Voice instruction recognition model customization method and device, computer equipment and storage medium

Abstract

The invention discloses a voice instruction recognition model customization method and device, computer equipment and a storage medium. The method comprises the steps that an instruction word list input by a user is acquired; performing splitting word segmentation on instruction words in the instruction word list to obtain a junk word list; converting the instruction word list words and the junk word list into a byte pair symbol sequence by using byte pair coding; based on the byte pair symbol sequence and the sensitivity parameter, using a weighted finite state converter to generate an instruction language model, and combining the instruction language model with a pre-trained acoustic model to form a complete voice instruction recognition model; synthesizing an audio sample of the instruction word by using a multi-speaker speech synthesis model, and automatically testing the recognition rate and the false recognition rate of the speech instruction recognition model; and packaging the voice instruction recognition model and delivering the voice instruction recognition model to the user. The instruction word list can be quickly customized, the voice instruction recognition model can be automatically adjusted and optimized, full-process automation is achieved, manual intervention is not needed, operation is convenient, and working efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of speech recognition technology, and in particular to a method, device, computer equipment and storage medium for customizing a speech command recognition model. Background Art

[0002] Voice command recognition refers to the user controlling electronic devices by speaking command words. Mobile phones, headphones, smart speakers, smart home appliances and other electronic products can use voice command recognition to allow users to control the device without touching the device, increasing the convenience and fun of electronic products.

[0003] Depending on whether the electronic device needs to be connected to the Internet, voice command recognition is divided into online recognition and offline recognition. Among them, online recognition can rely on the powerful computing power of the server, use a large vocabulary continuous speech recognition algorithm, and has a high recognition rate; while offline recognition is based on the computing power of the electronic device, and has certain limitations on speech recognition capabilities. Some low-cost devices, such as lamps, fans, or remote controls, usually use chips with low main frequency and memory capacity, and cannot run large vocabulary continuous speech recognition algorithms, and can only be limited to recognizing fixed command words. Equipment manufacturers need to customize command words according to different specifications and models of products. In the context of rapid iteration of electronic products, the command word list needs to be adjusted frequently. At present, there is a lack of methods to quickly customize the command word list and update the voice command recognition model, so the operation is cumbersome, difficult to produce, and inefficient. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a method for customizing a voice command recognition model, which can quickly customize a command vocabulary, automatically generate a voice command recognition model, and perform automated testing and tuning, which is convenient to operate and improves work efficiency.

[0005] On the one hand, an embodiment of the present invention provides a method for customizing a voice command recognition model, comprising: Acquire a command word list entered by a user, wherein the command word list includes command words and sensitivity; Decomposing and segmenting the instruction words in the instruction word list to obtain a junk word list; Converting the instruction word list words and the junk word list into a sequence of byte pair symbols using byte pair encoding; Based on the byte pair symbol sequence and the sensitivity parameter, a weighted finite state converter is used to generate a command language model, and the command language model is combined with a pre-trained acoustic model to form a complete voice command recognition model; Synthesizing an audio sample of the instruction word using a multi-speaker speech synthesis model, and passing the audio sample into the speech instruction recognition model for a recognition rate test; Using a preset test audio to perform an error rate test on the voice command recognition model; The voice command recognition model that has passed the test is packaged and delivered to the user.

[0006] According to some embodiments of the present invention, the instruction word list also includes instruction word replies.

[0007] According to some embodiments of the present invention, the weighted finite state converter has a grammatical converter, and the grammatical converter is a 2-gram language model including the instruction words and the junk words.

[0008] According to some embodiments of the present invention, the probability of all words in the grammar converter defaults to 1 / N, where N is the total number of entries. If the sensitivity of a certain instruction word is to be set higher than other words, the sensitivity of the instruction word is adjusted up to (1+D) / N, and the probability of other words is (1-D / (N-1)) / N, where D is the sensitivity coefficient.

[0009] According to some embodiments of the present invention, after the step of passing the audio sample into the voice command recognition model for recognition rate testing, the step includes: If the recognition rate of a certain instruction word does not meet the requirement, the weight of the instruction word in the grammar converter is increased and the instruction language model is regenerated.

