CN114974249A - Voice recognition method, device and storage medium - Google Patents

Abstract

The embodiment of the application provides a voice recognition method, a voice recognition device and a storage medium, wherein the method comprises the following steps: the method comprises the steps that an edge end obtains a cloud end identification result, wherein the cloud end identification result comprises at least one target sequence which is obtained by the cloud end aiming at target identification objects; and correcting the at least one target sequence based on the local reference text corresponding to the target recognition object to obtain an edge end recognition result. According to the method and the device, the recognition result of the cloud to the target recognition object is obtained through the edge, the cloud recognition result is corrected based on the local reference text corresponding to the target recognition object, the recognition result of the cloud is optimized, and the accuracy of voice recognition is improved.

Description

A kind of speech recognition method, device and storage medium

technical field

The present application relates to the field of computer technology, and in particular, to a speech recognition method and device.

Background technique

With the development of the Internet, speech recognition has played an increasingly important role. Automatic Speech Recognition (ASR) is a technology that enables machines to convert speech signals into corresponding text or commands through the process of recognition and understanding. Among them, Large Vocabulary Continuous Speech Recognition (LVCSR) technology has developed rapidly in recent years and has been widely used in many fields.

In the prior art, a speech recognition engine deployed in the cloud is usually used, and the corpus for training the language model generally adopts the corpus in the general field. However, due to the limited amount of data, it is impossible to cover all areas. When performing speech recognition tasks in specific fields, such as speech recognition tasks in medicine, architecture, artificial intelligence, etc., because the language model does not cover enough in these fields, or there are many words that do not appear in the dictionary (Out Of Vocabulary, OOV), which will lead to the degradation of ASR recognition performance and the reduction of recognition accuracy.

SUMMARY OF THE INVENTION

Embodiments of the present application provide a speech recognition method, device, and storage medium, so as to solve the defect that speech cannot be accurately recognized in the prior art, and improve the accuracy of speech recognition.

In a first aspect, an embodiment of the present application provides a speech recognition method, including:

The edge terminal obtains the cloud recognition result, and the cloud recognition result includes at least one target sequence obtained by the cloud for the target recognition object recognition; based on the local reference text corresponding to the target recognition object, the at least one target sequence is modified to obtain the edge end recognition result.

Optionally, according to a speech recognition method according to an embodiment of the present application, the at least one target sequence is modified based on the local reference text corresponding to the target recognition object to obtain an edge recognition result, including:

Based on the language model corresponding to the target sequence and the local reference text, vocabulary replacement is performed on a candidate sequence in the target sequence; wherein, the language model corresponding to the target sequence is obtained by training based on the local reference text of;

and / or,

Based on named entity recognition NER, lexical replacement is performed on one of the candidate sequences in the target sequence.

Optionally, according to a speech recognition method according to an embodiment of the present application, the vocabulary matching probability of the candidate sequence is the highest among the vocabulary matching probabilities of all the target sequences;

Wherein, for each of the target sequences, the vocabulary matching probability is calculated based on the language model corresponding to the target sequence and the target sequence;

The word matching probability of the target sequence is used to describe the frequency of words in the target sequence appearing in the local reference text.

Optionally, according to a speech recognition method according to an embodiment of the present application, the vocabulary replacement is performed on an alternative sequence in the target sequence based on the language model corresponding to the target sequence and the local reference text, include:

Based on the language model corresponding to the target sequence, obtain the continuous matching probability corresponding to the continuous word combination in the candidate sequence, wherein the continuous matching probability is used to indicate that the continuous matching probability corresponding to the continuous word combination is in the the frequency of occurrences in the local reference text;

If the continuous matching probability corresponding to any continuous word combination in the candidate sequence is lower than the first preset threshold, the continuous matching probability is replaced by the replacement text in the local reference text the first consecutive word combination of the preset threshold;

Wherein, the phoneme matching degree of the phoneme sequence corresponding to the replacement text and the phoneme sequence of the first continuous vocabulary combination is greater than a second preset threshold, and the replacement text is higher than the first continuous vocabulary combination in the local reference Occurs more frequently in text.

