An English end-to-end speech recognition method, apparatus, device and medium

Abstract

Embodiments of the present application provide an English end-to-end speech recognition method, device, equipment and medium, the method comprises: obtaining an English speech to be recognized; inputting the English speech to be recognized into a pre-trained CTC model for recognition, to obtain an English text corresponding to the English speech; wherein the CTC model is trained by taking English speech and corresponding English text as training samples, and the English text is obtained by splitting based on the correspondence between letter combinations and pronunciations. In the embodiments of the present application, since the English text used for training the CTC model is obtained by splitting based on the correspondence between letter combinations and pronunciations, that is, the English text has a certain correspondence with the actual pronunciation, which conforms to the actual language rules, so the CTC model trained based on the English text is more accurate, thereby improving the accuracy of English speech recognition.

Description

Technical Field

[0001] This invention relates to the field of speech recognition technology, and in particular to an end-to-end English speech recognition method, apparatus, device, and medium. Background Technology

[0002] Automatic Speech Recognition (ASR) is a technology that recognizes human speech and converts it into text. It can be applied to services such as voice dialing, voice navigation, indoor device control, voice document retrieval, and simple dictation data entry.

[0003] For end-to-end speech recognition in English, the training samples of the speech recognition model are generally processed based on BPE subwords. However, BPE subwords are segmented only according to the frequency of letter combinations, without considering the correspondence between letter combinations and actual pronunciations. This makes the speech recognition model training inaccurate, resulting in low speech recognition accuracy. Summary of the Invention

[0004] In view of the above problems, embodiments of the present invention provide an English end-to-end speech recognition method, apparatus, device and medium to overcome or at least partially solve the above problems.

[0005] A first aspect of this invention discloses an end-to-end English speech recognition method, the method comprising:

[0006] Obtain the English speech to be recognized;

[0007] The English speech to be recognized is input into a pre-trained CTC model for recognition, and the corresponding English text is obtained.

[0008] The CTC model is trained using English speech and corresponding English text as training samples. The English text is segmented based on the correspondence between letter combinations and pronunciations.

[0009] Optionally, the method further includes:

[0010] The unsegmented English text is input into a pre-trained segmentation and annotation model for segmentation processing to obtain segmented English text. The training samples of the segmentation and annotation model are obtained by segmenting according to the correspondence between letter combinations and pronunciations.

[0011] Optionally, the step of inputting the unsegmented English text into a pre-trained segmentation and annotation model for segmentation includes:

[0012] The segmentation and annotation model marks the position of each letter in each word of the unsegmented English text.

[0013] Optionally, the training process of the segmentation labeling model includes:

[0014] According to the correspondence between letter combinations and pronunciations, the training words are preprocessed to generate feature vectors corresponding to the training words. The feature vectors represent the positional relationship of each letter in the training word.

[0015] The feature vector is input into the segmentation and labeling model for training to obtain the trained segmentation and labeling model. The trained segmentation and labeling model has the function of segmenting words according to the correspondence between letter combinations and pronunciations.

[0016] Optionally, the preprocessing of the training words to generate feature vectors corresponding to the training words includes:

[0017] According to the order of the letters in the training word, the training word is split into multiple letter combinations, and the position of each letter in the letter combination is marked, and a feature vector corresponding to each letter is generated.

[0018] A second aspect of the present invention discloses an English end-to-end speech recognition device, the device comprising:

[0019] The acquisition module is used to acquire the English speech to be recognized;

[0020] The recognition module is used to input the English speech to be recognized into a pre-trained CTC model for recognition, and obtain the English text corresponding to the English speech; wherein, the CTC model is trained using English speech and the corresponding English text as training samples, and the English text is segmented based on the correspondence between letter combinations and pronunciations.

[0021] Optionally, the device further includes:

[0022] The sample processing module is used to input the unsegmented English text into a pre-trained segmentation and annotation model for segmentation processing to obtain segmented English text. The training samples of the segmentation and annotation model are obtained by segmenting according to the correspondence between letter combinations and pronunciations.