[0010] According to some embodiments of the present invention, after the step of using a preset test audio to perform an error rate test on the voice command recognition model, the following steps are included: If the error recognition rate of a certain instruction word exceeds the standard, the weight of the instruction word in the grammar converter is increased and the instruction language model is regenerated.

[0011] According to some embodiments of the present invention, the step of packaging the voice command recognition model that has passed the test and delivering it to the user includes: Package the voice command recognition model into an executable file on a computer, an update package, or firmware for an embedded device; Delivered to users via online download or email.

[0012] On the other hand, an embodiment of the present invention provides a voice command recognition model customization device, which is used to implement the above-mentioned voice command recognition model customization method, and the device includes: An instruction word acquisition module, used to acquire an instruction word list entered by a user, wherein the instruction word list includes instruction words and sensitivity; An instruction word conversion module, used for decomposing and segmenting instruction words in the instruction word list to obtain a junk word list, and converting the instruction word list words and the junk word list into a byte pair symbol sequence using byte pair encoding; A command recognition model generation module, for generating a command language model using a weighted finite state converter based on the byte pair symbol sequence and the sensitivity parameter, and combining the command language model with a pre-trained acoustic model to form a complete voice command recognition model; The command recognition model testing module is used to synthesize the audio sample of the command word using the multi-speaker speech synthesis model, pass the audio sample into the voice command recognition model for recognition rate testing, and use the preset test audio to perform an error recognition rate test on the voice command recognition model; The command recognition model packaging module is used to package the voice command recognition model that has passed the test and deliver it to the user.

[0013] On the other hand, an embodiment of the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program, the processor executes the above-mentioned voice command recognition model customization method.

[0014] On the other hand, an embodiment of the present invention provides a storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned voice instruction recognition model customization method are implemented.

[0015] The embodiments of the present invention have at least the following beneficial effects: The present invention proposes a method for customizing a voice command recognition model. The user submits a list of command words to be recognized online; the command words in the command word list are decomposed and segmented to obtain a list of junk words; the command word list words and the junk word list are converted into a byte pair symbol sequence using byte pair encoding; based on the byte pair symbol sequence and sensitivity parameters, a weighted finite state converter is used to generate a command language model, and the command language model is combined with a pre-trained acoustic model to form a complete voice command recognition model; a multi-speaker speech synthesis model is used to synthesize an audio sample of the command word, and the audio sample is passed into the voice command recognition model for recognition rate testing; the voice command recognition model is tested for error rate using a preset test audio; and the voice command recognition model that passes the test is packaged and delivered to the user. After the user submits the command word list, the server background automatically adjusts parameters according to the command word list and sensitivity level and generates a voice command recognition model, and then automatically tests and tunes the voice command recognition model. In the whole process, except that the command word list needs to be manually input, the whole process is automated without manual intervention; the work that originally took several days to complete is reduced to just a few minutes. Using speech synthesis technology to simulate recognition rate testing can accurately control the trade-off between recognition rate and false recognition rate. It can quickly customize the voice command recognition model, which is convenient to operate and improves work efficiency.

[0016] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 A flowchart of a method for customizing a voice command recognition model according to an embodiment of the present invention; Figure 2 A detailed flow chart of a method for customizing a voice command recognition model according to an embodiment of the present invention; Figure 3 A module diagram of a voice command recognition model customization device using an embodiment of the present invention; Figure 4 The figure is a schematic diagram of the internal structure of a computer device using an embodiment of the present invention. DETAILED DESCRIPTION

[0018] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0019] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., and orientations or positional relationships indicated are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0020] In the description of the present invention, "several" means one or more, "multiple" means more than two, greater than, less than, and exceeding are understood as not including the number itself, and "above", "below", and "within" are understood as including the number itself. If there is a description of "first", "second", etc., it is only used for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0021] In the description of the present invention, unless otherwise clearly defined, words such as “set”, “install”, “connect” and “connected” should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above words in the present invention based on the specific content of the technical solution.

[0022] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments.

