Teacher classroom speech recognition method and system based on hot word guidance, and readable storage medium
By constructing a hot word database and integrating it with acoustic features, the problem of low accuracy in recognizing professional terms and high-frequency words in teachers' classrooms by traditional speech recognition systems has been solved, achieving higher recognition accuracy and adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-06
- Publication Date
- 2026-04-10
AI Technical Summary
Traditional speech recognition systems have limited accuracy in recognizing subject-specific terminology and high-frequency teaching vocabulary in the classroom, and cannot effectively cope with the frequent appearance of new terms during the teaching process.
We construct a hot word library that is highly relevant to teaching scenarios. By preprocessing and noise suppression of teachers' classroom audio, we remove low-relevance words, convert them into hot word embedding vectors, and fuse them with acoustic features. These vectors are then input into a large language model for speech recognition.
It improves the recognition accuracy of professional terms and high-frequency words in teaching scenarios, enhances the model's adaptability to teaching content and recognition accuracy, and maintains a high recognition rate, especially under noise interference and accent changes.
Smart Images
Figure CN121838769A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of semantic recognition technology, and in particular to a method, system and readable storage medium for teacher classroom speech recognition based on hot word guidance. Background Technology
[0002] Teacher speech recognition technology is a core component of smart education systems. Its goal is to automatically convert teachers' lectures into text, supporting applications such as classroom content analysis, teaching assessment, and student review. Traditional speech recognition systems are typically trained on large-scale general corpora. While they perform well in certain scenarios, they lack the ability to adapt to the specific needs of teaching scenarios, especially in terms of limited accuracy when dealing with subject-specific terminology and high-frequency teaching vocabulary.
[0003] Currently, with the development of deep learning technology, some improved speech recognition solutions for specific domains have emerged. Domain-adaptive training-based speech recognition methods improve the ability to recognize domain-specific vocabulary by introducing domain-specific corpora for model fine-tuning. However, this method still suffers from insufficient adaptability to dynamically changing classroom content and cannot effectively cope with the frequent appearance of new terminology during teaching.
[0004] In view of this, this application proposes a new method for recognizing teacher classroom speech, which aims to improve the recognition accuracy of professional terms and high-frequency words in teaching scenarios. Summary of the Invention
[0005] The main purpose of this application is to provide a teacher classroom speech recognition method based on hot word guidance, which aims to solve the problem of how to improve the performance of teacher classroom speech recognition.
[0006] To achieve the above objectives, this application provides a teacher classroom speech recognition method based on hot word guidance, the method comprising: S10: Sort the words in the semantic dataset associated with the teacher's classroom audio according to their weights, and remove low-relevance words from the target words whose weight values are within the first preset proportion after sorting that do not match the reference words in the reference database, and use them as hot words. S20, the hot words are converted into hot word embedding vectors, and the attention weights corresponding to the hot word embedding vectors are calculated. The attention weights are then fused with the acoustic features in the teacher's classroom audio to obtain fused features. S30, the dataset composed of the various fused features is input into the large language model for speech recognition.
[0007] Optionally, before step S10, the method further includes: S40 extracts the audio track from the classroom teaching video, converts it to a standard audio format, and uses spectral subtraction for noise suppression. S50, perform speech activity detection on the denoised audio, retain the effective segments containing the teacher's speech, and remove silent and non-teaching speech segments to obtain the teacher's classroom audio.
[0008] Optionally, S10 specifically includes: S11, Calculate the weight value of the word in the semantic dataset. :
[0009] In the formula, Let be the frequency of word t in semantic dataset d. Let t be the number of documents containing word t, and N be the total number of words in the semantic dataset.
[0010] S12, Select words whose weight values are within the aforementioned preset ratio as target words; S13, determine the correlation between the target vocabulary and the reference vocabulary in the reference database, and remove low-correlation vocabulary with a correlation less than a preset threshold to obtain the hot words.
