Method for recognizing command words in a speech signal and related device

Abstract

This application belongs to the field of speech recognition and discloses a method and related equipment for recognizing command words in speech signals, which can improve the accuracy of command word recognition. The method includes: inputting a target continuous speech signal into a speech recognition model to obtain a set of phoneme sequence paths and a path score corresponding to each phoneme sequence path in the set; if there is a set of scoring paths in the set of phoneme sequence paths with a path score greater than a preset threshold, determining the output matrix sequence corresponding to each scoring path in the set; performing backtracking expansion on the output matrix sequence corresponding to each scoring path to obtain a target candidate region for each scoring path; determining the target phoneme of the target command word to be recognized and the non-target phonemes in the target candidate region; and determining whether the target command word is recognized in the target candidate region based on the target phoneme and the non-target phonemes.

Description

Technical Field

[0001] This application relates to the field of speech recognition, and in particular to a method and related equipment for recognizing command words in speech signals. Background Technology

[0002] While command word decoding algorithms can efficiently and quickly calculate scores, they are not effective for statements containing target command words (such as "lock"), even if the actual intention of the speech is not to perform the speech scenario corresponding to the command word. For example, statements such as "forgot to lock", "locked", "don't lock", and "locked" all mention the word "lock", but they are actually meant to remind, confirm, refuse, or state completion, not to actually perform the "lock" action.

[0003] Currently, decoding algorithms for command words only check if the sentence contains the command word "lock," without understanding the true intent of the sentence. The algorithm will treat it as "to execute the lock" and calculate the score, unaware that it is not actually necessary. In the misidentification problem, some misidentifications are in the form of corresponding speech that should not be recognized, such as the command word "lock." When the voice input is "forgot to lock," or "locked," or "not lock," or "locked," etc., although these voices contain the command word, the actual intent is not to execute the lock action, which leads to low accuracy in recognizing command words in speech. Summary of the Invention

[0004] This application provides a method and related equipment for recognizing command words in speech signals, which can improve the accuracy of command word recognition.

[0005] The first aspect of this application proposes a method for recognizing command words in speech signals, comprising: The target continuous speech signal is input into the speech recognition model to obtain a set of phoneme sequence paths and a path score corresponding to each phoneme sequence path in the set of phoneme sequence paths. If there is a set of scoring paths in the phoneme sequence path set whose path scores are greater than a preset threshold, then determine the output matrix sequence corresponding to each scoring path in the scoring path set. The output matrix sequence corresponding to each scoring path is backtracked and expanded to obtain the target candidate region for each scoring path; Determine the target phonemes of the target command word to be identified and the non-target phonemes in the target candidate region; Based on the target phonemes and the non-target phonemes, determine whether the target command word is identified in the target candidate region.

[0006] In one possible design, the backtracking expansion of the output matrix sequence corresponding to each scoring path to obtain the target candidate region for each scoring path includes: Determine the first frame, the last frame, and the number of backtracking frames corresponding to the target continuous speech signal for each scoring path in the output matrix sequence. The target candidate region for each scoring path is determined from the target continuous speech signal based on the first frame, the last frame, and the number of backtracking frames.

[0007] In one possible design, determining whether the target command word is identified in the target candidate region based on the target phoneme and the non-target phonemes includes: The first number of non-target phonemes in the backtracked frames and the second number of non-target phonemes in the target candidate region are counted. Based on the first quantity and the second quantity, determine whether the target command word is identified in the target candidate region.

[0008] In one possible design, determining whether the target command word is identified in the target candidate region based on the first quantity and the second quantity includes: Determine whether the first quantity is greater than a first preset value, and determine whether the second quantity is greater than a second preset value; If the first quantity is greater than the first preset value or the second quantity is greater than the second preset value, then it is determined that the target candidate region has not identified the target command word; If the first quantity is less than or equal to the first preset value, and the second quantity is less than or equal to the second preset value, then it is determined that the target candidate region has identified the target command word.