[0023] Optionally, the sample processing module includes:

[0024] The tagging module is used by the segmentation and tagging model to mark the position of each letter in the unsegmented English text.

[0025] A third aspect of the present invention discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executed, implements the English end-to-end speech recognition method as described in the first aspect of the present invention.

[0026] A fourth aspect of this invention discloses a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the English end-to-end speech recognition method described in the first aspect of this invention.

[0027] The embodiments of the present invention have the following advantages:

[0028] In this embodiment of the invention, a pre-trained CTC model is used to recognize English speech to obtain the corresponding English text. Since the English text used to train the CTC model is segmented based on the correspondence between letter combinations and pronunciations, that is, the English text has a certain correspondence with the actual pronunciation, which conforms to the actual language rules. Therefore, the CTC model trained based on this English text is more accurate, thereby improving the accuracy of English speech recognition. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart of the steps of an English end-to-end speech recognition method provided in an embodiment of the present invention;

[0031] Figure 2 This is a flowchart of the steps of a segmentation and annotation model training method provided in an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the structure of an English end-to-end speech recognition device provided in an embodiment of the present invention. Detailed Implementation

[0033] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] In English speech recognition technologies, training text for speech recognition models is generally processed in three ways: single English words, single characters, and BPE sub-words. However, training texts based on single English words are large and slow, making it difficult to handle rare words or words not appearing during training. Training texts based on single characters have too fine segmentation granularity, resulting in the loss of much semantic information inherent in the words themselves. While training texts based on BPE sub-words have a segmentation granularity between the word and character levels, BPE sub-words are segmented only based on the frequency of letter combinations, without considering the correspondence between letter combinations and actual pronunciation, leading to inaccurate training of the speech recognition model.

[0035] Therefore, in order to overcome the above limitations, the applicant proposes the following technical concept: combining actual language pronunciation rules, each word in the English training text used for speech recognition model is segmented according to the correspondence between letter combinations and pronunciations, so that the segmented English training text also meets the actual language rules, thereby making the CTC model trained based on the segmented training text more accurate.

[0036] Based on the above technical concept, embodiments of the present invention provide an English end-to-end speech recognition method, such as... Figure 1 As shown, Figure 1 A flowchart of an English end-to-end speech recognition method provided in this embodiment of the invention includes the following steps:

[0037] Step S101: Obtain the English speech to be recognized.

[0038] In this embodiment, the English speech to be identified can be English speech in a network system or English speech generated by real-time human speech. English speech can be obtained in a variety of ways, such as by capturing it from the network or by collecting it through a recording device. For example, for English speech in a network system, the corresponding English speech can be obtained by capturing it from the network, and for English speech generated by real-time human speech, it can be collected through a recording device to obtain the corresponding English speech.

[0039] Since the English speech to be recognized may contain the voices of multiple speakers and environmental noise, it needs to be processed after acquisition to partially eliminate noise and the influence of different speakers, making the processed English speech more reflective of the essential characteristics of speech. In this embodiment, endpoint detection and speech enhancement are used to process the English speech to be recognized. Endpoint detection refers to distinguishing between speech and non-speech signal segments in the speech signal, accurately determining the starting point of the speech signal. After endpoint detection, subsequent processing can focus only on the speech signal, which plays an important role in improving the accuracy and recognition rate of the speech recognition model. The main task of speech enhancement is to eliminate the influence of environmental noise on English speech. In subsequent steps, semantic recognition is performed on the processed English speech, improving the accuracy of English speech recognition.

[0040] Step S102: Input the English speech to be recognized into the pre-trained CTC model for recognition to obtain the English text corresponding to the English speech; wherein, the CTC model is trained using English speech and the corresponding English text as training samples, and the English text is segmented based on the correspondence between letter combinations and pronunciations.