[0011] Optionally, before executing S20, the following steps are also included: S60, calculate the first weight value corresponding to the current hot word or the second weight value corresponding to the newly added words in the semantic dataset every preset period; S70, retain / remove the current hot words according to the first weight value, and update / not update the newly added words as hot words according to the second weight value.
[0012] Optionally, S20 includes: S21, convert the hot words in the hot word library into hot word embedding vectors:
[0013] in, Let i be the hot word embedding vector, i be the hot word index, and E be the embedding matrix; S22, Calculate the attention weights between acoustic features and hot word embedding vectors:
[0014] Where A is the attention weight matrix, (Acoustic feature projection). (Hot word feature projection) This is the vector dimension scaling factor; S23, the attention-weighted hot word features are fused with the acoustic features to obtain the fused features:
[0015] in, As a feature of fusion, Acoustic characteristics, (Hot word value projection) These are learnable fusion weight coefficients.
[0016] Optionally, in step S30, the loss function used by the large language model satisfies the following expression:
[0017] in, For CTC losses, For hot word cross-entropy loss, These are the weighting coefficients.
[0018] Optionally, after S30, the method further includes: The language recognition results of the large language model are verified by at least one of the following indicators: word error rate, hot word recognition accuracy, hot word recognition recall rate, and hot word F1 score.
[0019] In addition, to achieve the above objectives, this application also provides a computer system, the computer system comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, it implements the steps of the teacher classroom speech recognition method based on hot word guidance as described in any of the preceding claims.
[0020] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the teacher classroom speech recognition method based on hot word guidance as described in any of the preceding claims.
[0021] This application has at least the following beneficial effects: By constructing a hot word library that is highly relevant to teaching scenarios, the hot words in the hot word library are processed to obtain fusion features that highlight the features. These features are then input into a large language model for speech recognition, enabling the model to improve recognition accuracy by utilizing hot word information. Attached Figure Description
[0022] Figure 1 This is a flowchart illustrating the teacher classroom speech recognition method based on hot word guidance involved in the embodiments of this application; Figure 2 This is a schematic diagram of the hardware operating environment of the computer system involved in the embodiments of this application.
[0023] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0024] To better understand the above technical solutions, exemplary embodiments of this disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.
[0025] First Embodiment Reference Figure 1 This embodiment provides a teacher classroom speech recognition method based on hot word guidance, the method including the following steps: S10: Sort the words in the semantic dataset associated with the teacher's classroom audio according to their weights, and remove low-relevance words from the target words whose weight values are within the first preset proportion after sorting that do not match the reference words in the reference database, and use them as hot words. In this embodiment, a hot word database is first constructed. Words in the semantic dataset associated with teachers' classroom audio are sorted by weight, and low-relevance words that do not match the reference words in the reference database are removed from the target words whose weight values are within the first preset proportion after sorting. These low-relevance words are then used as hot words.
[0026] In some alternative implementations, the teacher's classroom audio is obtained by separating the audio track from the classroom teaching video dataset, converting it to a standard audio format, performing noise suppression processing, performing speech activity detection on the denoised audio, retaining the effective segments containing the teacher's speech, and removing silent and non-teaching speech segments to obtain the teacher's classroom audio.
[0027] In addition, the teacher classroom video dataset includes classroom teaching videos covering at least one complete teaching cycle. These videos must record clear audio signals from the teacher during instruction, and the video format must be a common digital video format. In some optional implementations, the videos must fully cover at least one teaching cycle (e.g., a standard 45-minute class period), ensuring that the teacher's audio signal is clearly identifiable and free from severe environmental noise interference. The video format must be a common digital video format such as MP4 or AVI, with a resolution of at least 1280×720, a frame rate of at least 25fps, an audio sampling rate of at least 16kHz, and either mono or stereo channels.
[0028] Furthermore, and optionally, a denoising algorithm based on spectral subtraction is used to eliminate background noise, and the calculation formula is as follows:
[0029] in, This is the spectrum of the denoised speech signal. The amplitude spectrum of the noisy speech signal. For the estimated noise spectrum, This is the noise reduction factor (set to 1). This is the original speech phase spectrum.
