overall testing time, which is crucial for rapid iteration and optimization of speech parsing systems.

[0022] Figure 3 This is a schematic flowchart illustrating yet another example of allocating test tasks according to an embodiment of the present invention. Figure 3 As shown in this example, before generating test speech using the test model, the following may be included: Step 10: Divide the test task into at least two test subtasks; Step 14: Allocate corresponding computing resources to each of the test subtasks according to preset rules.

[0023] By allocating corresponding computing resources to each test subtask according to preset rules, the utilization of computing resources can be maximized and waste avoided. The preset rules are not specifically limited; they can be based on the severity of errors obtained from previous tests, generally allocating more computing resources to more severe errors. Load balancing technology can also be used to ensure that the load on each test subtask node is optimal, avoiding situations where some test nodes are overloaded while others are idle. Effective resource management strategies ensure efficient utilization of computing resources, reduce hardware resource requirements, and lower testing costs.

[0024] It should be understood that the present invention is not limited thereto. In some other examples, before generating test speech through the test model, the following may be included: Divide the test task into at least two test sub-tasks; Each test subtask is allocated corresponding computing resources according to preset rules; The test subtasks are executed in parallel.

[0025] By pre-allocating computing resources and then conducting parallel tests, both testing efficiency and the waste of computing resources can be guaranteed, which will not be elaborated further here.

[0026] It should be understood that there can be multiple methods for generating test speech using a test model. In some embodiments, step 102, generating test speech using a test model, may include: Based on the preset test text, the test model generates test speech.

[0027] For example, by using a test model, speech synthesis technology and text-to-speech technology, corresponding test speech can be generated from a preset test text. The preset test text can be user input, pre-stored text, or text generated by the model itself. In addition, background noise and acoustic environment can be added during the speech generation process to simulate speech in various real-world environments as much as possible. This enables the speech parsing system to perform well in various real-world environments.

[0028] In some other examples, step 102, generating test speech through a test model, may include generating test speech through a test model based on preset test text, language settings, speech rate settings, accent settings, and noise settings.

[0029] For example, the text, language, speech rate, accent, and noise of the test voice to be generated can be directly preset, so that the generated voice is in a completely controlled state and can obtain any test voice that the user wants.

[0030] For example, and without limitation, the generation process is based on a pre-defined multi-dimensional parameter system. Test text parameters not only support regular sentences but can also be specifically designed with challenging content containing easily confused phonemes, technical terms, or complex grammatical structures to test the system's linguistic processing capabilities. Language settings can precisely specify the target language and its dialect variants, ensuring that the generated speech conforms to specific language norms in phonology and prosody. Speech rate settings typically use relative value control (e.g., 0.5x to 2.0x normal speech rate), capable of simulating various speaking styles from slow and clear to rapid connected speech. Accent settings are achieved through speaker embedding vectors, which can reproduce accent characteristics of specific regions or create continuously varying accent effects through vector interpolation techniques. Noise settings support precise control of multiple noise types, including selecting noise types (white noise, babble noise, street noise, etc.), setting the signal-to-noise ratio level, and configuring the room impulse response to simulate different acoustic environments.

[0031] This fully parameterized generation method offers significant advantages. First, it achieves perfect reproduction of test conditions; any test case can be accurately reproduced with the same parameter settings, ensuring the comparability of test results. Second, the system supports large-scale parallel generation, capable of generating tens of thousands of hours of test speech covering various edge cases in a short time, greatly expanding the test coverage. More importantly, by systematically adjusting a single parameter (such as changing only the signal-to-noise ratio while keeping other parameters constant), users can analyze the specific impact of that factor on system performance in isolation, thereby conducting precise attribution analysis.

[0032] For example, testers can generate a series of speech samples with only gradient differences in accent intensity. By analyzing the curve of recognition accuracy as a function of accent intensity, they can quantitatively assess the system's adaptability to different accents. Similarly, by generating speech samples with the same text content but with added noise from a train station, restaurant, and office, respectively, the robustness of the system under different noise environments can be compared.

[0033] In another example of the method for generating test speech according to an embodiment of the present invention, step 102, generating test speech through a test model, may include: Test speech is generated directly from the test model, and corresponding test text is generated based on the test speech.

[0034] For example, you can first synthesize the sound, and then obtain the corresponding text based on the synthesized sound, avoiding cumbersome processes such as inputting test text and test conditions, and improving testing efficiency.