[0023] See also Figure 1 This embodiment provides a method for customizing a voice command recognition model, which mainly includes steps S101 to S107: S101, obtaining a command word list entered by a user, the command word list including command words and sensitivity; S102, breaking down and segmenting the instruction words in the instruction word list to obtain a junk word list; S103, using byte pair encoding to convert the instruction word list words and the junk word list into a byte pair symbol sequence; S104, based on the byte pair symbol sequence and the sensitivity parameter, using a weighted finite state transformer WFST to generate a command language model, and combining the command language model with a pre-trained acoustic model to form a complete voice command recognition model; S105, synthesizing an audio sample of the command word using a multi-speaker speech synthesis model, and passing the audio sample into a speech command recognition model for a recognition rate test; S106, using a preset test audio to perform an error recognition rate test on the voice command recognition model; S107: Package the voice command recognition model that has passed the test and deliver it to the user.

[0024] After the user submits the command word list on the command word entry page, the server backend automatically adjusts the parameters according to the command word list and sensitivity level and generates a voice command recognition model, and then automatically tests and tunes the voice command recognition model. In the whole process, except for the command word list, which needs to be manually input, the whole process is automated without manual intervention; the work that originally took several days to complete is reduced to just a few minutes. Using speech synthesis technology to simulate the recognition rate test can accurately control the trade-off between the recognition rate and the error recognition rate. The voice command recognition model can be quickly customized, which is convenient to operate and improves work efficiency.

[0025] See also Figure 2 The method for customizing a voice command recognition model in an embodiment of the present application may include the following detailed steps: S201. The user submits a list of command words that he wants to recognize through the command word entry page. The command word list includes command words, command word replies and sensitivity. Among them, the sensitivity parameter affects the sensitivity index of the recognition rate test. For example, the sensitivity is set to high, normal or low, corresponding to 95%, 90% and 85% recognition rates respectively. The command word reply field is the prompt word played when the command word is recognized, which is not a required item. It should be noted that the language of the command word can be Chinese or other languages, but if a certain language is to be recognized, the back end must first prepare a pre-trained acoustic model of the corresponding language. When the command word entry page is submitted, the command words filled in by the user must be checked to avoid repeated filling, the same pronunciation or too close command words.

[0026] S202. After receiving the command word list, the backend server decomposes all the command words in the command word list to obtain a junk word list. Among them, junk words belong to a subset of command words. For example, if the command word is "turn on the light", then the words "turn on", "turn on the light", "light" and so on can be added to the junk word list. If the junk word and the command word are the same, they are removed from the junk word list. Adding junk words can effectively reduce the error recognition rate of the command word without affecting the sensitivity of the command word.

[0027] S203. Use byte pair encoding to convert the instruction word list words and the junk word list into a byte pair symbol sequence. The byte pair encoding used must be consistent with the byte pair encoding used by the acoustic model. Byte pair encoding BPE (Byte Pair Encoding) is a data compression algorithm that iteratively replaces the most frequent byte pair in a sequence with an unused byte. For spelling languages ​​such as English, byte pair encoding can convert even non-words, such as certain brand names, into a symbol sequence. Compared with traditional phoneme or polyphone-based modeling units, byte pair encoding lowers the phonetic knowledge threshold for training models, can easily train models in multiple languages, and help models better handle unknown and rare words, improving the generalization ability of the model.

