Popular media asset voice recognition method, device, equipment and medium

Abstract

The present disclosure relates to a popular media resource voice recognition method, device, equipment and medium, the method comprising: continuously collecting text corpus about popular media resources in a preset period to enter a popular word table; obtaining a voice to be recognized; identifying the voice to be recognized based on the popular word table and a basic language model to obtain a recognition result. The present disclosure continuously collects text corpus about popular media resources in a preset period to enter a popular word table, and then uses the popular word table to help the basic language model identify the popular media resources that may be contained in the voice to be recognized when the voice to be recognized is obtained for voice recognition, thereby improving the recognition accuracy of popular media resources in voice recognition and reducing the time and resources consumed for updating the language model.

Description

Technical Field

[0001] This disclosure relates to the field of speech recognition technology, specifically to a speech recognition method, apparatus, device, and medium for popular media assets. Background Technology

[0002] Popular music charts are common in music and audiobook industries, featuring songs and audiobooks that have recently been frequently featured. In products like smart speakers and smart TVs, it's very common for users to issue commands to play popular songs or audiobooks. In these scenarios, if the ASR (Automatic Speech Recognition) system doesn't update for an extended period, it won't be able to respond to user demands for real-time popular songs or audiobooks, leading to user dissatisfaction. Therefore, recognizing popular songs and audiobooks has real-world and practical value. A common solution to this problem is to frequently add the corpus of popular songs and audiobooks to the existing language model training data for iterative training.

[0003] However, because the trending lists are updated very frequently, the language model also needs to be frequently updated and tested using the above method. Furthermore, the language model needs to be fused with the dictionary and acoustic model to generate a decoding graph before it can participate in audio decoding, and these processes are very time-consuming. Moreover, due to the rapid update frequency of the trending lists, the amount of text data collected each time the language model is updated is limited. This limited data is negligible compared to the previous language model's data, potentially causing the newly trained model to be ineffective. Summary of the Invention

[0004] To solve the above-mentioned technical problems, or at least partially solve them, this disclosure provides a method, apparatus, device, and medium for speech recognition of popular media assets.

[0005] Firstly, this disclosure provides a method for speech recognition of popular media assets, including:

[0006] Continuously collect text corpora about popular media resources within a preset period and input them into a popular vocabulary list;

[0007] Acquire the speech to be recognized;

[0008] The speech to be identified is identified based on the popular vocabulary and the basic language model, and the recognition result is obtained.

[0009] Optionally, the step of continuously collecting text corpora about popular media resources and inputting them into a popular vocabulary list within a preset period includes:

[0010] Within a preset period, a web crawler program is used to continuously capture text corpora of media resources located on the trending list and input them into the trending word list.

[0011] Optionally, the step of identifying the target word based on the popular vocabulary and the basic language model...

[0012] The speech is processed to obtain recognition results, including:

[0013] 0. Use the aforementioned popular word list to enhance the hot word decoding of the basic language model, in order to increase...

[0014] The probability of the basic language model recognizing the text corpus contained in the popular word list is added;

[0015] The speech to be recognized is decoded using the basic language model, and the word sequence corresponding to the highest probability of the decoding result of the speech to be recognized is taken as the recognition result.

[0016] Optionally, the method further includes:

[0017] 5. In response to the end of the preset period, the multiple text corpora contained in the popular word list are...

[0018] Enter the training text and clear the aforementioned popular word list;

[0019] The basic language model is iteratively trained using the training text to obtain a newly trained language model.

[0020] Replace the base language model with the newly trained language model.

[0021] 0. Optionally, the step of inputting multiple text corpora contained in the popular vocabulary into the training text.

[0022] include:

[0023] Based on the number of times each popular media resource was screened, multiple text corpora contained in the popular term list were entered into the training text, so that the number of text corpora corresponding to popular media resources with more screenings would be greater in the training text.

[0024] 5. Optionally, the step of iteratively training the basic language model using the training text...

[0025] Practice yields a newly trained language model, including:

[0026] The text corpus in the training text is segmented according to the word segmentation lexicon to obtain the word segmentation result. Newly emerging word groups on the Internet are added to the word segmentation lexicon.