[0035] For example, the test model integrates a dynamic text generation engine (e.g., based on a large language model) to automatically create grammatically correct, semantically fluent test texts covering different domains, lengths, and complexities. These generated texts, along with pre-defined, diverse test conditions (such as multiple languages, accents, speech rates, noisy environments, and vocal tract simulations), are then input into a subsequent speech generation module to synthesize a massive amount of highly diverse test speech. This speech, created by the test model, is then sent to the speech parsing system to be evaluated for analysis, resulting in recognized text. Finally, the test model automatically compares the original text initially generated by the dynamic text engine with the recognized text output by the speech parsing system. This process completely avoids the tedious manual preparation and input of test texts and configuration of test conditions, achieving end-to-end automation from test case generation to result comparison. This not only improves testing efficiency by several orders of magnitude but also exhaustively enumerates thousands of edge cases and complex scenarios that are difficult to conceive of in manual testing through programmatic means, greatly enhancing the breadth and depth of test coverage, thereby providing unprecedented stress testing and robustness evaluation of the speech parsing system.

[0036] When automatically generating voice data, a series of quality control measures can be used to ensure the accuracy and reliability of the data. For example, samples are manually reviewed, and preset programs are used to automatically detect generation errors and inconsistencies.

[0037] According to embodiments of the present invention, optionally, the test model may include an input control layer, a text processing module, a conditional and control coding layer, an acoustic model, and a vocoder; advantageously, generating test speech through the test model based on preset test text, language settings, speech rate settings, accent settings, and noise settings may include: The input control layer receives the test text, language settings, speech rate settings, accent settings, and noise settings. Based on the test text, a standard phoneme sequence is generated by the text processing module; Based on the language setting, speech rate setting, accent setting, and noise setting, a corresponding conditional control embedding vector is generated through the conditional and control coding layer. Based on the standard sonic sequence and the conditional control embedding vector, a Mel spectrogram is generated through the acoustic model, wherein the acoustic model includes a text encoder, a variant adapter, and a Mel spectrogram generator. The standard phoneme sequence is encoded by the text encoder, and the duration, fundamental frequency, and energy characteristics of the elements in the standard phoneme sequence are adjusted according to the conditional control embedding vector to obtain an adapted phoneme sequence. Based on the adapted phoneme sequence, the Mel spectrogram is generated through the Mel spectrogram generator. The test speech is obtained through the vocoder based on the Mel spectrogram.

[0038] The test model can be built upon a large-scale artificial intelligence model (i.e., a "big model," such as a large language model (LLM) or a multimodal big model). This allows the input control layer to transform from a passive parameter receiver into an active, intelligent test case generation engine. The big model can automatically generate massive amounts of diverse test text based on the test objectives; that is, preset test text, language settings, speech rate settings, accent settings, and noise settings can be automatically generated by the big model. For example, the big model can be instructed to "generate 100 business dialogues containing numbers, proper nouns, and complex grammatical structures," or "generate 50 sentences with near-homophones that are easily misheard in noisy environments." This greatly expands the coverage and relevance of test cases. The big model can understand natural language instructions and automatically synthesize complex test conditions. For example, testers only need to specify "simulate a scenario where a news report is broadcast at 1.5 times the normal speed with a slight Cantonese accent in a noisy subway environment," and the big model can parse the instruction and output the corresponding accent vector, noise and RIR configuration, and speech rate control parameters. By leveraging large models to gain a deep understanding of language and acoustics, we can proactively create rare but crucial edge cases in the real world to stress test the system's limits.

[0039] The test model can be a deep fusion of large-scale models and controllable TTS generation models. This invention can evolve from merely a speech generation tool into a highly intelligent automated test platform capable of understanding test requirements, autonomously designing experiments, and continuously driving performance optimization of the speech parsing system. This significantly improves the efficiency, depth, and breadth of testing, providing key infrastructure for the development of next-generation highly robust speech technologies.

[0040] For illustrative purposes only and not as a limitation, text data can be the text corresponding to the speech data to be generated by the model. Language data can include common languages ​​from various countries; speech rate data can include various common speech rates; accent data can include accent data from different age groups and genders, dialect data, pitch data, etc.; noise data can include various common noises, ambient sounds, and acoustic environments (such as echoes, far-field speech), etc. You can start with public datasets such as LibriSpeech or Common Voice, and then expand the dataset as needed to include more variations and scenarios. You can also assign different weights to text data, language data, speech rate data, accent data, and noise data, and if necessary, assign zero weight to at least one of the data types, and then generate the corresponding speech data for training.