[0028] S204, using the weighted finite state converter WFST to generate the instruction language model. The specific method is to synthesize and simplify the symbol converter T, the dictionary converter L and the grammar converter G. The symbol converter T is responsible for mapping the output of the acoustic model into a symbol sequence, which is a byte pair encoded symbol. The symbol converter T and the dictionary converter L are both determined by the byte pair encoding and do not need to be changed. The grammar converter G comes from the instruction word list and the junk word list; the sensitivity of each instruction word can be adjusted. Specifically, G is a 2-gram language model containing instruction words and junk words, in which the probability of all words defaults to 1 / N, and N is the total number of entries. If you want a certain instruction word to be more sensitive than other words, then the sensitivity of the word can be increased to (1+D) / N, and the probability of other words can be reduced to (1-D / (N-1)) / N. D is the sensitivity coefficient, and the sensitivity coefficient D is generally taken as 0.2 according to experience. The instruction language model is combined with the pre-trained acoustic model to form a complete voice instruction recognition model. The complete voice instruction recognition model also includes an acoustic model trained on a large amount of corpus. When a sequence of audio features is input, the acoustic model will output a symbol sequence, and the voice command recognition model can determine whether the symbol sequence contains a certain command word. If the command word is not included, the voice command recognition model will not return anything. Among them, the dictionary converter L maps the modeling unit sequence to a word sequence, and the mapping is determined by the byte pair encoding model. Common symbol converters T include temporal connection classification CTC, hidden Markov HMM, etc. If the command word filled in by the user is not a word already in the vocabulary, the byte pair encoding model will decompose the command word into a symbol sequence and add the mapping to the dictionary converter L. The symbol converter T, dictionary converter L and grammar converter G are synthesized and simplified according to the following formula: In the formula, represents the composition operation, min and det represent the minimization and determinization operations respectively. This sequence of operations can effectively reduce the number of states and edges, and reduce computing power and storage requirements without changing the output of the weighted finite state transformer WFST.

[0029] It should be noted that the weighted finite state transducer WFST (Weighted Finite State Transducer) can be regarded as an automaton in the form of a directed graph Q, consisting of a set of states and transitions between states. Unlike traditional finite state automata, the transitions in the weighted finite state transducer WFST are also accompanied by weights. These weights can represent the relative importance or probability of different transitions. Therefore, the weighted finite state transducer WFST can not only accept an input sequence, but also assign a weight to each possible output sequence. In the field of speech recognition, the weighted finite state transducer WFST is widely used in the construction of acoustic models and language models. The acoustic model is used to convert speech signals into text, while the language model is used to predict the most likely word sequence that appears in a given context. By combining these two models and using the weighted finite state transducer WFST to encode the relationship between them, an accurate speech recognition system can be achieved.

[0030] S205. Use a multi-speaker speech synthesis model to synthesize multiple command word audio samples, and pass the samples into the speech command recognition model to test its recognition rate. The multi-speaker speech synthesis model is trained with a data set containing multiple speakers, and can synthesize the timbre of multiple speakers to avoid the situation where the test results are inconsistent with the actual product experience due to a single timbre. If the recognition rate does not meet the requirements and the test fails, return to step S204, adjust the parameters of the grammar converter G, and if the recognition rate of a certain command word does not meet the standard, increase the weight of the command word in the 2-gram language model, for example, increase it by 20%, and the weight of the command word that has met the standard remains unchanged, and then re-synthesize the command language model.

[0031] It should be noted that the latest speech synthesis technology is sufficient to synthesize speech that is very close to human voice and can replace humans to test speech recognition systems. Multi-speaker speech synthesis models include SPEAR-TTS, Mellotron, CoVoMix, and ChatTTS. For example, SPEAR-TTS is used to divide the TTS (Text-to-Speech) task into two stages. The first stage maps the text to high-order semantic tokens, and the second stage maps the semantic tokens to low-order acoustic tokens. The continuous speech waveform is mapped to a token sequence with a limited vocabulary, thereby converting the speech synthesis problem into a language model task. Through pre-training and back-translation technology, SPEAR-TTS can stably reproduce the timbre of the target speaker with only 3 seconds of speech, and the synthesized speech is highly natural and comparable to real speech.

