Hot word extraction method and device for speech recognition, storage medium and product

Abstract

One or more embodiments of the specification provide a hot word extraction method, device, storage medium and product for speech recognition. The method comprises: obtaining a first data set, the first data set containing documents related to a target field; performing speech synthesis based on the original text of the documents contained in the first data set to generate simulated speech data; and performing recognition processing on the simulated speech data to obtain recognized text corresponding to the documents; comparing the original text of the documents with the corresponding recognized text, and counting the error words contained in the original text and the corresponding error frequencies to obtain an error word set; selecting a first candidate hot word set from the error word set; wherein the error frequency of the words selected into the first candidate hot word set is higher than the error frequency of the words not selected; and extracting a hot word related to the target field from the first candidate hot word set, the hot word reflecting the high-frequency words prone to errors of the speech recognition system in the target field.

Description

TECHNICAL FIELD

[0001] One or more embodiments of the present specification relate to the technical field of computer software, and particularly relate to a hot word extraction method for speech recognition, an electronic device, a computer readable storage medium, and a computer program product. BACKGROUND

[0002] In the practical application of a speech recognition system, for different application scenarios, it is often necessary to improve the recognition accuracy of domain-specific words. Since a general speech recognition model is often trained based on non-specific domain data, for some words that are common in a certain domain but do not often appear in general corpus, recognition errors are prone to occur.

[0003] To solve such problems, existing technologies usually introduce "hot words" to optimize recognition performance. Hot words generally refer to words that have high importance in a specific context, appear frequently, and may cause understanding bias if recognized incorrectly. By explicitly specifying such words as priority words for the recognition model, the recognition accuracy of related words can be improved.

[0004] However, existing hot word acquisition methods often rely on manual setting or expert experience. These methods are inefficient, prone to omissions, and difficult to cover changes caused by vocabulary evolution and domain expansion in a timely manner. SUMMARY

[0005] Therefore, one or more embodiments of the present specification provide a hot word extraction method for speech recognition, an electronic device, a computer readable storage medium, and a computer program product.

[0006] To achieve the above-mentioned purpose, one or more embodiments of the present specification provide technical solutions as follows:

[0007] According to a first aspect of one or more embodiments of the present specification, a hot word extraction method for speech recognition is provided, comprising:

[0008] obtaining a first data set containing documents related to a target domain;

[0009] performing speech synthesis based on the original text of the documents contained in the first data set to generate simulated speech data; and performing recognition processing on the simulated speech data to obtain the recognition text corresponding to the documents;

[0010] comparing the original text of the documents with the corresponding recognition text, and counting the error words contained in the original text and the corresponding error frequency to obtain an error word set;

[0011] A first candidate hot word set is selected from the set of incorrect words; wherein the error frequency of words selected into the first candidate hot word set is higher than that of words not selected.

[0012] Extract hot words related to the target domain from the first set of candidate hot words.

[0013] According to a second aspect of the embodiments of this specification, an electronic device is provided, comprising:

[0014] processor;

[0015] Memory used to store processor-executable instructions;

[0016] Wherein, when the processor executes the executable instructions, it is used to implement the method described in the first aspect.

[0017] According to a third aspect of the embodiments of this specification, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in the first aspect.

[0018] According to a fourth aspect of the embodiments of this specification, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0019] The technical solutions provided in the embodiments of this specification may include the following beneficial effects:

[0020] In this embodiment, by synthesizing speech from the original text of documents contained in the first dataset targeting the target domain, and then performing recognition processing on the synthesized simulated speech data, followed by comparison and analysis between the recognized text and the original text, the recognition process of a speech recognition system in a real-world application scenario can be effectively simulated, truly reflecting the recognition capability of the speech recognition system in the target domain. Based on the error word set obtained from the comparison results, high-frequency words prone to errors in the speech recognition system can be accurately identified. The resulting first hot word candidate set has a clear basis for recognition difficulty and can more specifically reflect the recognition shortcomings of the speech recognition system in the target domain. Furthermore, the hot words extracted from the first hot word candidate set have stronger applicability and representativeness, which helps to improve the subsequent speech recognition effect.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this specification. Attached Figure Description

[0022] Figure 1 This is a flowchart of a hot word extraction method for speech recognition provided in an exemplary embodiment.