[0041] Optionally, the test model may further include a post-processing module for further mixing the acoustic environment and / or noise obtained from the vocoder to obtain a variant speech of the test speech. Optionally, the test model may include at least one of deep neural networks, recurrent neural networks, and transformer networks. For example, deep neural networks may include DNN networks, which are suitable for advanced feature extraction; in speech and text conversion applications, DNNs can capture complex acoustic features. Recurrent neural networks may be RNN networks and their variants such as LSTM or GRU networks, which are superior for processing time-series data and are suitable for capturing temporal dependencies in speech data. Transformer networks may be Transformer networks, suitable for handling long-range dependencies, and can improve the model's ability to understand long-range context through self-attention mechanisms.

[0042] As another example, from a model structure perspective, the test model can employ a modular pipeline architecture to ensure precise injection and processing of control signals. For instance, the process begins with the input control layer, which receives a complete set of control parameters, including the test text, language settings, speech rate scale, accent (or speaker) vectors, noise, and room impulse response (RIR) configuration. Next, the text processing module cleans and normalizes the raw text, and performs multilingual segmentation and phonemeization based on the language settings, transforming it into a standard phoneme sequence. Simultaneously, the conditional and control coding layers work in parallel, encoding various discrete and continuous control parameters into dense embedding vectors, such as language embeddings, speaker / accent embeddings, style / prosodic tokens, and environmental vectors encoding the noise and RIR configuration. These conditional vectors are then injected into the core acoustic model, which can be based on a non-autoregressive architecture similar to FastSpeech2. The acoustic model first converts the phoneme sequence into a semantically rich hidden representation using a text encoder (which can be a Transformer encoder). Then, a crucial variant adapter uses control signals to precisely predict and adjust the duration, fundamental frequency (determining pitch), and energy (determining loudness) of each phoneme, specifically applying a speech rate control signal to scale the duration sequence. The adapted phoneme sequence is finally decoded into a Mel spectrogram by a Mel spectrogram generator. The generated Mel spectrogram is then fed into a neural vocoder (such as HiFi-GAN or DiffWave) to convert it into a high-quality time-domain waveform. In the final stage, the post-processing module strictly adheres to the input configuration, mixing the synthesized clean speech with specified background noise at a target signal-to-noise ratio, and applying RIR convolution to simulate reverberation effects in a specific space. The final output is a test speech waveform that meets all control conditions and closely resembles a real-world scenario.

[0043] Machine learning models can be trained using transfer learning strategies. For example, a model pre-trained on a large amount of general data can be used for a specific speech-to-text task; or a multilingual text-to-speech pre-trained model can be fine-tuned with multiple conditions (such as language, accent, speech rate, noise, etc.). This strategy leverages the powerful feature extraction capabilities of pre-trained models while reducing the amount of training data required for a specific task. During training, advanced optimizers such as Adam or AdaGrad can be used, which automatically adjust the learning rate during training, accelerating convergence and improving the performance of the final model. Since large-scale models are prone to overfitting, techniques such as Dropout, L2 regularization, and data augmentation can be used to mitigate this problem. For example, training samples can be enhanced by adding background noise, changing speech rate or pitch, SpecAugment-style spectrum enhancement, and multi-person speech mixing. Furthermore, the combined mechanism of mixed precision and gradient accumulation can be used to reduce the computational resource requirements during model training.

[0044] As another example, a multi-task fine-tuning strategy based on transfer learning can be employed for model training. The model is typically pre-trained on a large-scale, multilingual TTS dataset to learn basic speech generation capabilities. Subsequently, it is fine-tuned on data containing multiple languages, accents, speech rates, and noisy environments, enabling the model to learn to understand and respond to various control signals. The training process is optimized using a complex multi-task loss function that combines the acoustic model's spectral regression loss (L1 / L2), duration / fundamental frequency / energy prediction loss, as well as the vocoder's adversarial loss, feature matching loss, and Mel spectrum reconstruction loss, ensuring high quality in terms of auditory and acoustic features of the generated speech. To significantly improve the model's generalization ability and robustness, data augmentation techniques are extensively used during training, such as SpecAugment-style spectral masking, random noise and RIR mixing, velocity perturbation, and speaker feature mixing. At the optimization level, optimizers such as AdamW or Adafactor are used, along with adaptive learning rate scheduling such as cosine annealing, mixed-precision training, and gradient accumulation, ensuring the efficiency and stability of the training process.