Optionally, according to a speech recognition method according to an embodiment of the present application, the target sequence includes: a phoneme sequence, and/or a word sequence;

Correspondingly, the language model corresponding to the phoneme sequence includes a phoneme sequence language model; the language model corresponding to the word sequence includes a word sequence language model.

Optionally, according to a speech recognition method according to an embodiment of the present application, if the target sequence includes a phoneme sequence, the target sequence is based on the language model corresponding to the target sequence and the local reference text. An alternative sequence of lexical replacements, including:

Based on the phoneme sequence language model, vocabulary replacement is performed on a candidate sequence in the phoneme sequence to obtain a first phoneme sequence;

Based on the word sequence language model, vocabulary replacement is performed on a candidate sequence in the at least one word sequence corresponding to the first phoneme sequence.

Optionally, according to a speech recognition method according to an embodiment of the present application, if the target sequence further includes a word sequence, the target sequence is determined based on the language model corresponding to the target sequence and the local reference text. lexical replacement with an alternative sequence in , which also includes:

If the fourth matching probability corresponding to the first word sequence is greater than the fourth matching probability corresponding to the second word sequence, then determining that the first word sequence is the edge recognition result;

If the fourth matching probability corresponding to the second word sequence is greater than the fourth matching probability corresponding to the first word sequence, then determining that the second word sequence is the edge recognition result;

Wherein, the first word sequence is obtained after performing vocabulary replacement on a candidate sequence in at least one word sequence corresponding to the first phoneme sequence;

The second word sequence is obtained by performing lexical replacement on a candidate sequence in the word sequence in the target sequence based on the word sequence language model;

The fourth matching probability corresponding to the first word sequence is used to describe the degree of matching between the first word sequence and the local reference text;

The fourth matching probability corresponding to the second word sequence is used to describe the degree of matching between the second word sequence and the local reference text.

Optionally, according to a speech recognition method according to an embodiment of the present application, the NER-based named entity recognition, performing vocabulary replacement on an alternative sequence in the target sequence, and obtaining the edge recognition result, including:

Based on the NER vocabulary, replace the vocabulary in a candidate sequence in the target sequence by the first replacement vocabulary in the NER vocabulary to obtain the edge recognition result;

Wherein, the phoneme matching probability of the phoneme sequence corresponding to the replacement vocabulary and the phoneme sequence of the vocabulary in the candidate sequence is greater than a third preset threshold.

Optionally, according to a speech recognition method according to an embodiment of the present application, the method further includes:

Identifying the local reference text based on the NER, and generating an NER vocabulary;

Based on the dictionary and/or the character-to-phoneme G2P technology, the phonemes corresponding to the NER vocabulary are obtained.

Optionally, according to a speech recognition method according to an embodiment of the present application, the local reference text includes: professional information of the activity to which the local reference text belongs, participant information, activity background information, and activity content.

Optionally, according to a speech recognition method according to an embodiment of the present application, after the edge end recognition result is obtained, the method further includes:

Based on the professional information of the activity to which the local reference text belongs, the relevant information of the local reference text is saved to the server.

In a second aspect, an embodiment of the present application provides a speech recognition method, including:

The device end obtains the cloud identification result or the edge identification result, wherein the cloud identification result includes a target sequence obtained by the cloud identifying the target identification object;

Based on the local vocabulary on the device side, the one target sequence is modified to obtain the device-side recognition result.

Optionally, according to a speech recognition method according to an embodiment of the present application, modifying the target sequence based on a local vocabulary on the device side to obtain a device-side recognition result, including:

Based on the second replacement vocabulary in the local vocabulary, replace the vocabulary corresponding to the second replacement vocabulary in the target sequence, and obtain the device-side recognition result;

Wherein, the phoneme matching probability of the phoneme sequence of the vocabulary corresponding to the second replacement vocabulary and the phoneme sequence of the second replacement vocabulary is greater than a fourth preset threshold.