[0009] In one possible design, the method further includes: Step 1: Randomly select a set of initialization parameters, including an initial score threshold, an initial backtracking frame count, a first initial preset value, and a second initial preset value; Step 2: Test the initialization parameters based on the dataset corresponding to the target command word to obtain test results; Step 3: Adjust the initialization parameters according to the test results, and test the adjusted initialization parameters based on the dataset; Steps 2 to 3 are executed iteratively until a preset iteration termination condition is reached, thereby obtaining the preset threshold, the number of backtracking frames, the first preset value, and the second preset value.

[0010] In one possible design, determining the output matrix sequence corresponding to each scoring path in the scoring path set includes: Determine the time frame range information corresponding to each phoneme in the target scoring path, wherein the target scoring path is any one of the scoring paths in the set of scoring paths; The output matrix sequence is extracted from the target continuous speech signal based on the time frame range information corresponding to each phoneme.

[0011] A second aspect of this application provides a device for recognizing command words in a speech signal, comprising: The recognition module is used to input the target continuous speech signal into the speech recognition model to obtain a set of phoneme sequence paths and a path score corresponding to each phoneme sequence path in the set of phoneme sequence paths. The first determining module is used to determine the output matrix sequence corresponding to each scoring path in the phoneme sequence path set when there is a scoring path set in the phoneme sequence path set where the path score is greater than a preset threshold. The backtracking module is used to backtrack and expand the output matrix sequence corresponding to each scoring path to obtain the target candidate region for each scoring path. The second determining module is used to determine the target phonemes of the target command word to be identified and the non-target phonemes in the target candidate region; The judgment module is used to determine whether the target command word is recognized in the target candidate region based on the target phoneme and the non-target phoneme.

[0012] A third aspect of this application provides a device for recognizing command words in speech signals, comprising: processor; Memory, used to store computer programs; The processor executes the computer program to implement the method for recognizing command words in the speech signal as described in any of the above embodiments.

[0013] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for recognizing command words in a speech signal as described in any of the above embodiments.

[0014] The fifth aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the method for recognizing command words in a speech signal as described in any of the above embodiments.

[0015] In the embodiments provided in this application, after identifying several scoring paths, each scoring path is backtracked and expanded to expand the phoneme sequence. This allows the capture of contextual information such as the coprolative effects with the command word, and the counting of the number of non-target phonemes as a discriminative feature. This enables the system to keenly identify those segments that, although their acoustic features are similar to the command word, are actually other words or noise. Thus, without sacrificing response speed, the accuracy of command word recognition is significantly improved. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a system for recognizing command words in a speech signal according to an embodiment of this application; Figure 2 A schematic diagram of the structure of a device for recognizing command words in a speech signal according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application; The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] Those skilled in the art will understand that, unless explicitly stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of this application means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when an element is “connected” or “coupled” to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein may include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.

[0019] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0020] For ease of understanding, the following is a description of the specialized terms used in this application: Command words: In a speech recognition system, specific words or phrases that need to be recognized, such as "turn on the lights" or "turn off the lock". Phoneme: The smallest unit in speech, comprising two main categories: vowels and consonants.

[0021] The following describes the method for recognizing command words in a speech signal provided in this application from the perspective of a device for recognizing command words in a speech signal. The device for recognizing command words in a speech signal can be a server or a service unit within a server, and is not specifically limited. For the sake of simplicity, the following description uses a server as an example to illustrate the device for recognizing command words in a speech signal.

[0022] Reference Figure 1 , Figure 1 This is a flowchart illustrating a method for recognizing command words in a speech signal according to an embodiment of this application. The method includes: 101. Input the target continuous speech signal into the speech recognition model to obtain the phoneme sequence path set and the path score corresponding to each phoneme sequence path in the phoneme sequence path set.