[0045] The aforementioned test model, through its sophisticated conditional control mechanism, modular system structure, and comprehensive multi-task training paradigm, can transform speech synthesis from a simple generation technology into a powerful "test speech factory." It can not only generate test samples covering countless edge cases on demand, greatly reducing the cost of data collection and annotation, but also provide unprecedented controllability, repeatability, and scalability for speech technology research, development, and performance evaluation.

[0046] For example only, Figure 4 A schematic diagram illustrating yet another example of generating test speech for a test model according to an embodiment of the present invention. For example... Figure 4 As shown, the test model can include a control input module for inputting test text, language settings, speech rate settings, accent settings, and noise settings. A text processing module normalizes and performs multilingual word segmentation / phoneme conversion on the test text to obtain test text data. A conditional control module receives the language settings, speech rate settings, and accent settings and outputs corresponding control signals. An acoustic model encodes the text based on the test text data, language setting control signals, and accent setting control signals to obtain text-encoded data. Based on the speech rate control signal, the text-encoded data, and the accent setting control signal, it controls the duration, fundamental frequency, energy, and speech rate of the speech to be generated, thereby obtaining a variant adaptation signal. A corresponding Mel spectrum is generated based on the variant adaptation signal and the language setting control signal. This Mel spectrum is input into a vocoder to obtain clean speech data. Then, the clean speech data and the noise setting control signal are input into a post-processing module to mix the noise data and clean speech data to obtain test speech data. This allows for the synthesis of diverse test speech under controlled conditions, covering language, speech rate, accent, noise, and acoustic environment, supporting the construction of large-scale, reproducible, and quantifiable speech test sets.

[0047] In some embodiments, step 106, obtaining the test result based on the parsed text, may include: By comparing the parsed text and the test text, a first test result is obtained.

[0048] By comparing the parsed text obtained from the test speech of the speech parsing system with the test text of the test model, the first test result for the speech parsing system is obtained. The first test result may include: recognition errors (accuracy), substitution errors, insertion errors, deletion errors, the frequency of occurrence of different error types, the robustness of the system (such as performance stability when facing different environmental noise, accents, and changes in speaking style), and may even include a comprehensive evaluation report on the above indicators, which can provide in-depth information on the performance of the speech parsing system, enabling testers to accurately identify the system's strengths and weaknesses.

[0049] Specifically, the acquisition and analysis of the first test results is the core of performance evaluation. This process uses sophisticated text comparison algorithms to compare and analyze the parsed text output by the speech parsing system with the original test text word by word, thereby generating a comprehensive and quantitative performance evaluation report.

[0050] The test results not only include the basic word error rate, but also further subdivide them into three basic error types and their frequencies: substitution errors (such as recognizing "four" as "ten"), insertion errors (adding redundant words), and deletion errors (omitting keywords). This fine-grained data can accurately reveal the system's weaknesses in recognizing different linguistic units (phonemes, words, phrases). For example, high-frequency substitution errors may indicate that the acoustic model's ability to distinguish certain similar phonemes is insufficient, while a large number of insertion errors may reflect that the language model's confidence threshold is set too low.

[0051] More importantly, by analyzing the changing trends of these error metrics under different testing conditions, the robustness of the system can be evaluated in depth. Specifically, observing the rising curve of the error rate in a noisy environment with a gradually decreasing signal-to-noise ratio can quantify the degree of performance degradation of the system in noisy environments; comparing the recognition accuracy under different accents and speech rates can assess the system's adaptability to pronunciation diversity; and analyzing the difference in recognition performance between long and short texts can test the stability of the system when processing complex language structures.

[0052] The comprehensive evaluation report generated from these analyses not only includes quantitative data for various indicators but also identifies the system's strengths and weaknesses that urgently need improvement. For example, the report might show that the system achieves a 98% accuracy rate in recognizing standard pronunciation in quiet environments, but its performance degrades significantly in noisy environments, particularly in recognizing unstressed words. These insightful observations provide clear directions for subsequent system optimization, such as strengthening noise suppression algorithms, expanding accent training data, or adjusting the weight parameters of the language model.

[0053] In some embodiments, step 106, obtaining the test result based on the parsed text, may include: The second test result is obtained based on the time taken to obtain the parsed text.