Optionally, according to a speech recognition method according to an embodiment of the present application, the vocabulary includes: a preset error-prone vocabulary and/or a real-time NER vocabulary.

Optionally, according to a speech recognition method according to an embodiment of the present application, the method further includes: acquiring real-time NER vocabulary in the display information corresponding to the target recognition object in real time based on the NER technology.

In a third aspect, an embodiment of the present application provides a speech recognition device, including a memory, including a memory, a transceiver, and a processor:

a memory for storing a computer program; a transceiver for sending and receiving data under the control of the processor; a processor for reading the computer program in the memory and performing the following operations:

The edge terminal obtains a cloud recognition result, where the cloud recognition result includes at least one target sequence obtained by the cloud for target recognition object recognition;

Based on the local reference text corresponding to the target recognition object, the at least one target sequence is modified to obtain an edge end recognition result.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, the at least one target sequence is modified based on the local reference text corresponding to the target recognition object to obtain an edge recognition result, including:

Based on the language model corresponding to the target sequence and the local reference text, vocabulary replacement is performed on a candidate sequence in the target sequence; wherein, the language model corresponding to the target sequence is obtained by training based on the local reference text of;

and / or,

Based on named entity recognition NER, lexical replacement is performed on one of the candidate sequences in the target sequence.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, the vocabulary matching probability of the candidate sequence is the highest among the vocabulary matching probabilities of all the target sequences;

Wherein, for each of the target sequences, the vocabulary matching probability is calculated based on the language model corresponding to the target sequence and the target sequence;

The word matching probability of the target sequence is used to describe the frequency of words in the target sequence appearing in the local reference text.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, the vocabulary replacement is performed on an alternative sequence in the target sequence based on the language model corresponding to the target sequence and the local reference text, include:

Based on the language model corresponding to the target sequence, obtain the continuous matching probability corresponding to the continuous word combination in the candidate sequence, wherein the continuous matching probability is used to indicate that the continuous matching probability corresponding to the continuous word combination is in the the frequency of occurrences in the local reference text;

If the continuous matching probability corresponding to any continuous word combination in the candidate sequence is lower than the first preset threshold, the continuous matching probability is replaced by the replacement text in the local reference text the first consecutive word combination of the preset threshold;

Wherein, the phoneme matching degree of the phoneme sequence corresponding to the replacement text and the phoneme sequence of the first continuous vocabulary combination is greater than a second preset threshold, and the replacement text is higher than the first continuous vocabulary combination in the local reference Occurs more frequently in text.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, the target sequence includes:

phoneme sequences, and/or word sequences;

Correspondingly, the language model corresponding to the phoneme sequence includes a phoneme sequence language model;

The language model corresponding to the word sequence includes a word sequence language model.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, if the target sequence includes a phoneme sequence, the target sequence is based on the language model corresponding to the target sequence and the local reference text. An alternative sequence of lexical replacements, including:

Based on the phoneme sequence language model, vocabulary replacement is performed on a candidate sequence in the phoneme sequence to obtain a first phoneme sequence;

Based on the word sequence language model, vocabulary replacement is performed on a candidate sequence in the at least one word sequence corresponding to the first phoneme sequence.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, if the target sequence further includes a word sequence, the target sequence is determined based on the language model corresponding to the target sequence and the local reference text. lexical replacement with an alternative sequence in , which also includes:

If the fourth matching probability corresponding to the first word sequence is greater than the fourth matching probability corresponding to the second word sequence, then determining that the first word sequence is the edge recognition result;

If the fourth matching probability corresponding to the second word sequence is greater than the fourth matching probability corresponding to the first word sequence, then determining that the second word sequence is the edge recognition result;

Wherein, the first word sequence is obtained after performing vocabulary replacement on a candidate sequence in at least one word sequence corresponding to the first phoneme sequence;

The second word sequence is obtained by performing lexical replacement on a candidate sequence in the word sequence in the target sequence based on the word sequence language model;

