Easy-to-mix word confirmation method and device, equipment and storage medium

Abstract

The invention relates to the technical field of speech recognition, in particular to an easy-to-mix word confirmation method and device, equipment and a storage medium, and the method comprises the steps: carrying out the training on an initial speech recognition model based on a preset command word material, and obtaining an optimized recognition model; inputting a real command word material into the optimization recognition model, and extracting an embedded representation of each audio clip; constructing an embedded representation dictionary of each phoneme according to the extracted embedded representation; when a new voice input is received, generating an embedded sequence of a corresponding audio by optimizing the recognition model; on the basis of an embedded representation dictionary, obtaining a phoneme embedded representation sequence of all easy-to-mix command words in an easy-to-mix command word list corresponding to the embedded sequence; and calculating the similarity between the embedded sequence of the input audio and the phoneme embedded representation sequence of the easy-to-confuse command word to obtain a final recognition result. According to the invention, the recognition accuracy of the speech recognition system on the easily-confused command words is remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of speech recognition technology, and in particular to a method, device, equipment and storage medium for confirming easily confused words. Background Art

[0002] The current embedded speech recognition system generally adopts the CTC (Connectionist Temporal Classification) algorithm, which is widely used in resource-constrained scenarios such as smart homes and wearable devices because it does not require mandatory audio-text alignment through end-to-end training. However, the existing solution directly relies on the Top-1 result output by CTC as the final recognition result, which has significant defects: first, for command words with similar pronunciations (such as "twenty-seven degrees" and "twenty-one degrees"), it is easy to cause misrecognition due to the slight difference in CTC path probability; second, the static decision mechanism of single forward propagation lacks the ability to perform secondary verification for easily confused scenes; in addition, the phoneme-level embedded features contained in the middle layer of the model are not effectively utilized, missing the opportunity to improve the discrimination accuracy through fine-grained acoustic feature comparison.

[0003] Therefore, the existing CTC scheme is prone to misrecognition of command words with similar pronunciations, and the lack of a secondary verification mechanism leads to low recognition accuracy, which is a technical problem that needs to be solved urgently. Summary of the invention

[0004] The main purpose of this application is to provide a method, device, equipment and storage medium for confirming easily confused words, aiming to solve the technical problems that the existing CTC scheme is prone to misidentification of command words with similar pronunciations and lacks a secondary verification mechanism, resulting in low recognition accuracy.

[0005] In order to achieve the above-mentioned invention object, the present application proposes a method for confirming easily confused words, the method comprising: Based on the preset command word data, the initial speech recognition model is trained to obtain an optimized recognition model; Input the real command word data into the optimized recognition model and extract the embedding representation of each audio clip; Based on the extracted embedding representation, construct an embedding representation dictionary for each phoneme; When receiving new speech input, the corresponding audio embedding sequence is generated by optimizing the recognition model; Based on the embedding representation dictionary, obtaining the phoneme embedding representation sequences of all easily confused command words in the easily confused command word list corresponding to the embedding sequence; The similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word is calculated to obtain the final recognition result.

[0006] Furthermore, the step of training the initial speech recognition model based on the preset command word material to obtain an optimized recognition model includes: Training the speech recognition model based on the general corpus until the loss rate or word error rate on the validation set is reduced to a first preset threshold, thereby obtaining the initial speech recognition model; Training is performed on the basis of the initial speech recognition model based on the preset command word materials until the loss rate or word error rate of the initial speech recognition model on the verification set containing the command words is reduced to a second preset threshold, thereby obtaining the optimized recognition model.

[0007] Furthermore, the step of inputting the real command word material into the optimized recognition model to extract the embedded representation of each audio clip includes: Acquire preset real command word material, wherein the real command word material is an audio clip containing a command word; Inputting the real command word data into the optimized recognition model; The optimized recognition model is used to infer the input command word material to generate a feature vector corresponding to each audio clip, that is, the embedding representation.

[0008] Furthermore, the step of constructing an embedding representation dictionary for each phoneme based on the extracted embedding representation includes: For the audio segment included in the real command word material, calculating the path score based on the CTC algorithm; Based on the path score, using a backtracking algorithm to find the best alignment path between the audio and the phoneme by selecting the maximum path score; Create an initial phoneme embedding representation dictionary, where each phoneme corresponds to an initial embedding representation; When the embedding representation score of an audio clip is greater than a preset threshold, it is considered to be a high-quality audio embedding representation; According to the optimal alignment path, the high-quality audio embedding representation is mapped to the corresponding phoneme, and the embedding representation of the corresponding phoneme is updated by a moving average method to obtain the embedding representation dictionary.

[0009] Furthermore, when a new speech input is received, the step of generating an embedding sequence of the corresponding audio by optimizing the recognition model comprises: Preprocessing the voice input to obtain an audio segment to be recognized; The audio segment to be recognized is input into the optimized recognition model, and the embedded representation of the audio segment to be recognized is generated through forward propagation, that is, the embedded sequence of the corresponding audio.

[0010] Furthermore, the step of obtaining the phoneme embedding representation sequences of all easily confused command words in the easily confused command word list corresponding to the embedding sequence based on the embedding representation dictionary includes: Identify the easily confused command word list corresponding to the embedding sequence, and load all easily confused command words in the easily confused command word list; Convert the easily-confusable command words into corresponding phoneme sequences; According to each phoneme sequence, the corresponding phoneme embedding representation is obtained from a pre-constructed phoneme embedding representation dictionary to obtain the corresponding phoneme embedding representation sequence.

[0011] Furthermore, the step of calculating the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word to obtain the final recognition result includes: Align the embedding sequence of the input audio with the phoneme embedding representation sequence of the confusing command word; Using a similarity measurement method to calculate the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of each confusing command word; Compare all similarity scores and select the command word with the highest score as the final recognition result.