[0041] In this embodiment, the CTC model is used to recognize English speech. CTC (Connectionist Temporal Classification) is a method for recognizing continuous sequences. Compared with traditional speech recognition, which requires pre-alignment, CTC can directly learn from sequence data without pre-labeling the mapping relationship between the input and output sequences in the training data, achieving better results in sequence learning tasks such as speech recognition. For example, in speech recognition, the input sequence (speech) is mapped as X = [x1, x2, ..., xT], and the corresponding output sequence is Y = [y1, y2, ..., yU]. The training goal of CTC is to match X and Y as closely as possible, that is, to maximize the output probability P(Y|X). Maximizing P(Y|X) is equivalent to establishing an accurate mapping between X and Y. In addition, CTC introduces whitespace labels, which allows the network to buffer when judging the current input speech frame, solving the problem of aligning repeated characters and continuous labels.

[0042] When the CTC model receives the English speech to be recognized, it recognizes the English speech by mapping the English speech to the modeling unit (i.e., letter combination) and obtains the corresponding output English text, thus realizing the conversion of English speech into English text corresponding to the English speech and completing English speech recognition.

[0043] Furthermore, the CTC model in this embodiment is obtained through pre-training. Before using the CTC model for English recognition, it needs to be trained further. Specifically, the training method of the CTC model includes: first, acquiring the English speech to be trained and the corresponding English text, and segmenting the English text according to the correspondence between word letter combinations and pronunciations; second, inputting the English speech to be trained and the segmented English text into the CTC model for training, so that the CTC model can accurately establish the mapping relationship between the input English speech and the output English text, thereby obtaining the trained CTC model.

[0044] In this embodiment, considering the actual pronunciation rules of English, the English text used to train the CTC model is segmented according to the correspondence between letter combinations and pronunciations. That is, the trained English text has a certain correspondence with the actual pronunciation, conforming to actual language rules. Therefore, the CTC model trained based on this English text is more accurate, thus ensuring the accuracy of English speech recognition. Furthermore, word segmentation using the correspondence between letter combinations and pronunciations has a segmentation granularity between word segmentation and character segmentation. This overcomes the problems of previous methods that processed training English text based on single words, resulting in large training texts, slow training speed, and difficulty in recognizing rare words or words not seen during training in actual English speech recognition. It also solves the problem of word loss due to overly fine segmentation granularity when processing training English text based on characters.

[0045] The end-to-end English speech recognition method provided in this embodiment has high speech recognition accuracy and can be applied to commercial services with speech recognition requirements, such as speech recognition in conference scenarios and speech quality inspection, as well as scenarios such as voice dialing, voice navigation, indoor equipment control, voice document retrieval, and simple dictation data entry. For example, in a conference scenario, the terminal device collects English speech in real time, processes the collected English speech, and inputs it into the CTC model for recognition to obtain the corresponding English text content. The recognized English text content can be displayed through a display device. As another example, in a voice navigation scenario, the navigation terminal device receives the English speech of the destination, converts it into corresponding English text content, extracts destination information from the English text content, and the navigation system plans a route and performs navigation based on the destination information.

[0046] In one optional embodiment, segmenting the English text used to train the CTC model includes: inputting the unsegmented English text into a pre-trained segmentation annotation model for segmentation processing to obtain segmented English text, wherein the training samples of the segmentation annotation model are obtained by segmenting according to the correspondence between letter combinations and pronunciations.

[0047] In this embodiment, a segmentation and annotation model is used to segment the English text used to train the CTC model. This involves splitting each word in the English text according to the correspondence between letter combinations and pronunciations. This segmentation and annotation model is a word segmentation model. Unlike previous BPE sub-words that segment words based on the frequency of letter combinations, this model segments words based on the correspondence between letter combinations and pronunciations, resulting in a segmentation granularity between word-level and character-level. Furthermore, this segmentation method incorporates actual language pronunciation rules, leading to more accurate segmentation. For example, for the word "book," if "ok" also appears frequently in the training data, then the BPE sub-words will segment "book" into "bo". OK However, this classification does not correspond to the actual pronunciation of the words, because the pronunciation of the words is... oo They should be pronounced together as If it cannot be split, then a more accurate word segmentation should be divided into b. oo k.