The fourth matching probability corresponding to the first word sequence is used to describe the degree of matching between the first word sequence and the local reference text;

The fourth matching probability corresponding to the second word sequence is used to describe the degree of matching between the second word sequence and the local reference text.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, the NER-based named entity recognition, performing vocabulary replacement on an alternative sequence in the target sequence, and obtaining the edge recognition result, including:

Based on the NER vocabulary, replace the vocabulary in a candidate sequence in the target sequence by the first replacement vocabulary in the NER vocabulary to obtain the edge recognition result;

Wherein, the phoneme matching probability of the phoneme sequence corresponding to the replacement vocabulary and the phoneme sequence of the vocabulary in the candidate sequence is greater than a third preset threshold.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, the operation further includes:

Identifying the local reference text based on the NER, and generating an NER vocabulary;

Based on the dictionary and/or the character-to-phoneme G2P technology, the phonemes corresponding to the NER vocabulary are obtained.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, the local reference text includes:

Professional information about the event to which the local reference text belongs, participant information, event background information, and event content.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, after obtaining the edge recognition result, the operation further includes:

Based on the professional information of the activity to which the local reference text belongs, the relevant information of the local reference text is saved to the server.

In a fourth aspect, an embodiment of the present application further provides an apparatus, including a memory, including a memory, a transceiver, and a processor:

a memory for storing a computer program; a transceiver for sending and receiving data under the control of the processor; a processor for reading the computer program in the memory and performing the following operations:

The device end obtains the cloud identification result or the edge identification result, wherein the cloud identification result includes a target sequence obtained by the cloud identifying the target identification object;

Based on the local vocabulary on the device side, the one target sequence is modified to obtain the device-side recognition result.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, the target sequence is modified based on a local vocabulary on the device side to obtain a device-side recognition result, including:

Based on the second replacement vocabulary in the local vocabulary, replace the vocabulary corresponding to the second replacement vocabulary in the target sequence, and obtain the device-side recognition result;

Wherein, the phoneme matching probability of the phoneme sequence of the vocabulary corresponding to the second replacement vocabulary and the phoneme sequence of the second replacement vocabulary is greater than a fourth preset threshold.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, the vocabulary includes: a preset error-prone vocabulary and/or a real-time NER vocabulary.

Optionally, according to a speech recognition apparatus according to an embodiment of the present application, the operation further includes: acquiring real-time NER words in the display information corresponding to the target recognition object in real time based on the NER technology.

In a fifth aspect, an embodiment of the present application further provides a speech recognition device, including:

a first obtaining unit, used for the edge terminal to obtain a cloud recognition result, where the cloud recognition result includes at least one target sequence obtained by the cloud for target recognition object recognition;

The first correcting unit is configured to correct the at least one target sequence based on the local reference text corresponding to the target recognition object to obtain an edge end recognition result.

In a sixth aspect, an embodiment of the present application further provides a speech recognition device, including:

The second obtaining unit is used for the device end to obtain the cloud identification result or the edge identification result, wherein the cloud identification result includes a target sequence obtained by the cloud identifying the target identification object;

The second correcting unit is configured to correct the one target sequence based on the local vocabulary of the device to obtain the identification result of the device.

In a seventh aspect, an embodiment of the present application further provides a processor-readable storage medium, where the processor-readable storage medium stores a computer program, and the computer program is configured to cause the processor to perform the above-mentioned primary selection aspect The steps of the method of the first aspect.

In the speech recognition method, device and storage medium provided by the embodiments of the present application, the recognition result of the target recognition object in the cloud is obtained through the edge terminal, and the cloud recognition result is corrected based on the local reference text corresponding to the target recognition object, so as to realize the recognition of the target recognition object. The optimization of the recognition results in the cloud improves the accuracy of speech recognition.

Description of drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the accompanying drawings used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description These are some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without any creative effort.