[0023] Figure 2 This is a schematic diagram of a hot word extraction process provided in an exemplary embodiment.

[0024] Figure 3 This is a schematic diagram of another hot word extraction process provided in an exemplary embodiment.

[0025] Figure 4 This is a schematic diagram of another hot word extraction process provided in an exemplary embodiment.

[0026] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an exemplary embodiment. Detailed Implementation

[0027] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0028] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0029] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this manual are all information and data authorized by the user or fully authorized by all parties. The collection, use and processing of related data shall comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals shall be provided for users to choose to authorize or refuse.

[0030] Based on the issues in the relevant technologies, please refer to Figure 1 as well as Figure 2 This specification provides a method for hot word extraction in speech recognition. This method can be executed by an electronic device, including but not limited to physical servers, virtual servers, smartphones / mobile phones, tablet computers, personal digital assistants (PDAs), laptops, desktop computers, wearable devices, or any other type of device. The hot word extraction method for speech recognition includes:

[0031] In S101, obtain a first data set, which contains documents related to the target field.

[0032] Exemplarily, the first data set contains core text data within the target field, such as industry reports, technical documents, product descriptions, customer feedback records, etc. It is necessary to ensure that the data covers the core concepts, terms, frequent problems, etc. of the target field. The first data set can be constructed by means of information scraping, manual compilation, or provided by domain knowledge experts, and can be updated regularly according to different application scenarios. Thus, it is ensured that the subsequently extracted hot words are strongly relevant to the context of the target field, improving the fit between the hot words and the actual recognition scenarios, and reducing the interference of irrelevant and redundant words on the recognition performance.

[0033] In S102, perform speech synthesis based on the original text of the documents contained in the first data set to generate simulated speech data; and, perform recognition processing on the simulated speech data to obtain the recognition text corresponding to the documents.

[0034] Exemplarily, speech synthesis can be based on TTS (Text-to-Speech) technology, which is a technology that converts written text into speech output. TTS technology allows computers, smart devices or applications to "read out" the text content and is widely used in fields such as voice assistants, navigation systems, and e-book reading. Speech recognition can use an ASR (Automatic Speech Recognition) model to convert speech signals into corresponding text for application scenarios such as speech input, speech search, and speech control.

[0035] This step simulates the real speech recognition process through the "original text → simulated speech → recognition text" process, accurately exposing the recognition weaknesses of the current speech recognition system for specific vocabulary.

[0036] In S103, compare the original text of the documents with the corresponding recognition text, count the words in the original text that are misrecognized and the corresponding error frequencies, and obtain a set of error words.

[0037] This step compares the original text with the recognition text at the word level or character level to identify each misrecognized word in the original text, record the number of errors and the number of occurrences for each misrecognized word, and thus calculate the error frequency, where error frequency = number of errors / total number of occurrences. Optionally, a stop word list can be introduced to exclude meaningless high-frequency words in the set of error words, such as "de" (的), "shi" (是), "le" (了), etc. This step quantifies the weak links of the speech recognition system in the target field vocabulary through the comparison of the original text and the recognition text, providing a basis for subsequent hot word extraction.

[0038] In S104, a first candidate hot word set is selected from the set of incorrect words; among them, the error frequency of words selected into the first candidate hot word set is higher than that of words not selected.

[0039] For example, the screening strategy may include, but is not limited to: (1) screening words with an error frequency higher than a preset threshold from the set of error words; (2) sorting the words by error frequency and taking the top N words, where N can be set according to the actual application scenario. This ensures that the extracted candidate hot words are concentrated in words that are weak in the speech recognition system but have high semantic importance in the target domain, thereby improving the "blind spot filling" effect of the subsequent speech recognition system.

[0040] In S105, hot words related to the target domain are extracted from the first set of candidate hot words.

[0041] For example, at least a portion of the first set of candidate hot words can be identified as hot words related to the target domain.

[0042] In this embodiment, by synthesizing speech from the original text of documents contained in the first dataset targeting the target domain, and then using a speech recognition system to identify the synthesized simulated speech data, the identified text is compared and analyzed with the original text. This allows for a realistic reproduction of the speech recognition system's performance in the application scenario, and the identification of words with errors and their frequency can be statistically analyzed. The first hot word candidate set extracted in this way reflects the high-frequency words that the speech recognition system is prone to errors in the target domain. Therefore, the hot words extracted from the first hot word candidate set have actual recognition difficulty and are more targeted and representative.