[0012] The second aspect of the present application provides a device for confirming easily confused words, comprising: An optimization module, used for training the initial speech recognition model based on preset command word materials to obtain an optimized recognition model; An extraction module, which is used to input the real command word data into the optimized recognition model and extract the embedding representation of each audio clip; A construction module, for constructing an embedding representation dictionary for each phoneme based on the extracted embedding representation; A generation module, for generating an embedding sequence of the corresponding audio by optimizing the recognition model when receiving a new speech input; An acquisition module, configured to acquire, based on an embedding representation dictionary, phoneme embedding representation sequences of all easily confused command words in the easily confused command word list corresponding to the embedding sequence; The calculation module is used to calculate the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word to obtain the final recognition result.

[0013] The third aspect of the present application also includes a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0014] The fourth aspect of the present application also includes a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of any of the above methods when executed by a processor.

[0015] Beneficial effects: This solution can significantly improve the recognition accuracy of easily confused command words in the speech recognition system. First, based on the CTC algorithm and the backtracking algorithm, the optimal alignment path between audio and phonemes is found to ensure the generation of high-quality phoneme embedding representations. Combined with the command word material for training, the optimized model can more accurately identify specific command words and reduce the possibility of misrecognition. Using the phoneme embedding representation dictionary, the system can efficiently process newly input audio data in practical applications, and establish a list of easily confused words that are easily confused with the currently recognized command words to perform a secondary confirmation mechanism within the easily confused words. Specifically, the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command words is calculated to accurately distinguish the easily confused command words. This series of steps not only improves the recognition accuracy, but also reduces the computational complexity and enhances the adaptability and flexibility of the system by optimizing the model and the generation process of the embedding representation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flowchart of a method for confirming easily confused words according to an embodiment of the present application is shown; Figure 2 A schematic block diagram of the structure of an easily confused word confirmation device according to an embodiment of the present application; Figure 3 A schematic block diagram of the structure of a computer device according to an embodiment of the present application.

[0017] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0019] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "a", "an", "above", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to 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 we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be an intermediate element. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and / or" used herein includes all or any module and all combinations of one or more associated listed items.

[0020] Those skilled in the art will understand that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the field to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as here.

[0021] Reference Figure 1 The embodiment of the present invention provides a method for confirming easily confused words, including steps S1-S6, specifically: S1. Training the initial speech recognition model based on the preset command word data to obtain an optimized recognition model; S2, input the real command word data into the optimized recognition model to extract the embedding representation of each audio clip; S3. Construct an embedding representation dictionary for each phoneme based on the extracted embedding representation; S4, when receiving new speech input, generating an embedding sequence of the corresponding audio by optimizing the recognition model; S5, based on the embedding representation dictionary, obtaining the phoneme embedding representation sequences of all easily confused command words in the easily confused command word list corresponding to the embedding sequence; S6. Calculate the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word to obtain the final recognition result.

[0022] As described in step S1 above, first, in the basic training stage, the speech recognition model is trained using a general corpus. General corpus usually contains a large amount of daily speech data, aiming to allow the model to learn a wide range of language features. For example, a public speech dataset (such as LibriSpeech) containing multiple languages, different accents, and background noise can be used. During the training process, the model generates prediction outputs through forward propagation and calculates the loss function (such as CTC loss, ConnectionistTemporal Classification Loss), and then adjusts the model parameters through back propagation to minimize the loss value. This process continues until the loss rate or word error rate (WER) of the model on the validation set is reduced to a preset first threshold, and an initial speech recognition model is obtained that can be used directly or for subsequent fine-tuning. This step ensures that the model has basic speech recognition capabilities and can handle general speech input.

[0023] Next is the fine-tuning phase, where preset command word corpora are introduced into the training process. Command word corpora are data sets designed specifically for specific application scenarios (such as command words such as "turn on the lights" and "turn off the lights" in smart home control). The target command words of these corpora include easily confused words (such as "twenty-seven degrees" and "twenty-one degrees"). During the fine-tuning process, the initial speech recognition model continues to be used, and the model parameters are further adjusted in combination with the command word corpora. In this way, the model is able to learn more specific command word features and continue training on the validation set containing command words until the loss rate or WER drops to a second preset threshold. This step significantly improves the model's recognition accuracy for specific command words and reduces the possibility of misrecognition.

[0024] For example, suppose we want to train a speech recognition model for a smart speaker, which contains two command words, "play music" and "pause music". First, we use general corpus to perform basic training on the model so that it can recognize general voice commands. Then, we add preset command words such as "play music" and "pause music" that are used in actual applications to the training set, and fine-tune it to obtain an optimized recognition model. In this process, the model will learn how to accurately distinguish the corresponding command words, so as to accurately respond to user commands in actual applications.

[0025] The purpose of step S1 is to ensure that the model has both broad speech recognition capabilities and can accurately recognize specific command words, especially those that are easily confused, through a two-stage training method. This optimized model improves recognition accuracy, can run stably in complex speech environments, and improves user experience.

[0026] As described in steps S2-S3 above, first, real command word data need to be prepared. These corpora should cover all target command words and be as diverse as possible to ensure that the model can learn features in various situations. For example, for the speech recognition task of a smart home control system, the corpus should contain common command words such as "turn on the lights", "turn off the lights", "turn up the temperature", "turn down the temperature", and cover different pronunciations, accents, and background noise. Next, these corpora are passed as input to the fine-tuned optimized recognition model. Before input, the audio data usually needs to be preprocessed, such as noise reduction, normalization, and segmentation, to improve the stability and accuracy of the model processing.

[0027] Then, an embedding representation of each audio clip is generated through model inference. Specifically, the forward propagation process inside the model generates a series of high-dimensional feature vectors, which represent the feature representation of the audio clip at different time steps. The embedding representation here is usually extracted from the feature layer before the output layer of the model, such as the features before the linear layer or the fully connected layer. For the CTC algorithm, the model generates a probability matrix whose time step matches the length of the audio. The maximum score path is found through the backtracking algorithm, and the optimal alignment path between the audio sequence and the phoneme is determined. Based on this optimal alignment path, the audio embedding representation corresponding to each phoneme can be extracted.