[0027] The word segmentation results are fused with the basic pre-trained word segments according to a preset ratio to obtain a training word segmentation set;

[0028] The basic language model is iteratively trained using the training word segmentation set to obtain a newly trained language model.

[0029] Optionally, before replacing the base language model with the newly trained language model, the method further includes:

[0030] The newly trained language model is pruned according to a preset pruning coefficient.

[0031] Secondly, this disclosure provides a popular media asset voice recognition device, including:

[0032] The collection module is used to continuously collect text corpora about popular media resources within a preset period and input popular word lists;

[0033] The acquisition module is used to acquire the speech to be recognized;

[0034] The recognition module is used to recognize the speech to be recognized based on the popular vocabulary and the basic language model, and obtain the recognition result.

[0035] Optionally, when the collection module continuously collects text corpus of popular media resources into the popular word list within a preset period, it is specifically used to continuously crawl the text corpus of media resources located on the popular list and input it into the popular word list using a crawler program within the preset period.

[0036] Optionally, when the recognition module recognizes the speech to be recognized based on the popular word list and the basic language model to obtain the recognition result, it is specifically used to enhance the basic language model with hot word decoding using the popular word list, so as to increase the probability of the basic language model recognizing the text corpus contained in the popular word list; decode the speech to be recognized using the basic language model, and take the word sequence corresponding to the decoding result of the speech to be recognized with the highest probability as the recognition result.

[0037] Optionally, the device further includes a training module, configured to, in response to the end of the preset period, input multiple text corpora contained in the popular word list into training text and clear the popular word list; use the training text to iteratively train the basic language model to obtain a newly trained language model; and replace the basic language model with the newly trained language model.

[0038] Optionally, when the training module inputs multiple text corpora contained in the popular term list into the training text, it is specifically used to input multiple text corpora contained in the popular term list into the training text according to the number of times each popular media resource has been screened, so that the number of text corpora corresponding to popular media resources with more screened times is greater in the training text.

[0039] Optionally, when the training module uses the training text to iteratively train the basic language model to obtain a newly trained language model, it specifically performs word segmentation processing on the text corpus in the training text according to the word segmentation vocabulary list to obtain word segmentation results, wherein newly emerging word groups on the Internet are added to the word segmentation vocabulary list; the word segmentation results are fused with the basic pre-trained word segmentation according to a preset ratio to obtain a training word segmentation set; and the training word segmentation set is used to iteratively train the basic language model to obtain a newly trained language model.

[0040] Optionally, before replacing the base language model with the newly trained language model, the training module is used to prune the newly trained language model according to a preset pruning coefficient.

[0041] Thirdly, this disclosure provides an electronic device, including:

[0042] Memory;

[0043] Processor; and

[0044] Computer programs;

[0045] The computer program is stored in the memory and configured to be executed by the processor to implement the method as described in the first aspect.

[0046] Fourthly, this disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0047] The technical solution provided in this disclosure has the following advantages compared with the prior art:

[0048] The method, apparatus, device, and medium for speech recognition of popular media resources disclosed herein continuously collect text data about popular media resources within a preset period and input it into a popular vocabulary list. Then, when obtaining the speech to be recognized for speech recognition, the popular vocabulary list is used to help the basic language model identify the popular media resources that may be contained in the speech to be recognized, thereby improving the recognition accuracy of popular media resources in speech recognition and reducing the time and resources consumed in updating the language model. Attached Figure Description

[0049] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0050] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart of a popular media asset speech recognition method provided in this embodiment of the disclosure;

[0052] Figure 2 A schematic diagram illustrating an application scenario provided by an embodiment of this disclosure;

[0053] Figure 3 A schematic diagram of the structure of the popular media asset voice recognition device provided in the embodiments of this disclosure;

[0054] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0055] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0056] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.

[0057] Because the trending topics list updates very frequently, the language model also needs frequent updates and testing. The language model must be fused with a dictionary and acoustic model to generate a decoding graph before it can participate in audio decoding, a process that is very time-consuming. Furthermore, due to the rapid update frequency of the trending topics list, the amount of text data collected each time the language model is updated is limited; this data may be negligible compared to the previous language model's data, potentially causing the newly trained model to be ineffective. To address this problem, this disclosure provides a method for speech recognition of trending media assets, which will be described below with specific embodiments.