[0023] In this embodiment, the server can pre-train an end-to-end speech recognition model, such as Transformer or CNN-LSTM. This speech recognition model is trained on a large amount of corpus containing target command words. Thus, when recognizing a target continuous speech signal to determine whether it contains the target command word, the server can acquire the target continuous speech signal and input it into the pre-trained speech recognition model. The speech recognition model decodes the target continuous speech signal to identify the phoneme sequence (or word sequence) corresponding to the target continuous speech signal. The speech recognition model then calculates several possible phoneme sequence paths, i.e., a set of phoneme sequence paths, and assigns an overall score to each phoneme sequence path in the set of phoneme sequence paths, i.e., the path score corresponding to each phoneme sequence path.

[0024] It should be noted that when the speech recognition model decodes the target continuous speech, a search space will be constructed. The vertical axis of this search space is phonemes, states or words, the horizontal axis is time (or frames), and the path is each time point; select a state from all possible candidate states, and connect all the selected states in the time order of the states to form a complete sequence, which is the phoneme sequence path; The score of this path is the comprehensive score, including the acoustic score and the language model score. The acoustic score represents the probability that each frame of speech feature belongs to a certain phoneme or state, and represents the possibility that the generated word sequence conforms to language habits (grammar, semantics). For example, as a legal instruction, the language model score of "guan suo" will be much higher than that of a meaningless combination like "guan suo". The final path score is a weighted combination of these two scores.

[0025] 102. If there is a set of scored paths with path scores greater than the preset threshold in the set of phoneme sequence paths, determine the output matrix sequence corresponding to each scored path in the set of scored paths.

[0026] In this embodiment, after obtaining the path score, the server can determine whether there is a phoneme sequence path with a path score greater than the preset threshold in the set of phoneme sequence paths. If so, define the phoneme sequence paths with path scores greater than the preset threshold in the set of phoneme sequence paths as scored paths to form a set of scored paths; After that, the server can determine the output matrix sequence corresponding to the scored paths in the set of scored paths. Specifically, the server can first determine the time frame range information corresponding to each phoneme in the target scored path, where the target scored path is any one of the scored paths in the set of scored paths; extract the output matrix sequence from the target continuous speech signal based on the time frame range information; that is, after the server decodes the target continuous speech through the speech recognition model to obtain the scored path, the scored path already contains the time boundary information of each phoneme, and this scored path will show the time frame ranges corresponding to the phoneme sequences of the target command word (for example, if the target command word is "guan suo", then this scored path will show the phoneme sequence g, uan, s, uo of "guan suo", and the time frame ranges corresponding to the phoneme sequence g, uan, s, uo respectively, such as the phoneme "g" corresponding to frames 100 - 105). The server can use this time mapping relationship to extract the corresponding time frame ranges from the complete phoneme probability matrix, thereby extracting the time frames corresponding to each phoneme in the phoneme sequence from the target continuous speech signal, and splicing the time frames corresponding to each phoneme in the event order to obtain the output matrix sequence corresponding to each scored path in the set of scored paths.

[0027] 103. Backtrack and expand the output matrix sequence corresponding to each scoring path to obtain the target candidate region for each scoring path.

[0028] In this embodiment, the server determines the first frame, the last frame of the output matrix sequence corresponding to each scoring path, and the number of backtracking frames corresponding to the target continuous speech signal; determines the target candidate region from the target continuous speech signal based on the first frame, the last frame, and the number of backtracking frames.

[0029] It should be noted that the pronunciation of speech is a continuous process, and the pronunciation at the beginning of the target command word may be affected by the phonemes in front of the target command word (co-articulation effect). In order to obtain a more accurate command word recognition effect or perform subsequent misrecognition judgment, in this embodiment, the determined output matrix sequence is backtracked and expanded. Therefore, when performing command word recognition, the target continuous speech signal is not extracted starting from the starting frame (for example, the Nth frame) of the target command word determined by the scoring path, but is extracted starting from the starting frame of the target command word and backtracking k frames on the time axis. This k value is the number of backtracking frames corresponding to the target continuous speech signal. For example, backtracking 15 frames (about 600 milliseconds) to ensure that the speech context information before the start of the command word can be covered. For example, the first frame of this output matrix sequence is N, the last frame is M, and the number of backtracking frames is k. Then the first frame of the target candidate region is the (N - k)th frame, and the last frame of the target candidate region is M.