[0030] Specifically, the preprocessed speech segments are associated with the corresponding artificially transcribed text to form a semantic dataset of "speech-text" pairs, and the words in the semantic dataset are sorted according to their weights.
[0031] Further, and optionally, the TF-IDF algorithm is used to calculate word weights. Specifically, this includes: S11, Calculate the weight value of the word in the semantic dataset. :
[0032] In the formula, Let be the frequency of word t in semantic dataset d. Let t be the number of documents containing word t, and N be the total number of words in the semantic dataset.
[0033] S12, Select words whose weight values are within the aforementioned preset ratio as target words; S13, determine the correlation between the target vocabulary and the reference vocabulary in the reference database, and remove low-correlation vocabulary with a correlation less than a preset threshold to obtain the hot words.
[0034] In some optional implementations, the preset ratio is 15%.
[0035] In some alternative implementations, the reference database can be a teaching syllabus. Non-teaching-related vocabulary is removed, while subject-specific terms, core concepts, and frequently used teaching words are retained.
[0036] In some alternative implementations, hot words are divided into three levels: core hot words, important hot words, and general hot words, based on their importance in teaching.
[0037] In some alternative implementations, each hot word is given pinyin annotations, semantic explanations, and attribute information such as its frequency of occurrence in teaching content.
[0038] S20, the hot words are converted into hot word embedding vectors, and the attention weights corresponding to the hot word embedding vectors are calculated. The attention weights are then fused with the acoustic features in the teacher's classroom audio to obtain fused features. In this embodiment, after obtaining the hot words, the hot words are converted into high-dimensional hot word embedding vectors, and the attention weights corresponding to the hot word embedding vectors are calculated. The attention weights are then fused with the acoustic features in the teacher's classroom audio to obtain fused features.
[0039] Acoustic features refer to the low-dimensional, high-density vector representations containing speech temporal and content information extracted from preprocessed teacher classroom audio using a feature encoder based on the wav2vec 2.0 architecture. They form the acoustic foundation for end-to-end speech recognition and are fused with hot word features through an attention mechanism to achieve accurate hot word-guided recognition.
[0040] Further and optionally, this step may specifically include: S21, convert the hot words in the hot word library into hot word embedding vectors:
[0041] in, Let i be the hot word embedding vector, i be the hot word index, and E be the embedding matrix; S22, Calculate the attention weights between acoustic features and hot word embedding vectors:
[0042] Where A is the attention weight matrix, (Acoustic feature projection). (Hot word feature projection) This is the vector dimension scaling factor; S23, the attention-weighted hot word features are fused with the acoustic features to obtain the fused features:
[0043] in, As a feature of fusion, Acoustic characteristics, (Hot word value projection) These are learnable fusion weight coefficients.
[0044] It's worth noting that fusing hot word embedding vectors with acoustic features effectively enhances the model's ability to recognize keywords in teaching scenarios. Specifically, hot word embedding vectors carry semantic prior information from the teaching context, while acoustic features capture the physical characteristics of the speech signal. By weighted fusion using an attention mechanism, the model can dynamically focus on words highly relevant to the current teaching content during recognition, thus maintaining high recognition accuracy even under complex conditions such as noise interference, accent variations, or terminological ambiguity. Furthermore, the fused features not only preserve acoustic details but also incorporate semantic context, helping to improve the model's sensitivity to specialized terminology and high-frequency teaching expressions, ultimately achieving more accurate and robust classroom speech recognition results.
[0045] S30, the dataset composed of the various fused features is input into the large language model for speech recognition.
[0046] In this embodiment, after obtaining the fused features, the dataset composed of the fused features is input into the large language model for speech recognition.
[0047] In some alternative implementations, the architecture of the large language model is based on the wav2vec 2 model, which includes a feature encoder, a context network, and a quantizer; a hot word guidance module is introduced at its output.