[0043] In some embodiments, the electronic device further performs word segmentation on the original text of the documents contained in the first dataset to obtain an original vocabulary set; then, it evaluates the importance of each original vocabulary in the original vocabulary set to obtain an evaluation score for each original vocabulary. By segmenting the original text in the first dataset and evaluating the importance of each original vocabulary, core vocabulary with high information content and semantic value in the domain can be mined, providing a more accurate foundation for subsequent hot word selection. This processing method improves the ability to identify key information in a specific domain and helps to improve the accuracy and adaptability of the overall speech recognition system in professional contexts.

[0044] In one possible implementation, the electronic device can count the number of times each original word in the original vocabulary set appears in its respective document, as well as the total number of original words in the document containing each original word. Based on the ratio between the number of times each original word appears in its document and the total number of original words in the document containing each original word, the frequency of occurrence of that original word in the document can be calculated. Then, the electronic device can evaluate the importance of the original words based on their frequency of occurrence (TF, term frequency) in the document, thus obtaining an evaluation score for each original word. By counting the frequency of occurrence of words in their respective documents and evaluating their importance accordingly, the semantic weight of words in the document can be effectively reflected, enabling more accurate identification of words that significantly contribute to semantic expression and improving the accuracy of hot word extraction.

[0045] In another possible implementation, please refer to Figure 3 The electronic device acquires a pre-constructed second dataset, which includes documents related to multiple domains, such as internet text (e.g., news, encyclopedias, forums), general books, and publicly available text related to general domains, but is not limited to these. For each original word in the original vocabulary set, the electronic device can count the frequency of each original word in its respective documents. Furthermore, the electronic device can determine the number of documents in the second dataset containing that original word, and determine the inverse document frequency (IDF) of the original word based on the total number of documents in the second dataset and the number of documents containing that original word. The IDF is used to measure the rarity of the original word in the entire second dataset; for example, it can be the logarithm of the ratio of the total number of documents to the number of documents containing that original word. Finally, the electronic device determines the evaluation score of the original word based on the product of the frequency of each original word in its respective documents and the IDF of the original word.

[0046] When calculating TF-IDF (Term Frequency-Inverse Document Frequency) scores, using a separate second dataset (rather than the first dataset itself) to calculate the inverse document frequency (IDF) allows for a more accurate differentiation of the "uniqueness" of words within a specific domain. This embodiment calculates the inverse document frequency of words from the original first dataset using a second dataset covering multiple general domains, and combines this with the term frequency data from the first dataset to calculate the TF-IDF score. This helps to more accurately measure the "specificity" of words within the target domain. This method effectively avoids overestimating the importance of general words due to their frequent occurrence within a specific domain, thereby improving the recognizability and value of the selected hot words in practical applications.

[0047] For example, please refer to Figure 3After determining the evaluation scores of each original word in the original vocabulary set, the electronic device can select a second candidate hot word set from the original vocabulary set. The evaluation scores of the original words selected into the second candidate hot word set are higher than those of the original words that were not selected. For example, the selection strategy may include, but is not limited to: (1) selecting original words with evaluation scores higher than a preset threshold; (2) sorting the evaluation scores from smallest to largest and taking the N largest ones, where N can be set according to the actual application scenario. Selecting words with higher evaluation scores from the original vocabulary set as the second candidate hot word set can prioritize retaining words with high semantic value and strong domain representativeness. This method can further enhance the semantic coverage of the hot word set, optimize the interpretability of the hot words and their adaptation performance in the speech recognition system.

[0048] Finally, the electronic device can extract hot words related to the target domain from the first and second candidate hot word sets. For example, please refer to... Figure 3 The hot words related to the target domain include the intersection of the first and second candidate hot word sets. By extracting the intersection of the first candidate hot word set (obtained from misidentification statistics) and the second candidate hot word set (obtained from semantic importance assessment), both the difficulty of speech recognition and semantic representativeness can be comprehensively considered, thereby refining more representative hot words for the target domain. This method effectively enhances the practicality and reliability of hot words and improves the performance of the speech recognition system in real-world application scenarios.