[0048] Therefore, using a segmentation annotation model to segment the English text used to train the CTC model will result in segmentation results that are more consistent with the actual language rules of English. Consequently, when the CTC model is trained again based on the English text segmented by the segmentation annotation model, the mapping relationship between the input speech and the output English text of the CTC model will be more accurate, that is, the CTC model will be more accurate.

[0049] In one optional embodiment, the step of inputting the unsegmented English text into a pre-trained segmentation and annotation model for segmentation processing includes: the segmentation and annotation model marking the position of each letter in each word of the unsegmented English text.

[0050] In this embodiment, the segmentation and annotation model segments English text by marking the position of each letter in each word, using the marking results to characterize the word segmentation. Specifically, a corresponding position identifier is added to each letter in the word. The position identifier includes at least: a single letter identifier, a start letter identifier, a middle letter identifier, and an end letter identifier. A single letter identifier represents a letter that is pronounced alone (e.g., adding single letter identifiers to 'b' and 'k' in the word 'book'); while the start letter identifier, middle letter identifier, and end letter identifier represent letter combinations that are pronounced together and cannot be segmented (e.g., the 'ear' combination in the word 'heart' is pronounced [a:], so start letter identifiers, middle letter identifiers, and end letter identifiers are added to 'e', ​​'a', and 'r' respectively). Since a word's letter combination can be a single letter combination (pronounced alone) or a combination of multiple letters (pronounced together), the segmentation and annotation model adds corresponding position identifiers to each letter in the word according to the actual pronunciation rules of the word to achieve position marking and obtain the segmented word.

[0051] For example, the first letter of the words single, begin, middle, and end can be used as S, B, M, and E to represent single-letter identifiers, start-letter identifiers, middle-letter identifiers, and end-letter identifiers. For the word "book," based on the correspondence between letter combinations and pronunciations, it needs to be segmented into b. oo Therefore, each letter is labeled as b_S o_B o_E k_S, where b_S indicates that b is a single letter, and thus b is a separate letter combination during segmentation; o_B indicates that o is the first letter of the letter combination; and o_E indicates that o is the last letter of the letter combination. Therefore, o_B and o_E together represent the letter combination. oo Similarly, k_S indicates that k is a single letter, i.e., b_S o_B o_E. k_S indicates that the word segmentation result is b. oo k.

[0052] The segmentation annotation model is key to achieving accurate segmentation of English text. This model is pre-trained, such as... Figure 2 As shown, Figure 2 The flowchart of the training method for the segmentation and labeling model provided in this embodiment of the invention specifically includes the following steps:

[0053] Step 201: Preprocess the training words according to the correspondence between letter combinations and pronunciations to generate feature vectors corresponding to the training words. The feature vectors represent the positional relationship of each letter in the training word.

[0054] In this embodiment, the words used to train the segmentation and labeling model are split and labeled (i.e., preprocessed) according to an English pronunciation dictionary. That is, the words are split and labeled according to the correspondence between letter combinations and pronunciations to obtain corresponding feature vectors. The feature vectors refer to the feature vectors corresponding to each letter in the word.

[0055] In one optional embodiment, the preprocessing of the training words to generate the feature vectors corresponding to the training words includes: splitting the training words into multiple letter combinations according to the order of the letters in the training words, marking the position of each letter in the letter combination, and generating the feature vector corresponding to each letter.

[0056] In this embodiment, the feature vector corresponding to each letter in a word contains multiple letter position features. These letter position features include at least the preceding letter, the preceding letter, the current letter, the following letter, and the following letter. The feature vector of each letter reveals its position within the entire word and its position within letter combinations. The position within a letter combination refers to whether the letter is a standalone letter combination or its position within multiple letter combinations (i.e., the beginning, middle, and ending letter of a letter combination). Therefore, based on the word's feature vector, the specific word segmentation based on the correspondence between letter combinations and pronunciation can be determined.