[0048] In some optional implementations, the training process of the large language model employs a joint loss function for parameter optimization. Further, and optionally, the loss function is calculated as follows:
[0049] in, CTC loss (used for speech-text alignment). The hot word cross-entropy loss is used to optimize hot word recognition. This is the weighting coefficient (usually ranging from 0.5 to 0.8).
[0050] In the technical solution provided in this embodiment, a hot word library that is strongly related to the teaching scenario is constructed. After the hot words in the hot word library are processed to obtain feature-highlighting fusion features, they are input into a large language model for speech recognition, so that the model can use hot word information to improve recognition accuracy.
[0051] Second Embodiment Based on the above embodiments, this embodiment provides a method for updating a hot word database. The updating of the hot word database satisfies the following rules: S60, calculate the first weight value corresponding to the current hot word or the second weight value corresponding to the newly added words in the semantic dataset every preset period; S70, retain / remove the current hot words according to the first weight value, and update / not update the newly added words as hot words according to the second weight value.
[0052] Furthermore, the updated formula is:
[0053] in, This is the updated collection of trending words. This is the original collection of hot words. To add a new vocabulary set that meets the threshold, This is a set of outdated words to be removed.
[0054] The above steps are used to construct a teacher classroom hot word database that includes a hierarchical structure and dynamic updating capabilities.
[0055] Third Embodiment Based on the above embodiments, this embodiment provides a method for evaluating the performance of a large language model after training. In this embodiment, at least one of the following metrics—word error rate, hot word recognition accuracy, hot word recognition recall, and hot word F1 score—is used to verify the language recognition results of the large language model.
[0056] Specifically, the Word Error Rate (WER) is calculated using the following formula:
[0057] Where S is the number of substitutions, D is the number of deletions, I is the number of insertions, and N is the total number of words in the reference text; Hot word recognition accuracy The calculation formula is:
[0058] Hot word recognition recall rate The calculation formula is:
[0059] The F1 score of hot words is calculated using the following formula:
[0060] Fourth embodiment Based on the above embodiments, in order to further improve the model's ability to perceive the differences in importance of hot words in teaching scenarios, this embodiment proposes a multi-granularity hot word guidance mechanism, which specifically includes the following steps: S80 constructs a hierarchical hot word library, dividing hot words into three levels: "core hot words", "important hot words" and "general hot words", and assigning different attention guidance intensity coefficients to each level; S81 introduces a hierarchical perception bias term during the attention weight calculation process, enabling the model to apply different levels of attention to hot words at different levels when fusing acoustic features. S82 employs a hierarchical cross-entropy loss function, imposing a higher penalty on core hot word recognition errors during training to enhance the model's ability to recognize key terms.
[0061] Furthermore, the formula for calculating the hierarchical attention weights is as follows:
[0062] Where B is the hierarchical bias matrix, B_k = λ_k·I, λ_k is the intensity coefficient of the hot words in the k-th layer, decreasing in the order of "core hot words > important hot words > general hot words", and I is the identity matrix.
[0063] This embodiment introduces a hierarchical hot word structure and a differentiated attention mechanism, enabling the model to more accurately identify key concepts and terms in the teaching process, thereby further improving the professionalism and scene adaptability of speech recognition.
[0064] Verification of Examples This embodiment compares and verifies the performance of the model built based on the proposed method with that of the traditional baseline model. The experimental results are shown in Table 1 below: Table 1. Comparative Experiment Table
[0065] As can be seen, the experimental results confirm the effectiveness of the dynamic hot word database construction technology and hot word-speech collaborative training mechanism proposed in this application, which significantly improves the overall performance and professionalism of teachers' classroom speech recognition.
[0066] As one implementation scheme, Figure 2 This is a schematic diagram of the hardware operating environment of the computer system involved in the embodiments of this application.