[0049] In some embodiments, considering that sentiment may affect speech expression features (such as intonation, speech rate, stress, etc.) and the recognition accuracy of the speech recognition system, especially in actual interaction scenarios, the recognition results of the same text under different emotional states may differ. After acquiring the first dataset, the electronic device can synthesize speech on the original text of the documents contained in the first dataset according to a preset sentiment, generate simulated speech data with sentiment, and then perform recognition processing on the simulated speech data with sentiment to obtain the recognized text corresponding to the document. Next, the original text of the document is compared with the corresponding recognized text, and the words contained in the original text that are incorrectly recognized and their corresponding error frequencies are counted to obtain a set of incorrect words. A first candidate hot word set is selected from the set of incorrect words. Among them, the error frequency of words selected into the first candidate hot word set is higher than that of words not selected. Finally, hot words related to the target domain and sentiment are extracted from the first candidate hot word set. For example, the intersection between the first candidate hot word set and the aforementioned second candidate hot word set is determined as hot words related to the target domain and sentiment.

[0050] These hot words effectively reflect the high-risk vocabulary of a speech recognition system under specific emotional tendencies. They not only possess semantic representativeness but also cover sensitive or high-frequency words more prone to errors in emotional expression, resulting in stronger recognition and correction value and context matching capabilities. This embodiment, by introducing an emotional tendency dimension into the hot word extraction process, enhances the speech recognition system's adaptability to emotional changes in real-world interaction scenarios, improves the hot word hit rate and recognition accuracy of the recognition engine under specific intonation and emotional environments, thereby effectively reducing recognition errors, improving speech understanding accuracy, and further enhancing user experience and system response accuracy.

[0051] For example, a speech recognition system is used to automatically identify the content of conversations between users and customer service representatives. When users express similar content under different emotions (such as anger, anxiety, and satisfaction), their speech performance varies significantly, and the system's recognition accuracy also changes accordingly. The first dataset includes a large number of call transcripts from customer service centers (such as complaint dialogues, consultation dialogues, and praise feedback). Taking "anger" as an example, a speech synthesis model with an angry tone is used for text-to-speech synthesis to generate simulated angry speech; this speech is then fed into the speech recognition system to generate recognized text; compared with the original text, incorrect words are identified, such as "compensation," "delay," and "cancellation," which are high-frequency errors. The system statistically analyzes the words with high error frequencies under the "anger" emotion to form an "anger candidate hot word set"; and combines this with the TF-IDF method to further filter out "key hot words under the anger" emotion. For example, high-frequency words frequently misidentified under the "anger" emotion might include: "complaint," "compensation," "service," "attitude," "waiting too long," "delay," and "cancellation." The system's recognition engine can mark these hot words as key recognition words in the "anger" scenario, thereby improving recognition accuracy. For example, when a user says "I want to complain" in anger, it might have previously been misinterpreted as "I want to refactor," but now, because "complaint" is a popular word for anger, the system prioritizes and enhances its recognition of it.

[0052] To further improve the recognition performance of speech recognition for key domain vocabulary, hot word sets under different emotional states can be set up according to actual needs. Users may use different words, tones, or expressions when expressing the same or similar intentions under different emotional states (such as anger, anxiety, calmness, and satisfaction). By constructing hot word sets corresponding to these emotional states in the above manner, the speech recognition system can more accurately recognize these expressions, improving overall recognition robustness.

[0053] In practical deployments, the system can dynamically switch or prioritize loading corresponding emotional hot words based on the emotional tone of the current speech segment. For example, when the current speech is identified as having a "negative emotion," the system can increase the recognition weight of words with a high probability of misrecognition, such as "complaint," "refund," and "unusable." Combining emotional tone-based customized hot word libraries helps reduce the error rate of speech recognition in key scenarios (such as customer complaints and emergency feedback), thereby reducing comprehension bias and improving the accuracy of system responses and user satisfaction. As the amount of data grows and application scenarios become more diverse, more emotional dimensions (such as annoyance, surprise, and tension) can be continuously introduced into the hot word set, further enhancing the speech recognition model's adaptability to complex human-computer interaction scenarios.

[0054] By setting up hot word sets under different emotional tendencies, the speech recognition system can accurately identify high-frequency or key words in specific contexts based on emotional states. This not only enhances the recognition system's adaptability to complex contexts but also significantly improves the practical value and intelligent interactive experience of speech recognition in the target domain.