[0057] For example, we can use `prepre` to represent the letter before the letter before the letter, `pre` to represent the preceding letter, `cur` to represent the current letter, `next` to represent the following letter, and `nextnext` to represent the letter after the following letter. For the word "book," based on the correspondence between letter combinations and pronunciations, it needs to be broken down into "b". oo Let k represent the position of each letter, as follows: b_S o_B o_E k_S, where b_S indicates that b is a single letter combination, o_B indicates that o is the first letter of the letter combination, o_E indicates that o is the last letter of the letter combination, and k_S indicates that k is a single letter combination. The feature vectors corresponding to the letters b, o, o, and k are represented as follows:

[0058] b_S prepre_null pre_null cur_b next_o nextnext_o means that b is a single letter, the current letter is b and b is the first letter of the word, the next letter is o, the next letter is o, and since there are no other letters before the previous letter and before the previous letter, the letter b is the first letter of the word.

[0059] o_B prepre_null pre_b cur_o next_o nextnext_k means that o is the first letter of the letter combination, the current letter is o, there is no preceding letter, the preceding letter is b, the following letter is o, and the following letter is k.

[0060] o_E prepre_b pre_o cur_o next_k nextnext_null means that o is the ending letter of the letter combination, the current letter is o, the previous letter is b, the previous letter is o, the next letter is k, and there is no next letter.

[0061] k_S prepre_o pre_o cur_k next_null nextnext_null means that k is a single letter, the current letter is k and it is the last letter of the word, the previous letter is o, the previous letter is o, and since there are no other letters after it, the letter k is the last letter of the word.

[0062] Step 202: Input the feature vector into the segmentation and labeling model for training to obtain the trained segmentation and labeling model. The trained segmentation and labeling model has the function of segmenting words according to the correspondence between letter combinations and pronunciations.

[0063] In this embodiment, since the feature vector corresponding to each word is obtained by splitting and labeling based on the correspondence between letter combinations and pronunciations, the feature vector already carries the correspondence between letter combinations and pronunciations. When training the segmentation and labeling model, the training words and their corresponding feature vectors are input together into the model for training. This allows the model to learn the word segmentation rules, i.e., to statistically analyze the correspondence between letter combinations and pronunciations. Therefore, after training based on the feature vectors, the segmentation and labeling model has the ability to segment words according to the correspondence between letter combinations and pronunciations. Compared to other word segmentation methods, this method considers the actual pronunciation of English, resulting in more accurate segmentation results that better conform to the characteristics of English pronunciation. Subsequently, when processing the training English text for the CTC speech recognition model based on this segmentation and labeling model, the segmentation is performed according to the correspondence between letter combinations and pronunciations, which also conforms to actual language rules. This results in a more accurate CTC model and significantly improves the accuracy of speech recognition.

[0064] This invention also provides an English end-to-end speech recognition device, referring to... Figure 3 , Figure 3 This is a schematic diagram of the structure of an English end-to-end speech recognition device proposed in an embodiment of this application, as shown below. Figure 3 As shown, the device includes:

[0065] The acquisition module is used to acquire the English speech to be recognized;

[0066] The recognition module is used to input the English speech to be recognized into a pre-trained CTC model for recognition, and obtain the English text corresponding to the English speech; wherein, the CTC model is trained using English speech and the corresponding English text as training samples, and the English text is segmented based on the correspondence between letter combinations and pronunciations.

[0067] In an optional embodiment, the device further includes:

[0068] The sample processing module is used to input the unsegmented English text into a pre-trained segmentation and annotation model for segmentation processing to obtain segmented English text. The training samples of the segmentation and annotation model are obtained by segmenting according to the correspondence between letter combinations and pronunciations.