[0058] Figure 1 This is a flowchart illustrating a popular media asset speech recognition method provided in an embodiment of this disclosure. The method can be executed by a popular media asset speech recognition device, which can be implemented in software and / or hardware. This device can be configured in an electronic device, such as a server or terminal, where the terminal specifically includes a mobile phone, computer, or tablet computer. Furthermore, this method can be applied to… Figure 2 The application scenario shown includes a server 201, a smart speaker 202, and a user 203. It is understood that the popular media asset voice recognition method provided in this embodiment can also be applied to other scenarios.

[0059] The following is combined with Figure 2 The application scenarios shown are for Figure 1 This section introduces popular media asset speech recognition methods, for example, Figure 2 Server 201 in the system can execute this method. The specific steps of this method are as follows:

[0060] S101. Continuously collect text corpora about popular media resources within a preset period and input them into a popular word list.

[0061] Because trending lists are updated very frequently, and the popular media resources on those lists are constantly changing, training a language model based on the frequency of trending list updates is not only time-consuming and resource-intensive, but the trained language model also has limited longevity. To ensure that the intelligent product using automatic speech recognition can still accurately identify recent popular media resources, server 201 continuously collects text corpora related to popular media resources and inputs them into a popular vocabulary list within a preset period. This popular vocabulary list, including text corpora of songs, artists, and audiobooks from recent trending lists, improves the accuracy of speech recognition of popular media resources. The preset period is typically several times the update time of popular media resources; for example, if the trending list is updated weekly, then the preset period could be four weeks. Server 201 will collect text corpora of songs, artists, and audiobooks from the weekly trending lists and input them into the popular vocabulary list within those four weeks.

[0062] S102. Obtain the speech to be recognized.

[0063] For example, when user 203 wants to play a song, he tells the smart speaker 202 the song he wants to play, which is a recent popular song. The server 201 then obtains the voice to be recognized from user 203 through the smart speaker 202.

[0064] S103. Based on the popular vocabulary and basic language model, the speech to be recognized is identified, and the recognition result is obtained.

[0065] Server 201 increases the probability of recognizing popular media resources in the speech to be recognized by increasing the probability that the text corpus entered into the popular vocabulary list is recognized and decoded by the basic language model. When user 203 wants to play a recently popular song, the speech to be recognized by user 203 to smart speaker 202 contains the name of the popular song. Server 201 has already entered the text corpus of all popular songs, audiobooks, etc. on the popular charts into the popular vocabulary list, which will include the name of the popular song. This allows the basic language model to recognize the name of the popular song when recognizing the speech to be recognized. The basic language model is usually a pre-trained model, including NNLM, word2vec, FastText, etc. These language models have undergone preliminary training and have a wide range of applications, but they are not strong in language recognition for a single domain. They may have difficulty recognizing recently emerging hot words and new words.

[0066] This embodiment of the disclosure continuously collects text data about popular media resources within a preset period and inputs it into a popular word list. Then, when acquiring the speech to be recognized for speech recognition, the popular word list is used to help the basic language model identify the popular media resources that may be contained in the speech to be recognized, thereby improving the recognition accuracy of popular media resources in speech recognition and reducing the time and resources consumed in updating the language model.

[0067] Based on the above embodiments, the text corpus of popular media resources is continuously collected and entered into the popular word list within a preset period, including: within a preset period, using a web crawler program to continuously crawl the text corpus of media resources located on the popular list and enter it into the popular word list.

[0068] Because there are many types and categories of popular charts, server 201 can use a web crawler to collect text data of songs, artists and audiobooks from recent popular charts and enter this text data into a popular word list, which saves more time and cost compared to manually collecting this text.

[0069] This disclosure uses a web crawler to continuously collect text corpora from media resources on popular lists, thereby saving time and costs.

[0070] Based on the above embodiments, the speech to be identified is identified based on a popular vocabulary list and a basic language model to obtain the recognition result, including: using a popular vocabulary list to enhance the basic language model with hot word decoding to increase the probability of the basic language model recognizing the text corpus contained in the popular vocabulary list; using the basic language model to decode the speech to be identified, and taking the word sequence corresponding to the highest probability of the decoding result of the speech to be identified as the recognition result.