[0028] The goal of step S3 is to construct an embedding representation dictionary for each phoneme based on the extracted embedding representation. First, create an initial phoneme embedding representation dictionary. This dictionary can be initialized based on the embedding representation generated by the pre-trained model or randomly initialized. For example, suppose we have a list of common phonemes (such as "sh", "ang", "yi", etc.), we can assign an initial embedding vector to each phoneme. These initial embedding vectors can be obtained from a pre-trained language model or randomly generated high-dimensional vectors. For each input real command word material, the CTC algorithm is used to calculate the score of each possible path, and the best alignment path between the audio and the phoneme is found through the backtracking algorithm. This step ensures the accurate correspondence between the audio segment and the phoneme. A quality threshold is set, and only when the embedding representation score of the audio segment is higher than the threshold, it is considered to be a high-quality embedding representation. For example, if the embedding representation score of an audio segment is 0.9 and the threshold is set to 0.8, the embedding representation of the segment will be considered to be of high quality and used for subsequent updates. According to the best alignment path, the high-quality audio embedding representation is mapped to the corresponding phoneme, and the phoneme embedding representation is updated by the moving average method. Specifically, for each phoneme, all high-quality audio embedding representations are collected and the embedding representation of the phoneme is updated by weighted averaging. The formula example is as follows: ; where α is a weight coefficient between 0 and 1, which is used to control the speed and smoothness of the update. For example, suppose the embedding of the current phoneme "sh" is represented by E current , the newly obtained embedding is denoted as E new , then update it using the above formula. After the update is completed, save the final phoneme embedding representation dictionary and use the test set to evaluate it to check whether it can effectively reduce the recognition errors of easily confused words. If it is found that the embedding representation of some phonemes is not accurate enough or has large errors, the model parameters can be further adjusted or part of the data can be retrained to improve the overall performance.

[0029] Assume that "Twenty-seven degrees" is input into the optimized recognition model to update the corresponding content in the dictionary and build the dictionary. First, we preprocess the received speech sample in step S2 to remove background noise and normalize the volume, and then input it into the optimized speech recognition model. The model generates a corresponding embedding sequence, such as [E_1, E_2, ..., E_n], where E_i represents the embedding vector of the i-th time step. Then, in step S3, we gradually update the phoneme embedding representation dictionary based on these high-quality audio embedding representations. For example, the phoneme sequence of "Twenty-seven degrees" is ["er4", "shi2", "qi1", "du4"]. We update the high-quality embedding representation corresponding to each phoneme to the dictionary through the moving average method, where the audio is mapped from n length to m length, m is the length of the phoneme sequence and n is the length of the audio sequence. Finally, an accurate phoneme embedding representation dictionary can be obtained, which contains high-quality embedding representations of all phonemes.

[0030] The purpose of steps S2-S3 is to generate high-quality audio embedding representations that not only reflect the characteristics of the audio clips, but also align with specific phonemes, thereby providing accurate data support for subsequent steps. In this way, the system can effectively build a phoneme embedding representation dictionary and efficiently process newly input audio data in practical applications. High-quality embedding representations can be extracted from real speech samples and mapped to corresponding phonemes to build an accurate phoneme embedding representation dictionary. In this way, in subsequent steps, the system can distinguish these easily confused command words by calculating similarities, significantly improving the accuracy and reliability of recognition. Ultimately, this high-quality embedding representation provides a solid foundation for improving the performance of the entire speech recognition system.

[0031] As described in step S4 above, this step aims to convert the newly input audio data into a high-dimensional feature representation (embedded representation) for subsequent processing and identification of easily confused command words.

[0032] First, in the specific implementation process, necessary preprocessing operations need to be performed on the new voice input received. These preprocessing steps include noise reduction, normalization, and segmentation to ensure the quality and consistency of the audio data. For example, suppose a user issues a voice command of "play music" through a smart speaker. The system will first perform noise reduction on this audio signal, remove background noise, and normalize it to the same volume level. Next, the audio signal is converted into a format suitable for model input, such as a Mel-spectrogram or other commonly used audio feature representations.

[0033] Next, load the trained and fine-tuned speech recognition model. This model has been fully trained on general corpus and command word corpus, and performs well on the validation set. After loading the model, pass the preprocessed audio clip as input to the model. The model will process this data step by step according to its internal structure to generate a high-dimensional feature vector at each time step. Specifically, the model generates a series of embedding representations through forward propagation, usually extracting features before the output layer of the model (such as features before the linear layer or the fully connected layer). For example, assuming that the last layer of the model is a linear layer, the features before this layer are extracted as audio embedding representations.

[0034] To ensure that the generated embeddings are aligned with the phoneme embedding dictionary in subsequent steps, some additional processing steps may be required. For example, using the backtracking method of the CTC algorithm to align the audio clips and the phoneme sequence, or applying some smoothing techniques to reduce the impact of noise. This step helps improve the quality of the embeddings and makes the subsequent similarity calculations more accurate.

[0035] The role of step S4 is to convert the newly input audio data into high-quality embedding representations that not only reflect the characteristics of the audio clips, but also align with specific phonemes, thereby providing accurate data support for subsequent steps. In this way, the system can efficiently process newly input audio data in practical applications, and distinguish easily confused command words by calculating similarities, significantly improving the accuracy and reliability of recognition.

[0036] As described in step S5 above, first, at the beginning of step S5, the (embedded representation) embedding sequence of the input audio has been generated by step S4. Assume that the embedding sequence of the input audio is [E_1, E_2, ..., E_n], where E_i represents the embedding vector of the i-th time step.

[0037] Next, prepare a list of easily confused command words. For example, build or load a list containing easily confused command words, such as [[27 degrees, 21 degrees], [previous song, next song]]. The command words in each sublist are words that are easily misidentified. For smart home control systems, these easily confused command words may include "turn on the light" and "turn off the light", "raise the temperature" and "lower the temperature", etc.

[0038] Then, each easily confused command word is converted into a corresponding phoneme sequence. This step usually requires the use of a phoneme mapping tool (such as G2P, Grapheme-to-Phoneme converter) to complete. For example, "twenty-seven degrees" can be converted into a phoneme sequence ["er4", "shi2", "qi1", "du4"].