[0055] In some embodiments, when users use the speech recognition function, their historical behavioral data (such as access records, search preferences, operation paths, click frequency, usage time periods, and frequently visited locations) can reflect their focus, interests, or scenario preferences. Introducing this behavioral information as auxiliary information into the hot word extraction process can more accurately uncover frequently misused or highly relevant words in specific user scenarios or interest preferences, thereby improving the customized recognition effect.

[0056] Electronic devices can acquire user behavior data of target users; user behavior data is used to describe the usage behavior characteristics of target users in content related to the target domain; for example, user behavior data includes but is not limited to: (1) documents that users have browsed / accessed; (2) keywords that users click / query / voice input.

[0057] Please see Figure 4The electronic device can filter documents related to the user behavior data from the first dataset to form a data subset. In one implementation, documents containing keywords indicated by the user behavior data can be filtered from the first dataset to form a data subset. In another implementation, all documents in the first dataset can be assigned the same initial weight, and then the initial weights of the documents in the first dataset can be updated according to the user behavior data. For example, the weight of documents frequently accessed by users increases; if keywords involved in user clicks / queries / voice input frequently appear in certain documents, the weight of that document increases. Then, the electronic device can select a portion of documents with higher weights (e.g., weights higher than a preset threshold, or weights higher than the initial weights) from the first dataset to construct a data subset related to the target user. User behavior data can be regarded as a long-term accumulation and true reflection of their personal needs and concerns. By analyzing users' historical behavior data, the electronic device can prioritize the selection of documents highly related to the content of the target user's interests from the first dataset, construct a personalized data subset, and thus customize a set of hot words that better suits its usage scenario.

[0058] Electronic devices can perform speech synthesis based on the original text of documents contained in a data subset to generate simulated speech data. Then, they can perform recognition processing on the simulated speech data to obtain the corresponding recognized text of the document. The original text of the document is compared with the corresponding recognized text to statistically analyze the words in the original text that were incorrectly recognized and their corresponding error frequencies, resulting in a set of incorrect words. A first candidate hot word set is selected from the set of incorrect words, where the error frequency of words selected for the first candidate hot word set is higher than that of words not selected. From the first candidate hot word set, hot words relevant to the target domain and targeted at the target user are extracted. For example, the intersection of the first candidate hot word set and the second candidate hot word set obtained in the above manner is determined as the hot words relevant to the target domain and targeted at the target user. For instance, when these hot words are added to the "recognition vocabulary" or language model of a speech recognition engine, the recognition error rate in that user scenario can be significantly reduced. Hot words are no longer just "domain-general hot words," but rather high-frequency, high-error words finely adapted to a specific type of user in a specific scenario, making the recognition results more intelligent and reliable.

[0059] For example, suppose user A is a primary care physician who mainly uses voice recognition on mobile devices to issue medical records or query case information. His recent behavioral data is as follows: access records are concentrated on documents such as "coronary heart disease", "electrocardiogram analysis", and "bradycardia"; search keywords are "ventricular premature beats", "sinus bradycardia", and "myocardial ischemia".

[0060] During the process of extracting hot keywords, the electronic device first updates the initial weights of the documents in the first dataset. The aforementioned behavioral data leads to an increase in the weight of cardiovascular-related documents, such as "Guidelines for Chronic Stable Angina" and "Case Assessment of Cardiac Arrhythmias." The documents selected for the data subset are highly focused on cardiovascular-related content.

[0061] Electronic devices perform speech synthesis, recognition, and comparison on these documents in the data subset, identifying frequently misidentified words such as "ventricular premature beats" (identified as "four moves") and "myocardial infarction" (identified as "Xingang"). These frequently misidentified and highly relevant words constitute the first candidate hot word set. The second candidate hot word set can be obtained through the above description method. The intersection of the first and second candidate hot word sets forms the final hot word set, such as "ventricular premature beats," "sinus rhythm," and "electrocardiogram waveform."

[0062] In this embodiment, the introduction of personalized hot words improves the recognition accuracy in actual use scenarios for target users, especially in fields with a high concentration of technical terms (such as medicine, law, and finance). Users experience fewer speech recognition errors during interaction, reducing the need for repetitive input and manual correction, and significantly improving the efficiency of voice operation.