[0071] In speech recognition, a language model decodes each frame of the speech signal to be recognized. After decoding the last frame, it selects the active node with the highest cumulative probability as the optimal node and obtains its corresponding word sequence. Therefore, to increase the recognition accuracy of text corpora in the popular vocabulary list, it is necessary to enhance the basic language model with hot word decoding using the popular vocabulary list to increase the probability of the basic language model recognizing text corpora contained in the popular vocabulary list. Specifically, server 201 enhances the historical paths of active nodes based on the set of hot word corpora in the popular vocabulary list, increasing the cumulative historical path probability of the paths containing hot words, so that the hot word paths can be selected during the decoding process. This ensures that after the basic language model decodes the speech to be recognized, the word sequence corresponding to the decoding result with the highest probability can be the text corpora collected in the popular vocabulary list.

[0072] This embodiment of the disclosure enhances the basic language model by using a popular word list to decode hot words, thereby increasing the probability that the basic language model can recognize the text corpus contained in the popular word list. As a result, after the basic language model decodes the speech to be recognized, the word sequence corresponding to the speech decoding result obtained can be the text corpus collected in the popular word list with the highest probability, thereby improving the recognition accuracy of popular media resources in speech recognition.

[0073] Based on the above embodiments, the method further includes: in response to the end of a preset period, inputting multiple text corpora contained in the popular word list into the training text and clearing the popular word list; using the training text to iteratively train the basic language model to obtain a newly trained language model; and replacing the basic language model with the newly trained language model.

[0074] At the end of a preset period, server 201 uses the text corpus entered into the popular vocabulary list as a training set to iteratively train the language model, obtaining a new language model. The base language model is then replaced with the newly trained language model to enhance the recognition accuracy of previously popular words. Simultaneously, in the next preset period, server 201 continuously collects text corpus from popular media resources. Using the newly collected text corpus from the popular vocabulary list and the newly trained language model, server 201 identifies subsequent speech from user 203. This is equivalent to training a new version of the language model in each preset period. This not only allows for a larger collection of popular media resource text corpus within the preset period, improving the training effect of the language model, but also reduces the update frequency of the language model, saving manpower and resources.

[0075] For example, if the preset period is set to 4 weeks, when the 5th week begins, the server 201 will input multiple text corpora containing the popular words from the previous 4 weeks into the training text as the training set for the basic language model. Then, the basic language model will be iteratively trained using the training text. After a new language model is obtained, the basic language model will be replaced with the newly trained language model.

[0076] Based on the above embodiments, multiple text corpora contained in the popular term list are entered into the training text, including: according to the number of times each popular media resource is screened, multiple text corpora contained in the popular term list are entered into the training text, so that the number of text corpora corresponding to popular media resources with more screened times is greater in the training text.

[0077] When server 201 inputs text corpora from the popular term list into the training text, it can maintain the same number of text entries in the training text as the number of entries in each corpus in the popular term list. For example, if a popular song A appears on the trending charts 3 times in 4 weeks, then 3 entries for the name of popular song A will be input into the training text. Alternatively, it can assign different weights to these corpora based on the number of preview screenings of popular media resources; those with more preview screenings will appear more frequently in the training text, and those with fewer preview screenings will appear less frequently.

[0078] In real-world scenarios, when users listen to popular songs / audiobooks, they also search for and listen to songs of the same type, by the same artist, or audiobooks by the same author. To enable the ASR system to support this extended search recognition task, the training corpus includes songs and audiobooks by artists and authors from the hot word list. Furthermore, texts of related songs / audiobooks from the top four weeks' charts can be added to the training corpus using tags such as similar music style and genre. Therefore, the training text obtained in this step consists of the corpus of top-charting songs / audiobooks from the top four weeks, plus the corpus expanded from them.

[0079] This disclosure embodiment increases the number of text entries in the training text corresponding to popular media resources that have been screened more frequently, thereby keeping the training set data more closely aligned with the actual collected data and making the training results of the language model more realistic.

[0080] Based on the above embodiments, the basic language model is iteratively trained using training text to obtain a newly trained language model, including: segmenting the text corpus in the training text according to the word segmentation vocabulary to obtain the word segmentation result, and adding newly emerging words from the Internet to the word segmentation vocabulary; fusing the word segmentation result with the basic pre-trained word segmentation according to a preset ratio to obtain a training word segmentation set; and using the training word segmentation set to iteratively train the basic language model to obtain a newly trained language model.