[0039] According to each phoneme sequence, the corresponding phoneme embedding representation is obtained from the pre-constructed phoneme embedding representation dictionary. The specific steps are as follows: For each phoneme, find its corresponding embedding vector in the embedding representation dictionary. In the order of the phoneme sequence, the embedding representation vector of each phoneme is concatenated to form a complete phoneme embedding representation sequence for the command word. For example, the phoneme embedding representation sequence of "twenty-seven degrees" is [E_{er4}, E_{shi2}, E_{qi1}, E_{du4}], where E_{x} represents the embedding representation vector of phoneme x.

[0040] For example, suppose the user issues a voice command "twenty-seven degrees". The system first pre-processes the received audio signal and generates an embedding sequence [E_1, E_2, ..., E_n] through the optimized speech recognition model. Based on the recognition results of the sequence, a list of easily confused command words that may contain "twenty-seven degrees" is identified. Next, the system prepares a list of easily confused command words [[twenty-seven degrees, twenty-one degrees], [previous song, next song]], and uses the list of easily confused command words containing "twenty-seven degrees" as the operation list for the next step of secondary comparison. Then, "twenty-seven degrees" and "twenty-one degrees" under the easily confused command word list are converted into phoneme sequences ["er4", "shi2", "qi1", "du4"] and ["er4", "yi2", "shi2", "du4"] respectively.

[0041] Then, the system obtains the phoneme embedding representation sequences of the two command words from the phoneme embedding representation dictionary: the phoneme embedding representation sequence of "twenty-seven degrees" is [E_{er4}, E_{shi2}, E_{qi1}, E_{du4}]; the phoneme embedding representation sequence of "twenty-one degrees" is [E_{er4}, E_{yi2}, E_{shi2}, E_{du4}]; in this way, the system provides an accurate phoneme embedding representation sequence for subsequent steps. In step S6, the system calculates the similarity between the embedding sequence of the input audio and the two phoneme embedding representation sequences, and finally determines whether the command word issued by the user is "twenty-seven degrees" or "twenty-one degrees".

[0042] The purpose of step S5 is to ensure that the embedding sequence of the newly input audio can be accurately matched and compared with the easily confused command words, thereby improving the recognition accuracy. By constructing the phoneme embedding representation sequence of the easily confused command words, the system can efficiently process the newly input audio data in practical applications, making it easier to distinguish the easily confused command words by calculating the similarity later.

[0043] As described in step s6 above, first, at the beginning of step S6, the embedding sequence of the input audio has been generated by step S4, and the phoneme embedding representation sequence of each command word in the easily confused command word list has been obtained by step S5. Assume that the embedding sequence of the input audio is [E_1, E_2, ..., E_n], and the phoneme embedding representation sequence of "twenty-seven degrees" is [E_{er4}, E_{shi2}, E_{qi1}, E_{du4}], and the phoneme embedding representation sequence of "twenty-one degrees" is [E_{er4}, E_{yi2}, E_{shi2}, E_{du4}].

[0044] Next, a suitable similarity metric is used to calculate the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of each confusing command word.

[0045] Align the embedding sequence of the input audio with the phoneme embedding representation sequence of each easily confused command word. This can be achieved through dynamic time warping (DTW) or other alignment algorithms. DTW can process time series data of different lengths and find the best alignment path to minimize the difference between the two sequences. According to the aligned embedding sequence, the similarity score is calculated using a similarity metric method. For example: calculate the cosine value of the angle between two vectors, ranging from -1 to 1, and the closer the value is to 1, the more similar it is. Calculate the Euclidean distance between two vectors, and the smaller the distance, the more similar it is. Compare all similarity scores and select the command word with the highest score as the final recognition result. For example, assuming that the similarity score of "twenty-seven degrees" is 0.9 and the similarity score of "twenty-one degrees" is 0.8, then "twenty-seven degrees" is selected as the final recognition result. Output the command word with the highest score as the final recognition result, and perform subsequent processing or feedback to the user as needed. Step S6 ensures the accuracy and reliability of the speech recognition system through high-quality similarity calculation. This precise similarity calculation not only improves the recognition accuracy of the system, but also enhances the adaptability and flexibility of the system, enabling the system to operate stably in complex voice environments and improve user experience. Ultimately, this method ensures the efficiency and reliability of the voice recognition system in a variety of application scenarios such as smart homes and smart terminals.

[0046] Through steps S1-S6, the present application realizes an efficient and accurate speech recognition system, which is particularly suitable for processing the recognition problem of easily confused command words. First, in step S1, the initial speech recognition model is trained based on the preset command word material to obtain an optimized recognition model to ensure that the model can accurately recognize specific command words and reduce the possibility of misrecognition. Then, in step S2, the real command word material is input into the optimized recognition model, the embedding representation of each audio clip is extracted, and a high-quality audio embedding representation is generated, which provides a solid foundation for the subsequent construction of a phoneme embedding representation dictionary. In step S3, according to the extracted embedding representation, an embedding representation dictionary of each phoneme is constructed, the correspondence between audio and phoneme is found through the optimal alignment path, and the embedding representation of the same phoneme is averaged to generate a high-quality phoneme embedding representation dictionary. In step S4, when a new voice input is received, an embedding sequence of the corresponding audio is generated by optimizing the recognition model to ensure that the new input audio can be converted into a high-quality embedding representation. In step S5, based on the embedding representation dictionary, the phoneme embedding representation sequence of all easily confused command words in the easily confused command word list corresponding to the embedding sequence of the input audio is obtained, and data for similarity calculation is prepared. Finally, in step S6, the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command words is calculated, and the final recognition result is determined by cosine similarity, so as to accurately distinguish the easily confused command words. This series of steps significantly improves the accuracy, adaptability and flexibility of the speech recognition system, enabling the system to operate stably in complex speech environments, improve user experience, and be widely used in various application scenarios such as smart homes and smart terminals.

[0047] In one embodiment, the step of training the initial speech recognition model based on the preset command word material to obtain an optimized recognition model includes: S10, training the speech recognition model based on the general corpus until the loss rate or word error rate on the validation set is reduced to a first preset threshold, thereby obtaining the initial speech recognition model; S11. Perform training on the basis of the initial speech recognition model based on preset command word materials until the loss rate or word error rate of the initial speech recognition model on the verification set containing command words is reduced to a second preset threshold, thereby obtaining the optimized recognition model.