1 is a schematic flowchart of speech recognition provided by the prior art;

2 is one of the schematic flowcharts of speech recognition provided by an embodiment of the present application;

FIG. 3 is the second schematic flowchart of speech recognition provided by an embodiment of the present application;

FIG. 4 is the third schematic flowchart of speech recognition provided by an embodiment of the present application;

5 is one of the schematic structural diagrams of the speech recognition device provided by the embodiment of the present application;

FIG. 6 is a second schematic structural diagram of a speech recognition device provided by an embodiment of the present application;

7 is a third schematic structural diagram of a speech recognition device provided by an embodiment of the present application;

FIG. 8 is a fourth schematic structural diagram of a speech recognition apparatus provided by an embodiment of the present application.

Detailed ways

The term "and/or" in the embodiments of the present application describes the association relationship between associated objects, indicating that three relationships can exist. For example, A and/or B can indicate that A exists alone, A and B exist simultaneously, and B exists alone these three situations. The character "/" generally indicates that the associated objects are an "or" relationship.

In the embodiments of the present application, the term "plurality" refers to two or more than two, and other quantifiers are similar.

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

Embodiments of the present application provide a speech recognition method and apparatus, so as to improve the accuracy of speech recognition.

The method and the device are conceived based on the same application. Since the principles of the method and the device for solving the problem are similar, the implementation of the device and the method can be referred to each other, and repeated descriptions will not be repeated here.

The technical solutions provided in the embodiments of the present application can be applied to various systems, especially 5G systems. For example, applicable systems may be global system of mobile communication (GSM) system, code division multiple access (CDMA) system, wideband code division multiple access (WCDMA) general packet radio Service (general packet radio service, GPRS) system, long term evolution (long term evolution, LTE) system, LTE frequency division duplex (frequency division duplex, FDD) system, LTE time division duplex (time division duplex, TDD) system, advanced long-term Evolution (long term evolution advanced, LTE-A) system, universal mobile telecommunication system (UMTS), worldwide interoperability for microwave access (WiMAX) system, 5G New Radio (New Radio, NR) system, etc. . These various systems include terminal equipment and network equipment. The system may also include a core network part, such as an evolved packet system (Evloved Packet System, EPS), a 5G system (5GS), and the like.

The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

In the prior art, FIG. 1 is a schematic flowchart of speech recognition provided by the prior art. As shown in FIG. 1 , speech recognition is a process of converting speech into text.

Considering the auditory characteristics of the human ear, Mel cepstral coefficients or perceptual linear prediction coefficients have become one of the current mainstream speech feature vector extraction methods. After adding their first-order and second-order differences and normalizing the feature vectors, , which achieves good results on large-vocabulary continuous speech recognition problems.

The acoustic model is the underlying model of the speech recognition system and is the most critical part of the speech recognition system. The continuous speech signal is composed of some basic speech units. These basic speech units can be sentences, phrases, words, syllables, sub-syllables or phonemes. Which speech unit is selected as the modeling of the acoustic model The unit is determined by the specific application (such as the size of the vocabulary, the size of the voice library, the required performance and other objective factors). Generally speaking, it should be ensured that the selected modeling units meet the following conditions: 1) Robustness: each model has enough samples to estimate the model parameters; 2) Consistency: the modeling units should be stable under different conditions The lower acoustic characteristic changes are relatively small. In continuous speech, the implementation of modeling units in different contexts is sometimes very different due to the effect of co-articulation. In order to improve the accuracy of the model, it is necessary to consider the influence of the context on the modeling unit. In the study of acoustic models, context-dependent modeling units (such as diphones and triphones) have gradually received attention and have become the mainstream of the current acoustic model modeling units.