[0030] 104. Determine the target phonemes of the target command word to be recognized and the non-target phonemes in the target candidate region.

[0031] In this embodiment, the server can determine the target phonemes of the target command word to be recognized. These target phonemes are the phonemes that make up the target command word itself. For example, the target phonemes corresponding to the target command word "lock" are "g, uan, s, uo"; at the same time, the server can determine the non-target phonemes in the target candidate region. For example, if the phonemes in this target candidate region are "w oo o gg uan x iao ia s z uo", then the non-target phonemes in this target candidate region are: "w oo o x iao iaz".

[0032] 105. Judge whether the target command word is recognized in the target candidate region according to the target phonemes and the non-target phonemes.

[0033] In this embodiment, after initially determining the target candidate region, the server can evaluate the risk of command word misrecognition by quantifying the number of non-target phonemes in this target candidate interval. For the convenience of understanding, the following is a specific explanation: The server can count the first number of non-target phonemes in the backtracking frames and the second number of non-target phonemes in the target candidate region. As shown in the example above, the first number c1 is the number of non-target phonemes that appear in k frames before the first phoneme of the command word (such as "g") appears. This metric measures whether the speech environment in the target continuous speech signal contains too much irrelevant speech or noise before the target command word begins. If the first number c1 is too large, it means that the speech recognition model may incorrectly associate some irrelevant speech or noise before the target command word with the target candidate region, which leads to the first number c1 of non-target phonemes being too large. The second quantity c2 is the number of non-target phonemes that appear in the entire target candidate region (from frame N to frame M). The second quantity c2 measures whether the entire target candidate region where the target command word is located contains too many other phonemes. If c2 is too large, it indicates that the speech corresponding to the target candidate region may not be the target command word at all, but another word with similar pronunciation. Subsequently, the server determines whether the target command word is identified in the target candidate region based on the first quantity and the second quantity. Specifically, the server can determine whether the first quantity is greater than a first preset value and whether the second quantity is greater than a second preset value. If the first quantity is greater than the first preset value or the second quantity is greater than the second preset value, it is determined that the target command word is not identified in the target candidate region. If the first quantity is less than or equal to the first preset value and the second quantity is less than or equal to the second preset value, it is determined that the target command word is identified in the target candidate region.

[0034] In other words, the server pre-sets two preset values ​​t1 and t2. If too many non-target phonemes appear in the stage before the target command word begins (i.e., within the backtracking frame number) (the first quantity c1 exceeds the first preset value t1), it is determined that there is a lot of irrelevant speech in the backtracking frame number, and misidentification may be triggered by irrelevant speech in the backtracking frame number. If too many non-target phonemes are mixed in the target candidate area corresponding to the entire target command word (the second quantity c2 exceeds the second preset value t2), it can be determined that the target command word has not been identified in the target candidate area. That is, as long as any of the above is triggered, it will be determined that there are a lot of non-target phonemes in the target candidate area, which is likely to be a misidentification. Only when neither of the two defense lines is triggered (c1≤t1 and c2≤t2) is it determined that the target command word exists in the target candidate area, that is, the target command word has been identified in the target continuous speech signal.

[0035] It should be noted that the settings of the first preset value, the second preset value, the number of backtracking frames, and the preset threshold are crucial to achieving a balance between accuracy and false recognition rate in the entire system. These parameters cannot be directly derived through theoretical calculations but require extensive experimentation, debugging, and data analysis to determine. Specifically: Step 1: Randomly select a set of initialization parameters, including the initial score threshold, the initial number of backtracking frames, the first initial preset value, and the second initial preset value; Step 2: Test the initialization parameters based on the dataset corresponding to the target command word to obtain the test results; Step 3: Adjust the initialization parameters based on the test results, and test the adjusted initialization parameters based on the dataset; Steps 2 to 3 are executed iteratively until the preset iteration termination condition is reached, and the preset threshold, backtracking frame number, first preset value and second preset value are obtained.