[0048] In this embodiment, the process of training the initial speech recognition model based on the preset command word data to obtain the optimized recognition model is first performed by using the general corpus to perform basic training on the speech recognition model until the loss rate or word error rate of the model on the validation set is reduced to a preset first threshold, thereby obtaining the initial speech recognition model. The goal of this stage is to allow the model to learn a wide range of language features and have the basic ability to handle a variety of voice inputs. General corpus usually contains a large amount of daily speech data, covering different languages, accents and background noise, to ensure that the model has strong generalization ability. For example, the public LibriSpeech dataset can be used, which contains a large number of English speech clips to help the model learn basic speech features.

[0049] Next, after obtaining the initial speech recognition model, it is further fine-tuned using specific command word data. These command word data are specifically designed for specific command words in application scenarios, such as "turn on the light" and "turn off the light" in smart home control. In this way, the model can learn more specific command word features and significantly improve the recognition accuracy of these command words. During the fine-tuning process, the results of the initial model training on the general corpus are continued as a starting point, and the model parameters are further adjusted in combination with the command word data. Specifically, the command word data is input into the model, the score of each path is calculated, and the backtracking algorithm is used to find the best alignment path between audio and phonemes to generate high-quality embedding representations. This process continues until the loss rate or word error rate of the model on the validation set containing command words is reduced to a second preset threshold. This step significantly improves the recognition accuracy of the model for specific command words and reduces the possibility of misrecognition.

[0050] Through this two-stage training method, the system not only has a wide range of speech recognition capabilities, but also can accurately recognize specific command words, especially those that are easily confused. This method provides a solid foundation for subsequent steps (such as extracting embedded representations, building phoneme embedding representation dictionaries, etc.), and ultimately realizes an efficient and accurate speech recognition system. The system can operate stably in complex voice environments, improve user experience, and is widely used in various application scenarios such as smart homes and smart terminals. In this way, the system can efficiently process newly input audio data in practical applications, significantly improving the accuracy and reliability of recognition.

[0051] In one embodiment, the step of inputting the real command word material into the optimized recognition model and extracting the embedded representation of each audio segment includes: S20, obtaining preset real command word material, wherein the real command word material is an audio clip containing a command word; S21, inputting the real command word data into the optimized recognition model; S22. Use the optimized recognition model to infer the input command word material to generate a feature vector corresponding to each audio clip, that is, the embedded representation.

[0052] In this embodiment, the process of inputting real command word data into the optimized recognition model to extract the embedded representation of each audio clip first requires obtaining preset real command word data. These corpora contain command word audio clips in actual application scenarios, including collected corpora and corpora synthesized based on actual applications; the input command word data is inferred using the optimized recognition model to generate feature vectors corresponding to each audio clip, i.e., embedded representations. Specifically, the model generates a series of high-dimensional feature vectors through forward propagation, which represent the feature representations of audio clips at different time steps. In order to ensure high-quality embedded representations, the system calculates the score of each path and finds the best alignment path between audio and phonemes through a backtracking algorithm. Based on this best alignment path, the audio embedding representation corresponding to each phoneme can be extracted. For each audio clip, the probability matrix whose time step generated by the model matches the audio length is used to determine the best alignment path, thereby generating high-quality embedded representations.

[0053] In addition, to further improve the quality of the embedded representation, the system also sets a quality threshold. Only when the embedding representation score of the audio clip is higher than this threshold, it is considered to be a high-quality embedded representation. This step ensures that only those embedding representations with higher confidence are used to subsequently update the phoneme embedding representation dictionary. In this way, the system can efficiently and accurately process newly input audio data and provide a solid foundation for the subsequent construction of the phoneme embedding representation dictionary.

[0054] In one embodiment, the step of constructing an embedding representation dictionary for each phoneme based on the extracted embedding representation includes: S30, calculating a path score for the audio segment included in the real command word material based on a CTC algorithm; S31, based on the path score, using a backtracking algorithm to find the best alignment path between the audio and the phoneme by selecting the maximum path score; S32, creating an initial phoneme embedding representation dictionary, where each phoneme corresponds to an initial embedding representation; S33, when the embedding representation score of the audio clip is greater than a preset threshold, it is identified as a high-quality audio embedding representation; S34. According to the optimal alignment path, the high-quality audio embedding representation is mapped to the corresponding phoneme, and the embedding representation of the corresponding phoneme is updated by a moving average method to obtain the embedding representation dictionary.

[0055] In this embodiment, the process of constructing the embedding representation dictionary of each phoneme based on the extracted embedding representation first requires processing the audio clips in the real command word material. Specifically, for these audio clips containing command words, the path score is calculated based on the CTC (Connectionist Temporal Classification) algorithm. The CTC algorithm can handle the alignment problem of variable-length sequences and is particularly suitable for audio and phoneme alignment in speech recognition tasks. Next, based on the calculated path score, a backtracking algorithm is used to find the best alignment path for audio and phonemes by selecting the maximum path score. The best alignment path refers to the path with the highest score among all possible paths, representing the most likely correspondence between the audio sequence and the phoneme. For example, assuming that the best alignment path for a certain audio clip is ["er4", "shi2", "qi1"], the correspondence between the audio clip and the phoneme is determined.

[0056] Then, create an initial dictionary of phoneme embedding representations, where each phoneme corresponds to an initial embedding representation. This initial dictionary can be initialized based on the embedding representations generated by a pre-trained model or randomly initialized. For example, suppose we have a list of common phonemes (such as "sh", "ang", "yi", etc.), we can assign an initial embedding vector to each phoneme. These initial embedding vectors can be obtained from a pre-trained language model or randomly generated high-dimensional vectors.

[0057] To ensure high-quality embedding representations, the system sets a quality threshold, and only when the embedding representation score of an audio clip is greater than the threshold, it is considered a high-quality audio embedding representation. This step ensures that only those embedding representations with high confidence are used for subsequent updates of the phoneme embedding representation dictionary. For example, if the embedding representation score of an audio clip is 0.9 and the threshold is set to 0.8, the embedding representation of the clip will be considered high-quality and used for subsequent updates.