[0067] like Figure 2As shown, the computer system may include: a processor 1001, such as a CPU; a memory 1005; a user interface 1003; a network interface 1004; and a communication bus 1002. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.
[0068] Those skilled in the art will understand that Figure 2 The computer system architecture shown does not constitute a limitation on the computer system and may include more or fewer components than shown, or combine certain components, or have different component arrangements.
[0069] like Figure 2 As shown, the memory 1005, as a storage medium, may include an operating system, a network communication module, a user interface module, and computer programs. The operating system is a program that manages and controls the hardware and software resources of the computer system, as well as the operation of the computer programs and other software or programs.
[0070] exist Figure 2 In the computer system shown, the user interface 1003 is mainly used to connect to the terminal and communicate with the terminal; the network interface 1004 is mainly used to communicate with the backend server; and the processor 1001 can be used to call the computer program stored in the memory 1005.
[0071] In this embodiment, the computer system includes: a memory 1005, a processor 1001, and a computer program stored in the memory and executable on the processor, wherein: When processor 1001 calls a computer program stored in memory 1005, it performs the following operations: S10: Sort the words in the semantic dataset associated with the teacher's classroom audio according to their weights, and remove low-relevance words from the target words whose weight values are within the first preset proportion after sorting that do not match the reference words in the reference database, and use them as hot words. S20, the hot words are converted into hot word embedding vectors, and the attention weights corresponding to the hot word embedding vectors are calculated. The attention weights are then fused with the acoustic features in the teacher's classroom audio to obtain fused features. S30, the dataset composed of the various fused features is input into the large language model for speech recognition.
[0072] When processor 1001 calls a computer program stored in memory 1005, it performs the following operations: S40 extracts the audio track from the classroom teaching video, converts it to a standard audio format, and uses spectral subtraction for noise suppression. S50, perform speech activity detection on the denoised audio, retain the effective segments containing the teacher's speech, and remove silent and non-teaching speech segments to obtain the teacher's classroom audio.
[0073] When processor 1001 calls a computer program stored in memory 1005, it performs the following operations: S11, Calculate the weight value of the word in the semantic dataset. :
[0074] In the formula, Let be the frequency of word t in semantic dataset d. Let t be the number of documents containing word t, and N be the total number of words in the semantic dataset.
[0075] S12, Select words whose weight values are within the aforementioned preset ratio as target words; S13, determine the correlation between the target vocabulary and the reference vocabulary in the reference database, and remove low-correlation vocabulary with a correlation less than a preset threshold to obtain the hot words.
[0076] When processor 1001 calls a computer program stored in memory 1005, it performs the following operations: S60, calculate the first weight value corresponding to the current hot word or the second weight value corresponding to the newly added words in the semantic dataset every preset period; S70, retain / remove the current hot words according to the first weight value, and update / not update the newly added words as hot words according to the second weight value.
[0077] When processor 1001 calls a computer program stored in memory 1005, it performs the following operations: S21, convert the hot words in the hot word library into hot word embedding vectors:
[0078] in, Let i be the hot word embedding vector, i be the hot word index, and E be the embedding matrix; S22, Calculate the attention weights between acoustic features and hot word embedding vectors:
[0079] Where A is the attention weight matrix, (Acoustic feature projection). (Hot word feature projection) This is the vector dimension scaling factor; S23, the attention-weighted hot word features are fused with the acoustic features to obtain the fused features:
[0080] in, As a feature of fusion, Acoustic characteristics, (Hot word value projection) These are learnable fusion weight coefficients.
[0081] When processor 1001 calls a computer program stored in memory 1005, it performs the following operations: The loss function used in the large language model satisfies the following expression:
[0082] in, For CTC losses, For hot word cross-entropy loss, These are the weighting coefficients.
[0083] When processor 1001 calls a computer program stored in memory 1005, it performs the following operations: The language recognition results of the large language model are verified by at least one of the following indicators: word error rate, hot word recognition accuracy, hot word recognition recall rate, and hot word F1 score.