[0054] The response time of a speech parsing system can be evaluated by statistically analyzing the time consumed in parsing the test speech to obtain the parsed text. As an example, and not a limitation, in practice, a high-precision timer can be used to record the time at three key points: the start of speech input, the completion of speech data processing, and the final text output. By calculating the time difference between each stage, the total response time of the system can be accurately obtained, and further subdivided into audio preprocessing time, acoustic model inference time, and language model decoding time. This subdivision capability allows for more precise identification of performance bottlenecks. To ensure the comprehensiveness of the evaluation, the tests should cover a variety of typical scenarios. The system's basic performance should be tested in a quiet environment, its robustness evaluated in a high-noise environment, its adaptability tested using speech at different speeds, and its stability verified through speech inputs of varying lengths. Through statistical analysis of a large number of samples, the average response time, response time variance, and maximum / minimum response time can be obtained, thus comprehensively evaluating the system's performance under different loads and conditions. The advantage of this evaluation method is that it enables automated and standardized performance testing, systematically analyzing the impact of various factors on response time by controlling variables. Test results can provide clear directions for system optimization, such as whether it is necessary to optimize the acoustic model structure, improve the decoding strategy, or enhance the noise suppression capability, thereby promoting the performance improvement of the speech parsing system in practical applications.

[0055] In some embodiments, step 106, obtaining the test result based on the parsed text, may include: By comparing the parsed text and the test text, a first test result is obtained; The second test result is obtained based on the time taken to obtain the parsed text.

[0056] In other examples, test results may include a comprehensive evaluation report on metrics such as accuracy, response time, error type, error frequency, robustness, etc.

[0057] According to an example of an embodiment of the present invention, after obtaining the test result based on the parsed text, the following may be included: Based on the test results, an optimization strategy for the speech parsing system is output.

[0058] For example, test results can provide optimization strategies for the speech parsing system, such as model tuning, data preprocessing improvements, increasing training data, and algorithm optimization. For instance, adjusting model parameters based on test results, such as the number of layers, the size of hidden units, or the regularization strength during training; optimizing data preprocessing processes, such as improving speech signal denoising, enhancement, or feature extraction methods; if the model performs poorly on specific types of input, it may be necessary to increase the amount of training data for those types; and optimizing existing algorithms to improve processing speed or accuracy, for example, by introducing more efficient acoustic or language models. This provides guidance for continuous system improvement, contributing to enhanced overall system performance and reliability.

[0059] According to another example of an embodiment of the present invention, after obtaining the test result based on the parsed text, the following may be included: The speech parsing system is optimized based on the test results.

[0060] This is merely illustrative; the speech parsing system can be directly optimized based on the test results, such as by fine-tuning the model, improving data preprocessing, increasing training data, and optimizing the algorithm.

[0061] According to another example of an embodiment of the present invention, after obtaining the test result based on the parsed text, the following may be included: The speech parsing system is optimized based on the test results until the test results meet the preset expectations.

[0062] For example, a continuous optimization feedback loop can be established, adjusting the speech parsing system based on the results of each test and retesting until a satisfactory performance level is achieved. This is merely an example, not a limitation; in practical implementation, a deep analysis of the test results is necessary first. When problems are identified, such as word error rates exceeding thresholds, significant performance degradation in specific scenarios (e.g., high-noise environments), or substandard response times, engineers can pinpoint the system's weaknesses. For instance, if test results show a high confusion rate for specific phoneme pairs, targeted training of the acoustic model may be necessary; if frequent errors are removed in fast speech scenarios, adjusting the decoder's hyperparameter settings may be required.

[0063] Based on these analytical results, optimization measures will be implemented in a targeted manner. This may include: updating the training data of the acoustic model, especially supplementing speech samples corresponding to weak links; adjusting the weight parameters of the language model to reduce specific types of insertion errors; optimizing the front-end signal processing algorithm to improve speech enhancement in noisy environments; or improving the decoding strategy to seek a better balance between accuracy and response time.

[0064] After each optimization, the system re-enters the testing process and is validated using the same benchmark. This iterative process continues until all key performance indicators—including but not limited to word error rate, sentence error rate, response time, and performance in specific scenarios—reach the preset expected levels. Through this rigorous optimization cycle, not only can the steady improvement of system performance be ensured, but performance imbalances caused by over-optimization can also be effectively avoided, ultimately delivering a robust speech parsing system that performs well in various application scenarios.