[0036] That is, the server builds a dataset that includes positive and negative samples. The positive samples are recordings of the target command word (such as "lock") pronounced clearly by different speakers, at different speeds, and in different environments (quiet, slightly noisy). The negative samples include words with similar pronunciation to the target command word (such as "unlock"), phrases or sentences containing some phonemes of the target command word (such as "help me lock it"), and common environmental noises (such as keyboard sounds, door closing sounds, television sounds, and conversations between multiple people). The server tests on the dataset using the current initialization parameters (initial score threshold, initial backtracking frame count, first initial preset value, and second initial preset value) and calculates the False Accept Rate (FAR) and False Reject Rate (FRR). Based on the calculated FAR and FRR, the initialization parameters are adjusted. If there are too many false acceptances (FAR is too high), the initial score threshold is increased, or the first and second initial preset values ​​are decreased. If there are too many false rejections (FRR is too high), the initial score threshold is decreased, or the first and second initial preset values ​​are increased. At the same time, the initial backtracking frame count is fine-tuned to see if the adjusted initial backtracking frame count makes the context information more conducive to differentiation. This process is repeated until a parameter combination that achieves an acceptable balance between the FAR and FRR is found, resulting in the preset threshold, backtracking frame count, first preset value, and second preset value.

[0037] It should also be noted that during the recognition process, the environmental signal-to-noise ratio can be monitored in real time, and the first preset value and the second preset value can be adjusted according to the environmental signal-to-noise ratio. The target command word can then be recognized based on the adjusted first preset value and the second preset value.

[0038] Additionally, it's important to note that for large lists of command words, a hierarchical recognition strategy can be employed. First, a coarse-grained score for the "first-level command word" is calculated. If the score of this first-level command word exceeds a threshold, the corresponding finer-grained "second-level command words" are then activated for further calculation and threshold determination, thereby improving the efficiency of command word recognition.

[0039] In summary, it can be seen that in the embodiments provided in this application, after identifying several scoring paths, each scoring path is backtracked and expanded to expand the phoneme sequence. This allows the capture of contextual information such as the coprolative effects with the command word, and the counting of the number of non-target phonemes as a discriminative feature. This enables the system to keenly identify those segments that, although their acoustic features are similar to the command word, are actually other words or noise. Thus, without sacrificing response speed, the accuracy of command word recognition is significantly improved.

[0040] The embodiments of this application have been described above from the perspective of the method for recognizing command words in speech signals. The embodiments of this application will now be described below from the perspective of the device for recognizing command words in speech signals.

[0041] Please see Figure 2 , Figure 2 A virtual structural diagram of a device for recognizing command words in a speech signal provided in this application is shown. The device 200 for recognizing command words in a speech signal includes: The recognition module 201 is used to input the target continuous speech signal into the speech recognition model to obtain a set of phoneme sequence paths and a path score corresponding to each phoneme sequence path in the set of phoneme sequence paths. The first determining module 202 is used to determine the output matrix sequence corresponding to each scoring path in the phoneme sequence path set when there is a scoring path set in the phoneme sequence path set where the path score is greater than a preset threshold. The backtracking module 203 is used to backtrack and expand the output matrix sequence corresponding to each scoring path to obtain the target candidate region of each scoring path. The second determining module 204 is used to determine the target phonemes of the target command word to be identified and the non-target phonemes in the target candidate region; The judgment module 205 is used to determine whether the target command word is recognized in the target candidate region based on the target phoneme and the non-target phoneme.

[0042] In one possible design, the backtracking module 203 is specifically used for: Determine the first frame, the last frame, and the number of backtracking frames corresponding to the target continuous speech signal for each scoring path in the output matrix sequence. The target candidate region for each scoring path is determined from the target continuous speech signal based on the first frame, the last frame, and the number of backtracking frames.