[0084] Furthermore, those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in a computer system to implement the process steps of the embodiments of the above methods.
[0085] Therefore, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the various steps of the hot word-guided teacher classroom speech recognition method described in the above embodiments.
[0086] The computer-readable storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.
[0087] It should be noted that, since the storage medium provided in the embodiments of this application is the storage medium used to implement the methods of the embodiments of this application, those skilled in the art can understand the specific structure and variations of the storage medium based on the methods described in the embodiments of this application, and therefore will not be repeated here. All storage media used in the methods of the embodiments of this application fall within the scope of protection of this application.
[0088] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application 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.
[0089] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0090] 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.
[0091] 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.
[0092] It should be noted that any reference signs placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. This application can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.
[0093] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0094] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A method for teacher classroom speech recognition based on hotword bootstrapping, characterized in that, The method comprises the following steps: S10, sorting the vocabulary in the semantic data set associated with the teacher classroom audio according to the weight, and removing low correlation vocabulary that does not match the reference vocabulary in the reference database from the target vocabulary in the front preset proportion of the weight value after sorting as a hot word; S20, converting the hot word into a hot word embedding vector, calculating the attention weight corresponding to the hot word embedding vector, fusing the attention weight with the acoustic feature in the teacher classroom audio to obtain a fusion feature; S30, inputting the data set composed of each fusion feature into a large language model for speech recognition.
2. The method of claim 1, wherein, Before S10, it further comprises: S40, separating the audio track from the classroom teaching video, converting it into a standard audio format and using spectral subtraction for noise suppression processing; S50, performing voice activity detection on the de-noised audio, retaining the effective segment containing the teacher's voice, and removing the silent and non-teaching voice segment to obtain the teacher classroom audio.
3. The method of claim 1, wherein, S10 specifically comprises: S11, calculating a weight value of the vocabulary in the semantic dataset : ； wherein, is the frequency of occurrence of the word t in the semantic dataset d, is the number of documents containing the word t, N is the total number of words in the semantic dataset; S12, selecting the vocabulary in the front preset proportion of the weight value as the target vocabulary; S13, determining the correlation degree between the target vocabulary and the reference vocabulary in the reference database, removing the low correlation vocabulary with a correlation degree less than a preset threshold to obtain the hot word.
4. The method of claim 1, wherein, Before performing S20, it further comprises: S60, calculating the first weight value corresponding to the current hot word or the second weight value corresponding to the newly added vocabulary in the semantic data set every preset period; S70, retaining / removing the current hot word according to the first weight value, and updating / not updating the newly added vocabulary as a hot word according to the second weight value.
5. The method of claim 1, wherein, S20 comprises: S21, converting the hot word in the hot word library into a hot word embedding vector: ； wherein, is a hotword embedding vector, i is a hotword index, and E is an embedding matrix. S22, calculating the attention weight of the acoustic feature and the hot word embedding vector: ； wherein A is an attention weight matrix, (acoustic feature projection), (hotword feature projection), is a vector dimensionality scaling factor; S23, fusing the hot word feature after attention weighting and the acoustic feature to obtain a fusion feature: ； wherein, is a fusion feature, is an acoustic feature, (heat word value projection), is a learnable fusion weight coefficient.
6. The method of claim 1, wherein, In S30, the loss function used by the large language model satisfies the following expression: ； wherein, is a CTC loss, is a hotword cross-entropy loss, is a weight coefficient.
7. The method of claim 1, wherein, After S30, it further comprises: Verifying the language recognition result of the large language model by at least one of the following indicators: word error rate, hot word recognition accuracy, hot word recognition recall rate and hot word F1 value.
8. A computer system, characterized by The computer system comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and the computer program implements the steps of the hot word guided teacher classroom speech recognition method according to any one of claims 1 to 7 when executed by the processor.
9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program implements the steps of the hot word guided teacher classroom speech recognition method according to any one of claims 1 to 7 when executed by the processor.