[0069] In one optional embodiment, the sample processing module includes:

[0070] The tagging module is used by the segmentation and tagging model to mark the position of each letter in the unsegmented English text.

[0071] In an optional embodiment, the sample processing module further includes a segmentation and labeling model training module, the segmentation and labeling model training module comprising:

[0072] The preprocessing module is used to preprocess the training words according to the correspondence between letter combinations and pronunciations, and generate the feature vector corresponding to the training words. The feature vector represents the positional relationship of each letter in the training word.

[0073] The feature training module is used to input the feature vector into the segmentation and labeling model for training, so as to obtain the trained segmentation and labeling model. The trained segmentation and labeling model has the function of segmenting words according to the correspondence between letter combinations and pronunciations.

[0074] In one optional embodiment, the preprocessing module includes:

[0075] The splitting module is used to split the training word into multiple letter combinations according to the order of the letters in the training word, mark the position of each letter in the letter combination, and generate a feature vector corresponding to each letter.

[0076] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes an English end-to-end speech recognition method as described in this invention.

[0077] This application also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements an English end-to-end speech recognition method as described in this embodiment of the invention.

[0078] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0079] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

[0083] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0084] The present invention provides a detailed description of an English end-to-end speech recognition method, apparatus, device, and medium. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An end-to-end English speech recognition method, characterized in that, The method includes: Obtain the English speech to be recognized; The English speech to be recognized is input into a pre-trained CTC model for recognition, and the corresponding English text is obtained. The CTC model is trained using English speech and corresponding English text as training samples. The English text is segmented based on the correspondence between letter combinations and pronunciations. The correspondence between letter combinations and pronunciations is determined based on the actual pronunciation rules of English. The method further includes: The unsegmented English text is input into a pre-trained segmentation and annotation model for segmentation processing to obtain segmented English text. The training samples of the segmentation and annotation model are obtained by segmenting according to the correspondence between letter combinations and pronunciations.

2. The method according to claim 1, characterized in that, The step of inputting the unsegmented English text into a pre-trained segmentation and annotation model for segmentation includes: The segmentation and annotation model marks the position of each letter in each word of the unsegmented English text.

3. The method according to claim 1, characterized in that, The training process of the segmentation and labeling model includes: According to the correspondence between letter combinations and pronunciations, the training words are preprocessed to generate feature vectors corresponding to the training words. The feature vectors represent the positional relationship of each letter in the training word. The feature vector is input into the segmentation and labeling model for training to obtain the trained segmentation and labeling model. The trained segmentation and labeling model has the function of segmenting words according to the correspondence between letter combinations and pronunciations.

4. The method according to claim 3, characterized in that, The preprocessing of the training words to generate feature vectors corresponding to the training words includes: According to the order of the letters in the training word, the training word is split into multiple letter combinations, and the position of each letter in the letter combination is marked, and a feature vector corresponding to each letter is generated.

5. An end-to-end English speech recognition device, characterized in that, The device includes: The acquisition module is used to acquire the English speech to be recognized; The recognition module is used to input the English speech to be recognized into a pre-trained CTC model for recognition, and obtain the English text corresponding to the English speech; wherein, the CTC model is trained using English speech and the corresponding English text as training samples, and the English text is segmented based on the correspondence between letter combinations and pronunciations; the correspondence between letter combinations and pronunciations is determined based on the actual pronunciation rules of English. The device further includes: The sample processing module is used to input the unsegmented English text into a pre-trained segmentation and annotation model for segmentation processing to obtain segmented English text. The training samples of the segmentation and annotation model are obtained by segmenting according to the correspondence between letter combinations and pronunciations.

6. The apparatus according to claim 5, characterized in that, The sample processing module includes: The tagging module is used by the segmentation and tagging model to mark the position of each letter in the unsegmented English text.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes, it implements the English end-to-end speech recognition method as described in any one of claims 1 to 4.

8. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the English end-to-end speech recognition method as described in any one of claims 1 to 4.

Images