[0058] In this way, the system can efficiently and accurately build a high-quality phoneme embedding representation dictionary, providing a solid foundation for subsequent steps. This method not only improves the recognition accuracy of the system, but also enhances the adaptability and flexibility of the system, enabling the system to operate stably in complex voice environments, improve user experience, and be widely used in smart homes, smart terminals and other application scenarios. Ultimately, this high-quality phoneme embedding representation ensures the efficiency and reliability of the speech recognition system in practical applications.

[0059] In one embodiment, when a new speech input is received, the step of generating an embedding sequence of the corresponding audio by optimizing the recognition model comprises: S40, preprocessing the voice input to obtain an audio segment to be recognized; S41, input the audio segment to be recognized into the optimized recognition model, and generate an embedded representation of the audio segment to be recognized through forward propagation, that is, an embedded sequence of the corresponding audio.

[0060] In this embodiment, first, the received new voice input is preprocessed to obtain the audio segment to be recognized. The preprocessing step generally includes operations such as noise reduction, normalization and segmentation to ensure the quality and consistency of the audio data. Next, the preprocessed audio segment to be recognized is input into the optimized recognition model that has been trained and fine-tuned. This optimized recognition model has been fully trained on general corpus and command word corpus, and performs well on the validation set. After loading the model, the audio segment to be recognized is passed to the model as input. The model will process these data step by step according to its internal structure, and generate a series of high-dimensional feature vectors through forward propagation, which represent the feature representation of the audio segment at different time steps. Specifically, the model generates an embedded representation at each time step through forward propagation, and usually extracts the features before the output layer of the model (such as the features before the linear layer or the fully connected layer). In order to ensure that the generated embedded representation can be aligned with the phoneme embedding representation dictionary in the subsequent steps, some additional processing steps may be required. For example, the audio segment and the phoneme sequence are aligned using the backtracking method of the CTC algorithm, or a smoothing technique is applied to reduce the impact of noise. This step helps to improve the quality of the embedding representation, making the subsequent similarity calculation more accurate. Finally, the embedding sequence generated by the model can be expressed as [E_1, E_2, ..., E_n], where E_i represents the embedding vector of the i-th time step. In this way, the system can efficiently and accurately process the newly input audio data and generate high-quality embedding representations. These embedding representations not only reflect the characteristics of the audio clips, but also align with specific phonemes, thereby providing accurate data support for subsequent steps. For example, in the subsequent steps, the system can accurately distinguish these easily confused command words by calculating the similarity between the embedding sequence of the newly input audio and the phoneme embedding representation sequence of the easily confused command words, significantly improving the accuracy and reliability of recognition. In short, by preprocessing the received new voice input and generating high-quality embedding sequences using the optimized recognition model, the system can efficiently process various complex voice inputs in practical applications. This method not only improves the recognition accuracy of the system, but also enhances the adaptability and flexibility of the system, enabling the system to operate stably in complex voice environments, improve user experience, and is widely used in various application scenarios such as smart homes and smart terminals. Ultimately, this high-quality embedding representation ensures the efficiency and reliability of speech recognition systems in practical applications.

[0061] In one embodiment, the step of obtaining the phoneme embedding representation sequences of all easily confused command words in the easily confused command word list corresponding to the embedding sequence based on the embedding representation dictionary includes: S50, identifying a list of easily confused command words corresponding to the embedded sequence, and loading all easily confused command words in the list of easily confused command words; S501 converts the easily-confusable command words into corresponding phoneme sequences; S52. According to each phoneme sequence, a corresponding phoneme embedding representation is obtained from a pre-constructed phoneme embedding representation dictionary to obtain a corresponding phoneme embedding representation sequence.

[0062] In this embodiment, first, a list of easily confused command words corresponding to the embedding sequence is identified, and all easily confused command words in the list are loaded. This step involves determining a set of easily confused command words that the current input audio may correspond to. For example, in a smart home control system, common easily confused command words may include "turn on the lights" and "turn off the lights", "raise the temperature" and "lower the temperature", etc. By analyzing the context of the input audio or preset rules, the system can identify a set of potential easily confused command words. Next, the easily confused command words are converted into corresponding phoneme sequences. This process usually requires the use of a phoneme mapping tool (such as G2P, Grapheme-to-Phoneme converter) to complete. For example, "twenty-seven degrees" can be converted into a phoneme sequence ["er4", "shi2", "qi1", "du4"], and "twenty-one degrees" can be converted into a phoneme sequence ["er4", "yi2", "shi2", "du4"]. The phoneme sequence provides a pronunciation feature representation of each command word, so that the subsequent steps can perform similarity calculations by comparing the phoneme embedding representations. Then, according to each phoneme sequence, the corresponding phoneme embedding representation is obtained from the pre-constructed phoneme embedding representation dictionary to obtain the corresponding phoneme embedding representation sequence. The specific steps are as follows: For each phoneme, find its corresponding embedding vector in the embedding representation dictionary. For example, suppose that the phoneme embedding representation dictionary contains the embedding vector E_{er4} of the phoneme "er4", the embedding vector E_{shi2} of the phoneme "shi2", etc. In the order of the phoneme sequence, the embedding representation vectors of each phoneme are concatenated to form a complete phoneme embedding representation sequence for the command word. For example, the phoneme embedding representation sequence of "twenty-seven degrees" is [E_{er4}, E_{shi2}, E_{qi1}, E_{du4}], while the phoneme embedding representation sequence of "twenty-one degrees" is [E_{er4}, E_{yi2}, E_{shi2}, E_{du4}]. In this way, the system provides an accurate phoneme embedding representation sequence for subsequent steps. These phoneme embedding representation sequences will be used in the next step to calculate similarity scores to distinguish easily confused command words. For example, assuming that the user issues a voice command "twenty-seven degrees", the system first pre-processes the received audio signal and generates an embedding sequence [E_1, E_2, ..., E_n] through the optimized speech recognition model. Then, the system identifies a list of potential easily confused command words [[twenty-seven degrees, twenty-one degrees], [previous song, next song]] and converts "twenty-seven degrees" and "twenty-one degrees" into phoneme sequences ["er4", "shi2", "qi1", "du4"] and ["er4", "yi2", "shi2", "du4"] respectively.Then, the system obtains the phoneme embedding representation sequences of the two command words from the phoneme embedding representation dictionary: the phoneme embedding representation sequence of "twenty-seven degrees" is [E_{er4}, E_{shi2}, E_{qi1}, E_{du4}]; the phoneme embedding representation sequence of "twenty-one degrees" is [E_{er4},E_{yi2}, E_{shi2}, E_{du4}]; these phoneme embedding representation sequences will be used for similarity calculation in subsequent steps. By calculating the similarity between the embedding sequence of the input audio and these phoneme embedding representation sequences, the system can accurately identify whether the command word issued by the user is "twenty-seven degrees" or "twenty-one degrees".