[0063] The various technical features in the above embodiments can be combined arbitrarily, as long as there is no conflict or contradiction between the combinations of features. However, due to space limitations, they are not described one by one. Therefore, the arbitrary combination of various technical features in the above embodiments is also within the scope of this specification.

[0064] In some embodiments, this specification also provides an electronic device, including: a processor; and a memory for storing processor-executable instructions; wherein the processor implements the method described in any one of the above embodiments by executing the executable instructions.

[0065] Figure 5 This is a schematic structural diagram of a device provided in an exemplary embodiment. Please refer to... Figure 5 At the hardware level, the device includes a processor 502, an internal bus 504, a network interface 506, memory 508, and non-volatile memory 510, and may also include other hardware required for its functions. One or more embodiments of this specification can be implemented in software, for example, the processor 502 reads the corresponding computer program from the non-volatile memory 510 into memory 508 and then runs it. Of course, in addition to software implementation, one or more embodiments of this specification do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0066] In some embodiments, the hot word extraction device for speech recognition can be applied to, for example... Figure 5 The device shown implements the technical solution described in this specification. The hot word extraction device for speech recognition may include:

[0067] The dataset acquisition module is used to acquire the first dataset, which contains documents related to the target domain.

[0068] The text acquisition module is used to perform speech synthesis based on the original text of the documents contained in the first dataset to generate simulated speech data; and to perform recognition processing on the simulated speech data to obtain the recognized text corresponding to the document.

[0069] The text comparison module is used to compare the original text of the document with the corresponding recognized text, count the words in the original text that were incorrectly recognized and their corresponding error frequencies, and obtain a set of incorrect words.

[0070] The dataset acquisition module is also used to filter out the first candidate hot word set from the set of error words; wherein, the error frequency of words selected into the first candidate hot word set is higher than that of words not selected.

[0071] The hot word extraction module is used to extract hot words related to the target domain from the first set of candidate hot words.

[0072] In one implementation, an evaluation module is used to perform word segmentation on the original text of the documents contained in the first dataset to obtain an original vocabulary set; and to evaluate the importance of each original vocabulary in the original vocabulary set to obtain an evaluation score for each original vocabulary.

[0073] The dataset acquisition module is also used to filter out a second candidate hot word set from the original vocabulary set. The evaluation scores of the original words selected into the second candidate hot word set are higher than the evaluation scores of the original words that were not selected.

[0074] The hot word extraction module is specifically used to extract hot words related to the target domain from the first candidate hot word set and the second candidate hot word set.

[0075] In one implementation, the hot words related to the target domain include the intersection of the first set of candidate hot words and the second set of candidate hot words.

[0076] In one implementation, the evaluation module is specifically used to count the frequency of occurrence of each original word in the original vocabulary set in its respective document; based on the frequency of occurrence of each original word in the original vocabulary set in its respective document, the importance of the original words is evaluated to obtain the evaluation score of the original words.

[0077] In one implementation, the dataset acquisition module is also used to acquire a second dataset, which includes documents related to multiple domains.

[0078] The evaluation module is specifically used to determine the number of documents in the second dataset containing each original word in the original vocabulary set, and to determine the inverse document frequency of the original word based on the total number of documents in the second dataset and the number of documents containing the original word; and to determine the evaluation score of the original word based on the product of the occurrence frequency of each original word in its respective document and the inverse document frequency of the original word.

[0079] In one implementation, the text acquisition module is specifically used to perform speech synthesis on the original text of the documents contained in the first dataset according to a preset sentiment tendency, so as to generate simulated speech data with sentiment tendency.

[0080] The hot word extraction module is specifically used to extract hot words that are relevant to the target domain and target sentiment from the first set of candidate hot words.

[0081] In one implementation, a filtering module is used to acquire user behavior data of the target user; the user behavior data is used to describe the usage behavior characteristics of the target user in content related to the target domain; and documents related to the user behavior data are filtered from the first dataset to form a data subset.

[0082] The text acquisition module is specifically used to perform speech synthesis based on the original text of the documents contained in the data subset, generating simulated speech data.

[0083] The hot word extraction module is specifically used to extract hot words that are relevant to the target domain and targeted at the target users from the first set of candidate hot words.