[0081] The text corpus input into the training text is segmented into words. During segmentation, new words can be added to the segmentation vocabulary as needed to ensure the accuracy of segmentation and the language model. Since hot words are usually newly emerging phrases or product names on the internet, these new phrases need to be added to the segmentation vocabulary to avoid splitting these new words or product names into multiple words during segmentation, thus preventing segmentation errors. Then, the segmentation results are fused with the segmentation training corpus of the basic language model for iterative training. By fusing the segmentation results with the basic pre-trained segmentation, the training segmentation set includes not only the segmentation results from the hot vocabulary but also the segmentation training corpus originally used for pre-training the basic language model, thus avoiding overfitting and increasing the generality of the language model. By adjusting the ratio of segmentation results from the hot vocabulary to the pre-trained segmentation training corpus, the language model is tested and iterated multiple times to train a high-performance language model.

[0082] This embodiment of the disclosure performs word segmentation on the text corpus in the popular word list to obtain the word segmentation results. Then, the word segmentation results are fused with the basic pre-trained word segments according to a preset ratio to obtain a training word segmentation set. The training word segmentation set is then used to iteratively train the basic language model. This ensures that the training word segmentation set not only includes the word segmentation results in the popular word list, but also includes the word segmentation training corpus originally used to pre-train the basic language model, thereby avoiding overfitting and increasing the generality of the language model.

[0083] Based on the above embodiments, before replacing the base language model with the newly trained language model, the method further includes: pruning the newly trained language model according to preset pruning coefficients.

[0084] Since the base language model is typically quite large, the server 201 can prune the newly trained model to make it suitable for devices with lower performance. For example, the preset pruning coefficient can be set to 10^-7. Pruning the newly trained language model using this preset coefficient can make it suitable for devices such as smart speakers and smart TVs.

[0085] This embodiment of the disclosure prunes the newly trained language model according to a preset pruning coefficient, which makes the newly trained language model applicable to devices with lower performance and increases its versatility.

[0086] Figure 3This is a schematic diagram of the structure of a popular media asset voice recognition device provided in an embodiment of this disclosure. The popular media asset voice recognition device may be a component or assembly in the terminal as described in the above embodiment. The popular media asset voice recognition device provided in this embodiment of the disclosure can execute the processing flow provided in the popular media asset voice recognition method embodiment, such as... Figure 3 As shown, the popular media resource speech recognition device 300 includes: a collection module 301, used to continuously collect text data about popular media resources within a preset period and input it into a popular word list; an acquisition module 302, used to acquire the speech to be recognized; and a recognition module 303, used to recognize the speech to be recognized based on the popular word list and a basic language model, and obtain the recognition result.

[0087] In some embodiments, when the collection module 301 continuously collects text data about popular media resources and inputs it into the popular word list within a preset period, it is specifically used to continuously crawl the text data of media resources located on the popular list and input it into the popular word list using a crawler program within the preset period.

[0088] In some embodiments, when the recognition module 303 recognizes the speech to be recognized based on the popular vocabulary and the basic language model and obtains the recognition result, it is specifically used to enhance the basic language model with hot word decoding using the popular vocabulary, so as to increase the probability of the basic language model recognizing the text corpus contained in the popular vocabulary; and to decode the speech to be recognized using the basic language model, and take the word sequence corresponding to the highest probability of the decoding result of the speech to be recognized as the recognition result.

[0089] In some embodiments, the popular media asset speech recognition device 300 further includes a training module 304, which, in response to the end of a preset period, inputs multiple text corpora contained in the popular word list into training text and clears the popular word list; iteratively trains the basic language model using the training text to obtain a newly trained language model; and replaces the basic language model with the newly trained language model.

[0090] In some embodiments, when the training module 304 inputs multiple text corpora contained in the popular terminology into the training text, it is specifically used to input multiple text corpora contained in the popular terminology into the training text according to the number of times each popular media resource is screened, so that the number of text corpora corresponding to popular media resources with more screened times is more in the training text.