[0063] In summary, by identifying the list of easily confused command words corresponding to the embedding sequence, converting these command words into phoneme sequences, and then obtaining the corresponding phoneme embedding representation sequence from the embedding representation dictionary, the system can efficiently process new input audio data in practical applications and distinguish easily confused command words by calculating similarity. This method significantly improves the accuracy, adaptability and flexibility of the speech recognition system, enabling the system to operate stably in complex voice environments, improve user experience, and be widely used in a variety of application scenarios such as smart homes and smart terminals. Ultimately, this high-quality phoneme embedding representation ensures the efficiency and reliability of the speech recognition system in practical applications.

[0064] In one embodiment, the step of calculating the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word to obtain the final recognition result includes: S61, aligning the embedding sequence of the input audio with the phoneme embedding representation sequence of the easily confusing command word; S62, using a similarity measurement method to calculate the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of each easily confused command word; S63. Compare all similarity scores and select the command word with the highest score as the final recognition result.

[0065] In this embodiment, the process of calculating the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word to obtain the final recognition result is as follows: First, the embedding sequence of the input audio is aligned with the phoneme embedding representation sequence of each easily confused command word using the dynamic time warping (DTW) algorithm to ensure that the two are compared at the same time step. Then, a similarity measurement method such as cosine similarity or Euclidean distance is used to calculate the similarity score of the two. For example, assuming that the embedding sequence of the input audio is [E_1, E_2, ..., E_n], it is compared with the phoneme embedding representation sequence of "twenty-seven degrees" and "twenty-one degrees" respectively to obtain their respective similarity scores. Finally, the system compares all similarity scores and selects the command word with the highest score as the final recognition result. In this way, the system can efficiently and accurately process the newly input audio data and distinguish easily confused command words, significantly improving the accuracy, adaptability and user experience of the speech recognition system.

[0066] Reference Figure 2 , is a structural block diagram of an easily confused word confirmation device in an embodiment of the present application, the device comprises: The optimization module 100 is used to train the initial speech recognition model based on the preset command word material to obtain an optimized recognition model; An extraction module 200, used to input the real command word material into the optimized recognition model and extract the embedded representation of each audio segment; A construction module 300, for constructing an embedding representation dictionary for each phoneme according to the extracted embedding representation; A generation module 400, configured to generate an embedding sequence of the corresponding audio by optimizing the recognition model when receiving a new speech input; An acquisition module 500 is used to acquire, based on an embedding representation dictionary, phoneme embedding representation sequences of all easily confused command words in the easily confused command word list corresponding to the embedding sequence; The calculation module 600 is used to calculate the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word to obtain the final recognition result.

[0067] In one embodiment, the optimization module 100 includes a training unit for: Training the speech recognition model based on the general corpus until the loss rate or word error rate on the validation set is reduced to a first preset threshold, thereby obtaining the initial speech recognition model; Training is performed on the basis of the initial speech recognition model based on the preset command word materials until the loss rate or word error rate of the initial speech recognition model on the verification set containing the command words is reduced to a second preset threshold, thereby obtaining the optimized recognition model.

[0068] In one embodiment, the extraction module 200 includes an inference unit, which is used to: Acquire preset real command word material, wherein the real command word material is an audio clip containing a command word; Inputting the real command word data into the optimized recognition model; The optimized recognition model is used to infer the input command word material to generate a feature vector corresponding to each audio clip, that is, the embedding representation.

[0069] In one embodiment, the building module 300 includes an updating unit, which is used to: For the audio segment included in the real command word material, calculating the path score based on the CTC algorithm; Based on the path score, using a backtracking algorithm to find the best alignment path between the audio and the phoneme by selecting the maximum path score; Create an initial phoneme embedding representation dictionary, where each phoneme corresponds to an initial embedding representation; When the embedding representation score of an audio clip is greater than a preset threshold, it is considered to be a high-quality audio embedding representation; According to the optimal alignment path, the high-quality audio embedding representation is mapped to the corresponding phoneme, and the embedding representation of the corresponding phoneme is updated by a moving average method to obtain the embedding representation dictionary.

[0070] In one embodiment, the generation module 400 includes a sequence generation unit, which is used to: Preprocessing the voice input to obtain an audio segment to be recognized; The audio segment to be recognized is input into the optimized recognition model, and the embedded representation of the audio segment to be recognized is generated through forward propagation, that is, the embedded sequence of the corresponding audio.

[0071] In one embodiment, the acquisition module 500 includes a phoneme embedding acquisition unit, which is used to: Identify the easily confused command word list corresponding to the embedding sequence, and load all easily confused command words in the easily confused command word list; Convert the easily-confusable command words into corresponding phoneme sequences; According to each phoneme sequence, the corresponding phoneme embedding representation is obtained from a pre-constructed phoneme embedding representation dictionary to obtain the corresponding phoneme embedding representation sequence.

[0072] In one embodiment, the calculation module 600 includes a result output unit for: Align the embedding sequence of the input audio with the phoneme embedding representation sequence of the confusing command word; Using a similarity measurement method to calculate the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of each confusing command word; Compare all similarity scores and select the command word with the highest score as the final recognition result.