[0032] S206. Use pre-recorded audio to test the error recognition rate of the voice command recognition model. The test audio content contains voice and various environmental sounds, and the length needs to be at least 24 hours to more accurately count the error recognition rate of each command word. Perform feature extraction on a longer audio segment and use the pre-trained acoustic model for model inference, and store the model inference results on the server. This step takes a long time, but it only needs to be performed once, and the model inference results can be directly read in each subsequent test. The error recognition rate refers to the number of times a sound segment that does not contain a command word is mistakenly recognized as a command word per unit time. Each time the error recognition rate test is performed, the model inference result is input into the weighted finite state converter WFST, and the number of times each command word is recognized is counted; the number is divided by the total duration of the audio, which is the error recognition rate for each command word. Because it is run on the server, it can be completed in as little as a few minutes. If the error recognition rate exceeds the standard and the test fails, return to step S204 to adjust the parameters of the grammar converter G. Adjustment methods include increasing the 2-gram weight of junk words, for example, by 20%, or reducing the 2-gram weight of command words with high false positive rates, for example, by 20%. If the recognition rate and false positive rate cannot meet the standards at the same time no matter how the parameters of the grammar converter G are adjusted, the recognition rate will be prioritized. Using speech synthesis technology to simulate recognition rate testing can accurately control the trade-off between recognition rate and false positive rate.

[0033] S207, packaging the voice command recognition model into an executable file on a computer, an update package, or firmware for an embedded device; and delivering it to the user by online download or email. If the user is satisfied with the voice command recognition effect, the user can directly develop a product based on the voice command recognition model. If the user is still not satisfied with the voice command recognition effect, the user can adjust the sensitivity of the command and regenerate the recognition model.

[0034] See also Figure 3The embodiment of the present application also provides a voice command recognition model customization device, which is used to implement the above-mentioned voice command recognition model customization method, and the device includes: The instruction word acquisition module 301 is used to acquire an instruction word list entered by a user, wherein the instruction word list includes instruction words and sensitivity; An instruction word conversion module 302 is used to decompose and segment instruction words in the instruction word list to obtain a junk word list, and convert the instruction word list words and the junk word list into a byte pair symbol sequence using byte pair encoding; The instruction recognition model generation module 303 is used to generate an instruction language model based on the byte pair symbol sequence and the sensitivity parameter using a weighted finite state converter, and combine the instruction language model with the pre-trained acoustic model to form a complete voice instruction recognition model; The command recognition model testing module 304 is used to synthesize an audio sample of a command word using a multi-speaker speech synthesis model, pass the audio sample to the speech command recognition model for recognition rate testing, and use a preset test audio to perform an error rate test on the speech command recognition model; The instruction recognition model packaging module 305 is used to package the voice instruction recognition model that has passed the test and deliver it to the user.

[0035] It should be noted that the present embodiment of the voice command recognition model customization device and the above-mentioned voice command recognition model customization method embodiment are based on the same inventive concept. The contents not involved in the present embodiment of the voice command recognition model customization device can be referred to the above-mentioned voice command recognition model customization method embodiment and will not be repeated here.

[0036] See also Figure 4 The embodiment of the present application further provides a computer device, which includes a memory 401 and a processor 402. The processor 402 is used to run the computer program instructions in the memory 401 to implement the voice command recognition model customization method.

[0037] The memory 401 includes at least one type of readable storage medium, which includes a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory 401 may be an internal storage unit of a computer device, such as a hard disk of a computer device. In other embodiments, the memory 401 may also be a storage device of an external computer device, such as a plug-in hard disk configured in a computer device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. The memory 401 may also include both an internal storage unit of a computer device and an external storage device. The memory 401 may be used not only to store application software and various types of data installed in a computer device, but also to temporarily store data that has been output or is to be output.

[0038] The computer device also includes a bus 403. The bus 403 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0039] The computer device may further include a display component 404. The display component 404 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display component 404 may also be appropriately referred to as a display device or a display unit, which is used to display information processed in the computer device and to display a visual user interface.

[0040] The computer device may further include a communication component 405. The communication component 405 may optionally include a wired communication component and / or a wireless communication component (such as a WI-FI communication component, a Bluetooth communication component, etc.), which is generally used to establish a communication connection between the computer device and other electronic devices.

[0041] Figure 4 Only a computer device with some components is shown, and those skilled in the art will understand that Figure 4 The illustrated structure does not constitute a limitation on the computer device 100 , and the computer device 100 may include fewer or more components than illustrated, or may combine certain components, or may have different component arrangements.