[0091] In some embodiments, when the training module 304 iteratively trains the basic language model using training text to obtain a newly trained language model, it specifically performs word segmentation on the text corpus in the training text according to the word segmentation lexicon to obtain word segmentation results, wherein newly emerging words on the Internet are added to the word segmentation lexicon; the word segmentation results are fused with the basic pre-trained word segmentation according to a preset ratio to obtain a training word segmentation set; and the training word segmentation set is used to iteratively train the basic language model to obtain a newly trained language model.

[0092] In some embodiments, before replacing the base language model with the newly trained language model, the training module 304 prunes the newly trained language model according to a preset pruning coefficient.

[0093] Figure 3 The popular media asset voice recognition device shown in the embodiment can be used to execute the technical solution of the above method embodiment. Its implementation principle and technical effect are similar, and will not be described again here.

[0094] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. The electronic device can be a server as described in the above embodiment. The electronic device provided in this embodiment of the disclosure can execute the processing flow provided in the embodiment of the popular media asset speech recognition method, such as... Figure 4 As shown, the electronic device 400 includes: a memory 401, a processor 402, a computer program, and a communication interface 403; wherein the computer program is stored in the memory 401 and is configured to be executed by the processor 402 using the popular media asset speech recognition method described above.

[0095] In addition, this disclosure also provides a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the popular media asset speech recognition method described in the above embodiments.

[0096] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0097] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for speech recognition of popular media assets, characterized in that, include: Textual data about popular media resources is continuously collected and entered into a popular word list within a preset period, wherein the preset period is n times the update time of the popular media resources, and n is an integer greater than 1; Acquire the speech to be recognized; The speech to be identified is identified based on the aforementioned popular vocabulary and basic language model to obtain the recognition result; In response to the end of the preset period, the multiple text corpora contained in the popular word list are entered into the training text, and the popular word list is cleared. The basic language model is iteratively trained using the training text to obtain a newly trained language model. Replace the base language model with the newly trained language model.

2. The method as described in claim 1, characterized in that, The process of continuously collecting text corpora about popular media resources and inputting them into a popular vocabulary list within a preset period includes: Within a preset period, a web crawler program is used to continuously capture text corpora of media resources located on the trending list and input them into the trending word list.

3. The method as described in claim 1, characterized in that, The process of recognizing the speech to be recognized based on the popular vocabulary and the basic language model to obtain the recognition result includes: The hot word list is used to enhance the basic language model by decoding hot words, thereby increasing the probability that the basic language model can recognize the text corpus contained in the hot word list. The speech to be recognized is decoded using the basic language model, and the word sequence corresponding to the highest probability of the decoding result of the speech to be recognized is taken as the recognition result.

4. The method as described in claim 1, characterized in that, The step of inputting multiple text corpora contained in the popular vocabulary into the training text includes: Based on the number of times each popular media resource was screened, multiple text corpora contained in the popular term list were entered into the training text, so that the number of text corpora corresponding to popular media resources with more screenings would be greater in the training text.

5. The method as described in claim 1, characterized in that, The step of iteratively training the base language model using the training text to obtain a newly trained language model includes: The text corpus in the training text is segmented according to the word segmentation lexicon to obtain the word segmentation result. Newly emerging word groups on the Internet are added to the word segmentation lexicon. The word segmentation results are fused with the basic pre-trained word segments according to a preset ratio to obtain a training word segmentation set; The basic language model is iteratively trained using the training word segmentation set to obtain a newly trained language model.

6. The method as described in claim 5, characterized in that, Before replacing the base language model with the newly trained language model, the method further includes: The newly trained language model is pruned according to a preset pruning coefficient.

7. A voice recognition device for popular media assets, characterized in that, include: The collection module is used to continuously collect text corpora about popular media resources and input them into a popular word list within a preset period, wherein the preset period is n times the update time of the popular media resources, and n is an integer greater than 1; The acquisition module is used to acquire the speech to be recognized; The recognition module is used to recognize the speech to be recognized based on the popular vocabulary and the basic language model, and obtain the recognition result; The training module is used to input multiple text corpora from the popular vocabulary into the training text in response to the end of the preset period, and clear the popular vocabulary; iteratively train the basic language model using the training text to obtain a newly trained language model; and replace the basic language model with the newly trained language model.

8. An electronic device, characterized in that, include: Memory; processor; as well as Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

Images