[0073] Reference Figure 3 In an embodiment of the present application, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store usage data in the process of the easily confused word confirmation method, etc. The network interface of the computer device is used to communicate with an external terminal through a network connection. Furthermore, the above-mentioned computer device can also be provided with an input device and a display screen, etc. When the above computer program is executed by the processor to implement the easily confused word confirmation method, the method includes the following steps: training on the initial speech recognition model based on the preset command word material to obtain an optimized recognition model; inputting the real command word material into the optimized recognition model to extract the embedding representation of each audio segment; constructing an embedding representation dictionary for each phoneme based on the extracted embedding representation; when receiving new speech input, generating an embedding sequence of the corresponding audio through the optimized recognition model; based on the embedding representation dictionary, obtaining the phoneme embedding representation sequence of all easily confused command words in the easily confused command word list corresponding to the embedding sequence; calculating the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word to obtain the final recognition result. Those skilled in the art can understand that Figure 3 The structure shown in is merely a block diagram of a portion of the structure related to the present application solution and does not constitute a limitation on the computer device to which the present application solution is applied.

[0074] An embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, a method for confirming easily confused words is implemented, including the following steps: training on an initial speech recognition model based on a preset command word material to obtain an optimized recognition model; inputting the real command word material into the optimized recognition model to extract the embedding representation of each audio segment; constructing an embedding representation dictionary for each phoneme based on the extracted embedding representation; when receiving a new voice input, generating an embedding sequence of the corresponding audio through the optimized recognition model; based on the embedding representation dictionary, obtaining the phoneme embedding representation sequence of all easily confused command words in the easily confused command word list corresponding to the embedding sequence; calculating the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word, and obtaining the final recognition result. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0075] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0076] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the existence of other identical elements in the process, device, article or method including the element.

[0077] The above description is only a preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for confirming easily confused words, characterized in that: The method comprises: Based on the preset command word data, the initial speech recognition model is trained to obtain an optimized recognition model; Input the real command word data into the optimized recognition model and extract the embedding representation of each audio clip; Based on the extracted embedding representation, construct an embedding representation dictionary for each phoneme; When receiving new speech input, the corresponding audio embedding sequence is generated by optimizing the recognition model; Based on the embedding representation dictionary, obtaining the phoneme embedding representation sequences of all easily confused command words in the easily confused command word list corresponding to the embedding sequence; The similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word is calculated to obtain the final recognition result.

2. The method for confirming easily confused words according to claim 1, characterized in that: The step of training the initial speech recognition model based on the preset command word material to obtain an optimized recognition model includes: Training the speech recognition model based on the general corpus until the loss rate or word error rate on the validation set is reduced to a first preset threshold, thereby obtaining the initial speech recognition model; Training is performed on the basis of the initial speech recognition model based on the preset command word materials until the loss rate or word error rate of the initial speech recognition model on the verification set containing the command words is reduced to a second preset threshold, thereby obtaining the optimized recognition model.

3. The method for confirming easily confused words according to claim 1, characterized in that: The step of inputting the real command word material into the optimized recognition model and extracting the embedded representation of each audio clip includes: Acquire preset real command word material, wherein the real command word material is an audio clip containing a command word; Inputting the real command word data into the optimized recognition model; The optimized recognition model is used to infer the input command word material to generate a feature vector corresponding to each audio clip, that is, the embedding representation.

4. The method for confirming easily confused words according to claim 1, characterized in that: The step of constructing an embedding representation dictionary for each phoneme according to the extracted embedding representation comprises: For the audio segment included in the real command word material, calculating the path score based on the CTC algorithm; Based on the path score, using a backtracking algorithm to find the best alignment path between the audio and the phoneme by selecting the maximum path score; Create an initial phoneme embedding representation dictionary, where each phoneme corresponds to an initial embedding representation; When the embedding representation score of an audio clip is greater than a preset threshold, it is considered to be a high-quality audio embedding representation; According to the optimal alignment path, the high-quality audio embedding representation is mapped to the corresponding phoneme, and the embedding representation of the corresponding phoneme is updated by a moving average method to obtain the embedding representation dictionary.

5. The method for confirming easily confused words according to claim 1, characterized in that: The step of generating an embedding sequence of the corresponding audio by optimizing the recognition model when a new voice input is received comprises: Preprocessing the voice input to obtain an audio segment to be recognized; The audio segment to be recognized is input into the optimized recognition model, and the embedded representation of the audio segment to be recognized is generated through forward propagation, that is, the embedded sequence of the corresponding audio.

6. The method for confirming easily confused words according to claim 1, characterized in that: The step of obtaining the phoneme embedding representation sequences of all easily confused command words in the easily confused command word list corresponding to the embedding sequence based on the embedding representation dictionary comprises: Identify the easily confused command word list corresponding to the embedding sequence, and load all easily confused command words in the easily confused command word list; Convert the easily-confusable command words into corresponding phoneme sequences; According to each phoneme sequence, the corresponding phoneme embedding representation is obtained from a pre-constructed phoneme embedding representation dictionary to obtain the corresponding phoneme embedding representation sequence.

7. The method for confirming easily confused words according to claim 1, characterized in that: The step of calculating the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word to obtain the final recognition result includes: Align the embedding sequence of the input audio with the phoneme embedding representation sequence of the confusing command word; Using a similarity measurement method to calculate the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of each confusing command word; Compare all similarity scores and select the command word with the highest score as the final recognition result.

8. A device for confirming easily confused words, characterized in that: include: An optimization module, used for training the initial speech recognition model based on preset command word materials to obtain an optimized recognition model; An extraction module, which is used to input the real command word data into the optimized recognition model and extract the embedding representation of each audio clip; A construction module, for constructing an embedding representation dictionary for each phoneme based on the extracted embedding representation; A generation module, for generating an embedding sequence of the corresponding audio by optimizing the recognition model when receiving a new speech input; An acquisition module, configured to acquire, based on an embedding representation dictionary, phoneme embedding representation sequences of all easily confused command words in the easily confused command word list corresponding to the embedding sequence; The calculation module is used to calculate the similarity between the embedding sequence of the input audio and the phoneme embedding representation sequence of the easily confused command word to obtain the final recognition result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images