[0084] The specific implementation process of the functions and roles of each module in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0085] Based on the same concept as the methods described above, this specification also provides an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein the processor performs the steps of the method as described in any of the above embodiments by executing the executable instructions.

[0086] Based on the same concept as the methods described above, this specification also provides a computer-readable storage medium having computer instructions stored thereon that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

[0087] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0088] Based on the same concept as the methods described above, this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

[0089] The above description is merely a preferred embodiment of one or more embodiments of this specification and is not intended to limit the scope of one or more embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the protection scope of one or more embodiments of this specification.

Claims

1. A method for hot word extraction in speech recognition, comprising: Obtain the first dataset, which contains documents related to the target domain; Speech synthesis is performed on the original text of the documents contained in the first dataset to generate simulated speech data; and the simulated speech data is processed for recognition to obtain the recognized text corresponding to the document. The original text of the document is compared with the corresponding recognized text, and the words in the original text that are incorrectly recognized and their corresponding error frequencies are counted to obtain a set of incorrect words. A first candidate hot word set is selected from the set of incorrect words; wherein the error frequency of words selected into the first candidate hot word set is higher than that of words not selected. Extract hot words related to the target domain from the first candidate hot word set; The step of generating simulated speech data by performing speech synthesis on the original text of the documents contained in the first dataset includes: performing speech synthesis on the original text of the documents contained in the first dataset according to a preset emotional tendency to generate simulated speech data with the emotional tendency. The step of extracting hot words related to the target domain from the first candidate hot word set includes: extracting hot words related to the target domain and targeting the sentiment from the first candidate hot word set to obtain a hot word set, so that the speech recognition system can enhance the recognition weight of the hot word set related to the sentiment of the speech to be recognized during the speech recognition process.

2. The method according to claim 1, further comprising: The original text of the documents contained in the first dataset is segmented to obtain the original vocabulary set; The important procedures of each original word in the original vocabulary set are evaluated to obtain an evaluation score for each original word; A second candidate hot word set is selected from the original word set. The evaluation scores of the original words selected into the second candidate hot word set are higher than the evaluation scores of the original words that were not selected. The step of extracting hot words related to the target domain from the first candidate hot word set includes: Extract hot words related to the target domain from the first candidate hot word set and the second candidate hot word set.

3. The method according to claim 2, wherein the hot words related to the target field include: The intersection of the first set of candidate hot words and the second set of candidate hot words.

4. The method according to claim 2, wherein evaluating the importance of each original word in the original vocabulary set to obtain an evaluation score for each original word includes: Calculate the frequency of occurrence of each original word in the original vocabulary set in its respective document; Based on the frequency of occurrence of each original word in the original vocabulary set in its respective document, the importance of the original words is evaluated, and an evaluation score is obtained for each original word.

5. The method according to claim 4, further comprising: Obtain a second dataset, which includes documents related to multiple domains; For each original word in the original vocabulary set, determine the number of documents in the second dataset that contain that original word, and determine the inverse document frequency of the original word based on the total number of documents in the second dataset and the number of documents that contain that original word. The assessment of the importance of each original word in the original vocabulary set based on its frequency of occurrence in its respective document, to obtain an assessment score for the original word, includes: The evaluation score of each original word is determined based on the product of the frequency of each original word in its document and the inverse document frequency of the original word.

6. The method according to claim 1, further comprising: Obtain user behavior data from the target users; The user behavior data is used to describe the usage behavior characteristics of the target user in content related to the target domain; Documents related to the user behavior data are selected from the first dataset to form a data subset; The step of performing speech synthesis based on the original text of the documents contained in the first dataset to generate simulated speech data includes: Speech synthesis is performed on the original text of the documents contained in the data subset to generate simulated speech data; The step of extracting hot words related to the target domain from the first candidate hot word set includes: From the first set of candidate hot words, extract hot words that are related to the target domain and targeted at the target user.

7. An electronic device, comprising: processor; A memory for storing processor-executable instructions; wherein the processor implements the steps of the method as described in any one of claims 1 to 6 by executing the executable instructions.

8. A computer-readable storage medium having stored thereon computer instructions that, when executed by a processor, implement the steps of the method as claimed in any one of claims 1 to 6.

9. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method as claimed in any one of claims 1 to 6.

Images