[0042] The embodiment of the present application also provides a storage medium, when the instructions in the storage medium are executed by the processor of the computer, the computer can execute the voice instruction recognition model customization method provided in the above embodiment. For example, the storage medium can be ROM, RAM, CD-ROM, magnetic tape, floppy disk, USB flash drive and optical data storage device. It is worth noting that the storage medium mentioned in the embodiment of the present application can be a non-volatile storage medium, in other words, it can be a non-transient storage medium.

[0043] It should be understood that all or part of the steps to implement the above embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions; the computer instructions can be stored in the above storage medium. That is, in some embodiments, a computer program product containing instructions is also provided, which, when run on a computer, enables the computer to execute the voice instruction recognition model customization method provided in the above embodiments.

[0044] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A method for customizing a voice command recognition model, characterized in that: include: Acquire a command word list entered by a user, wherein the command word list includes command words and sensitivity; Decomposing and segmenting the instruction words in the instruction word list to obtain a junk word list; Converting the instruction word list words and the junk word list into a sequence of byte pair symbols using byte pair encoding; Based on the byte pair symbol sequence and the sensitivity parameter, a weighted finite state converter is used to generate a command language model, and the command language model is combined with a pre-trained acoustic model to form a complete voice command recognition model; Synthesizing an audio sample of the instruction word using a multi-speaker speech synthesis model, and passing the audio sample into the speech instruction recognition model for a recognition rate test; Using a preset test audio to perform an error rate test on the voice command recognition model; The voice command recognition model that has passed the test is packaged and delivered to the user.

2. The method for customizing a voice command recognition model according to claim 1, characterized in that: The instruction word list also includes instruction word replies.

3. The method for customizing a voice command recognition model according to claim 1, characterized in that: The weighted finite state converter has a grammatical converter, and the grammatical converter is a 2-gram language model including the instruction words and the junk words.

4. The method for customizing a voice command recognition model according to claim 3, characterized in that: The default probability of all words in the grammar converter is 1 / N, where N is the total number of entries. If you want to set the sensitivity of a certain instruction word higher than other words, then the sensitivity of the instruction word is adjusted to (1+D) / N, and the probability of other words is (1-D / (N-1)) / N, where D is the sensitivity coefficient.

5. The method for customizing a voice command recognition model according to claim 4, characterized in that: After the step of transferring the audio sample to the voice command recognition model for recognition rate testing, the method further comprises: If the recognition rate of a certain instruction word does not meet the requirement, the weight of the instruction word in the grammar converter is increased and the instruction language model is regenerated.

6. The method for customizing a voice command recognition model according to claim 5, characterized in that: After the step of using the preset test audio to perform an error rate test on the voice command recognition model, the method further comprises: If the error recognition rate of a certain instruction word exceeds the standard, the weight of the instruction word in the grammar converter is increased and the instruction language model is regenerated.

7. The method for customizing a voice command recognition model according to claim 1, characterized in that: The step of packaging the voice command recognition model that has passed the test and delivering it to the user includes: Package the voice command recognition model into an executable file on a computer, an update package, or firmware for an embedded device; Delivered to users via online download or email.

8. A voice command recognition model customization device, characterized in that: include: An instruction word acquisition module, used to acquire an instruction word list entered by a user, wherein the instruction word list includes instruction words and sensitivity; An instruction word conversion module, used for decomposing and segmenting instruction words in the instruction word list to obtain a junk word list, and converting the instruction word list words and the junk word list into a byte pair symbol sequence using byte pair encoding; A command recognition model generation module, for generating a command language model using a weighted finite state converter based on the byte pair symbol sequence and the sensitivity parameter, and combining the command language model with a pre-trained acoustic model to form a complete voice command recognition model; The command recognition model testing module is used to synthesize the audio sample of the command word using the multi-speaker speech synthesis model, pass the audio sample into the voice command recognition model for recognition rate testing, and use the preset test audio to perform an error recognition rate test on the voice command recognition model; The command recognition model packaging module is used to package the voice command recognition model that has passed the test and deliver it to the user.

9. A computer device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and when the processor runs the computer program, the processor executes the voice instruction recognition model customization method according to any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the voice instruction recognition model customization method described in any one of claims 1 to 7 are implemented.

Images