[0043] In one possible design, the judgment module 205 is specifically used for: The first number of non-target phonemes in the backtracked frames and the second number of non-target phonemes in the target candidate region are counted. Based on the first quantity and the second quantity, determine whether the target command word is identified in the target candidate region.

[0044] In one possible design, the judgment module 205 determines whether the target command word is identified in the target candidate region based on the first quantity and the second quantity, including: Determine whether the first quantity is greater than a first preset value, and determine whether the second quantity is greater than a second preset value; If the first quantity is greater than the first preset value or the second quantity is greater than the second preset value, then it is determined that the target candidate region has not identified the target command word; If the first quantity is less than or equal to the first preset value, and the second quantity is less than or equal to the second preset value, then it is determined that the target candidate region has identified the target command word.

[0045] In one possible design, the identification module 201 is further used for: Step 1: Randomly select a set of initialization parameters, including an initial score threshold, an initial backtracking frame count, a first initial preset value, and a second initial preset value; Step 2: Test the initialization parameters based on the dataset corresponding to the target command word to obtain test results; Step 3: Adjust the initialization parameters according to the test results, and test the adjusted initialization parameters based on the dataset; Steps 2 to 3 are executed iteratively until a preset iteration termination condition is reached, thereby obtaining the preset threshold, the number of backtracking frames, the first preset value, and the second preset value.

[0046] In one possible design, the first determining module is specifically used for: Determine the time frame range information corresponding to each phoneme in the target scoring path, wherein the target scoring path is any one of the scoring paths in the set of scoring paths; The output matrix sequence is extracted from the target continuous speech signal based on the time frame range information corresponding to each phoneme.

[0047] Reference Figure 3This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, memory, network interface, and database connected via a bus. The processor is designed to provide computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores operations, computer programs, and a database. The internal memory provides an environment for the operation of the operations and computer programs stored in the non-volatile storage medium. The database stores data such as methods for recognizing command words in speech signals. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps of a method for recognizing command words in speech signals, which includes: The target continuous speech signal is input into the speech recognition model to obtain a set of phoneme sequence paths and a path score corresponding to each phoneme sequence path in the set of phoneme sequence paths. If there is a set of scoring paths in the phoneme sequence path set whose path scores are greater than a preset threshold, then determine the output matrix sequence corresponding to each scoring path in the scoring path set. The output matrix sequence corresponding to each scoring path is backtracked and expanded to obtain the target candidate region for each scoring path; Determine the target phonemes of the target command word to be identified and the non-target phonemes in the target candidate region; Based on the target phonemes and the non-target phonemes, determine whether the target command word is identified in the target candidate region.

[0048] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements a method for recognizing command words in a speech signal. The method for recognizing command words in a speech signal includes: The target continuous speech signal is input into the speech recognition model to obtain a set of phoneme sequence paths and a path score corresponding to each phoneme sequence path in the set of phoneme sequence paths. If there is a set of scoring paths in the phoneme sequence path set whose path scores are greater than a preset threshold, then determine the output matrix sequence corresponding to each scoring path in the scoring path set. The output matrix sequence corresponding to each scoring path is backtracked and expanded to obtain the target candidate region for each scoring path; Determine the target phonemes of the target command word to be identified and the non-target phonemes in the target candidate region; Based on the target phonemes and the non-target phonemes, determine whether the target command word is identified in the target candidate region.