[0065] This invention also provides a testing device for a large-model-based speech parsing system, as described in the following embodiments. Since the principle by which this device solves the problem is similar to the testing method for a large-model-based speech parsing system, the implementation of this device can refer to the implementation of the testing method for a large-model-based speech parsing system; repeated details will not be elaborated further.

[0066] Figure 5 This is a schematic structural diagram of an example test apparatus for a large-model-based speech parsing system according to an embodiment of the present invention. Figure 5 As shown, the test apparatus for a large-model-based speech parsing system according to an embodiment of the present invention includes: The generation module 2 is used to generate test speech through a test model, wherein the test model is obtained by training a machine learning model based on the speech data and the corresponding text data, language data, speech rate data, accent data and noise data. Input module 4 is used to input the test speech into the speech parsing system to obtain the parsed text; and Module 6 is used to obtain test results based on the parsed text.

[0067] This invention also provides a computer device. Figure 6 This is a schematic structural diagram of an example of a computer device according to an embodiment of the present invention. Figure 6 As shown, the computer device 500 according to an embodiment of the present invention includes a memory 510, a processor 520, and a computer program 530 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 530, it implements the above-described test method for the speech parsing system based on a large model.

[0068] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described test method for a large-model-based speech parsing system.

[0069] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for testing a large-model-based speech parsing system.

[0070] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0071] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0072] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0073] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0074] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A testing method for a speech parsing system based on a large model, characterized in that, include: Test speech is generated through a test model, wherein the test model is obtained by training a machine learning model based on speech data and corresponding text data, language data, speech rate data, accent data and noise data. The test speech is input into the speech parsing system to obtain the parsed text; The test results are obtained based on the parsed text.

2. The method as described in claim 1, characterized in that, After obtaining the test results based on the parsed text, the process includes: Output an optimization strategy for the speech parsing system based on the test results; or Optimize the speech parsing system based on the test results; or The speech parsing system is optimized based on the test results until the test results meet the preset expectations.

3. The method as described in claim 1, characterized in that, Generate test speech using the test model, including: Based on the preset test text, the test model generates test speech; or Test speech is generated directly from the test model, and corresponding test text is generated based on the test speech; or Based on preset test text, language settings, speech rate settings, accent settings, and noise settings, the test model generates test speech.

4. The method as described in claim 3, characterized in that, The test results are obtained based on the parsed text, including: By comparing the parsed text and the test text, a first test result is obtained; and / or The second test result is obtained based on the time taken to obtain the parsed text.

5. The method as described in claim 3, characterized in that, The test model includes an input control layer, a text processing module, a conditional and control coding layer, an acoustic model, and a vocoder; Based on preset test text, language settings, speech rate settings, accent settings, and noise settings, the test model generates test speech, including: The input control layer receives the test text, language settings, speech rate settings, accent settings, and noise settings. Based on the test text, a standard phoneme sequence is generated by the text processing module; Based on the language setting, speech rate setting, accent setting, and noise setting, a corresponding conditional control embedding vector is generated through the conditional and control coding layer. Based on the standard sonic sequence and the conditional control embedding vector, a Mel spectrogram is generated through the acoustic model, wherein the acoustic model includes a text encoder, a variant adapter, and a Mel spectrogram generator. The standard phoneme sequence is encoded by the text encoder, and the duration, fundamental frequency, and energy characteristics of the elements in the standard phoneme sequence are adjusted according to the conditional control embedding vector to obtain an adapted phoneme sequence. Based on the adapted phoneme sequence, the Mel spectrogram is generated through the Mel spectrogram generator. The test speech is obtained through the vocoder based on the Mel spectrogram.

6. The method as described in claim 5, characterized in that, The test model also includes a post-processing module; based on preset test text, language settings, speech rate settings, accent settings, and noise settings, the test model generates test speech, and further includes: The post-processing module mixes noise with the test speech obtained from the vocoder to obtain a variant speech of the test speech.

7. The method as described in claim 1, characterized in that, Before generating test speech using the test model, the test task is divided into at least two test sub-tasks; as well as Each test subtask is allocated corresponding computing resources according to preset rules; and / or The test subtasks are executed in parallel.

8. A test apparatus for a speech parsing system based on a large model, characterized in that, include: The generation module is used to generate test speech through a test model, wherein the test model is obtained by training a machine learning model based on the speech data and the corresponding text data, language data, speech rate data, accent data and noise data. An input module is used to input the test speech into the speech parsing system to obtain parsed text; and The acquisition module is used to obtain test results based on the parsed text.

9. A computer 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 computer program, it implements the method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images