[0049] In some possible implementations, various aspects of the methods provided in this application can also be implemented as a program product, which can be implemented using any combination of one or more readable media. The program product includes program code that, when run on a computer device, is configured to cause the computer device to perform the steps of the methods described above according to various exemplary embodiments of this application. The computer device can perform the method for recognizing command words in a speech signal as described in the embodiments of this application, which includes: The target continuous speech signal is input into the speech recognition model to obtain a set of phoneme sequence paths and a path score corresponding to each phoneme sequence path in the set of phoneme sequence paths. If there is a set of scoring paths in the phoneme sequence path set whose path scores are greater than a preset threshold, then determine the output matrix sequence corresponding to each scoring path in the scoring path set. The output matrix sequence corresponding to each scoring path is backtracked and expanded to obtain the target candidate region for each scoring path; Determine the target phonemes of the target command word to be identified and the non-target phonemes in the target candidate region; Based on the target phonemes and the non-target phonemes, determine whether the target command word is identified in the target candidate region.

[0050] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0051] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method of recognizing a command word in a speech signal, characterized by, include: The target continuous speech signal is input into the speech recognition model to obtain a set of phoneme sequence paths and a path score corresponding to each phoneme sequence path in the set of phoneme sequence paths. If there is a set of scoring paths in the phoneme sequence path set whose path scores are greater than a preset threshold, then determine the output matrix sequence corresponding to each scoring path in the scoring path set. Determine the first frame, the last frame, and the number of backtracking frames corresponding to the target continuous speech signal for each scoring path in the output matrix sequence. Based on the first frame, the last frame, and the number of backtracking frames, the target candidate region of each scoring path is determined from the target continuous speech signal; Determine the target phonemes of the target command word to be identified and the non-target phonemes in the target candidate region; The first number of non-target phonemes in the backtracked frames and the second number of non-target phonemes in the target candidate region are counted. Determine whether the first quantity is greater than a first preset value, and determine whether the second quantity is greater than a second preset value; If the first quantity is greater than the first preset value or the second quantity is greater than the second preset value, then it is determined that the target candidate region has not identified the target command word; If the first quantity is less than or equal to the first preset value, and the second quantity is less than or equal to the second preset value, then it is determined that the target candidate region has identified the target command word.

2. The method of claim 1, wherein, The method further includes: Step 1: Randomly select a set of initialization parameters, including an initial score threshold, an initial backtracking frame count, a first initial preset value, and a second initial preset value; Step 2: Test the initialization parameters based on the dataset corresponding to the target command word to obtain test results; Step 3: Adjust the initialization parameters according to the test results, and test the adjusted initialization parameters based on the dataset; Steps 2 to 3 are executed iteratively until a preset iteration termination condition is reached, thereby obtaining the preset threshold, the number of backtracking frames, the first preset value, and the second preset value.

3. The method of any one of claims 1-2, wherein, Determining the output matrix sequence corresponding to each scoring path in the scoring path set includes: Determine the time frame range information corresponding to each phoneme in the target scoring path, wherein the target scoring path is any one of the scoring paths in the set of scoring paths; The output matrix sequence is extracted from the target continuous speech signal based on the time frame range information corresponding to each phoneme.

4. An apparatus for recognizing a command word in a speech signal for implementing the method according to any one of claims 1 to 3, characterized in that include: The recognition module is used to input the target continuous speech signal into the speech recognition model to obtain a set of phoneme sequence paths and a path score corresponding to each phoneme sequence path in the set of phoneme sequence paths. The first determining module is used to determine the output matrix sequence corresponding to each scoring path in the phoneme sequence path set when there is a scoring path set in the phoneme sequence path set where the path score is greater than a preset threshold. The backtracking module is used to backtrack and expand the output matrix sequence corresponding to each scoring path to obtain the target candidate region for each scoring path. The second determining module is used to determine the target phonemes of the target command word to be identified and the non-target phonemes in the target candidate region; The judgment module is used to determine whether the target command word is recognized in the target candidate region based on the target phoneme and the non-target phoneme.

5. A device for recognizing a command word in a speech signal, characterized by include: processor; Memory, used to store computer programs; The processor, when executing the computer program, implements the method for recognizing command words in the speech signal according to any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for recognizing command words in a speech signal according to any one of claims 1 to 3.

7. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for recognizing command words in the speech signal according to any one of claims 1 to 3.

Images