With the continuous development of speech recognition technology, the role of language models in speech recognition is becoming more and more important. Due to the dynamic time-varying, instantaneous and random nature of acoustic signals, it is impossible to recognize and understand speech without errors only by matching and judging acoustic patterns. The utilization of some higher-level linguistic knowledge can reduce the ambiguity of pattern matching at the level of acoustic recognition, thereby improving the accuracy of recognition. And a large-vocabulary continuous speech recognition system must detect at every moment whether a speech sound boundary is encountered, so that many different words or words will be recognized from different speech streams. To remove the ambiguity between these words or words, a language model is essential. Language models can provide contextual information and semantic information between words or words. Language models are not only used in speech recognition systems, but also in research fields such as machine translation and information retrieval. With the development of statistical language processing methods, statistical language models have become the mainstream technology of language processing in speech recognition.

Search is the process of finding the optimal sentence in the space composed of sentences, according to certain optimization criteria, that is, using the acquired knowledge (acoustic knowledge, phonetics knowledge, dictionary knowledge, language model knowledge, grammar and semantic knowledge, etc.) , find the optimal state sequence in the state (referring to the state of the phrase, word, modeling unit or HMM) space. The acoustic model, pronunciation dictionary and language model are closely combined through the Finite State Transducer (FST), and the search is performed on the FST.

The language model of the cloud speech recognition engine requires a large amount of text corpus during training. These text corpora are mainly corpora in general fields, and some corpora in other fields will also be added, but it is impossible to cover all fields. When performing speech recognition tasks in specific fields, such as speech recognition tasks in medicine, architecture, artificial intelligence and other fields, because the language model does not cover enough in these fields, the performance of automatic speech recognition technology (Automatic Speech Recognition, ASR) is caused. decreased significantly.

Therefore, for a specific recognition task, if some text corpora in this field can be added to optimize the recognition results, the recognition accuracy of the current recognition task can be significantly improved. In addition, it has become a trend to link the cloud, edge computing, and device-side computing on the Internet to build an Internet of Things (IoT) core computing capability. For the Internet of Things, breakthroughs in edge computing technology mean that many controls (calculations) will be implemented through local devices without having to be handed over to the cloud, and the processing will be completed at the local edge computing layer. This will undoubtedly greatly improve processing efficiency and reduce the load on the cloud. Due to being closer to the user, it can also provide users with a faster response and solve their needs at the edge. As a key technology in the 5G era, edge computing will become one of the indispensable infrastructures in the future. Therefore, after obtaining the basic ASR recognition results in the cloud, we can further improve the ASR recognition accuracy at the edge and device sides. The present application adopts the method of combining the cloud, the edge and the terminal to improve the recognition accuracy.

FIG. 2 is one of the schematic flowcharts of the speech recognition method provided by the embodiment of the present application. As shown in FIG. 2 , the method includes the following steps:

Step 201, the edge terminal obtains a cloud recognition result, and the cloud recognition result includes at least one target sequence obtained by the cloud for target recognition object recognition;

Optionally, the target recognition objects include but are not limited to audio or video files that can be acquired by the cloud or sent to the cloud. The cloud recognizes the target audio or video files to obtain a cloud recognition result that includes at least one target sequence, and the edge terminal obtains the Describe the cloud recognition result.

For example, the target recognition object can come from the real-time speech of one or more speakers in a professional conference, and the speech of the real-time speech can be recognized as the target recognition object in sentence-unit or segment-unit.

Step 202, modifying at least one target sequence based on the local reference text corresponding to the target recognition object to obtain an edge end recognition result.

With the cloud-side-end speech recognition method in this embodiment, the cloud recognition result is corrected to improve the accuracy of speech recognition.

Optionally, the local reference text includes: professional information of the activity to which the local reference text belongs, participant information, activity background information, and activity content.

Optionally, the application scenario of the recognition object, that is, the activity to which the local reference text belongs, may be activities such as conferences, lectures, and lectures.

Optionally, the edge terminal may obtain the cloud recognition result, including at least one target sequence obtained by the cloud for target recognition object recognition, and then may modify the at least one target sequence based on the local reference text to obtain the edge terminal recognition result.

Optionally, in this application, the combination of the cloud and the edge terminal is used to improve the accuracy of speech recognition, and the local reference text may not be uploaded to the cloud. The data optimizes the ASR recognition results returned from the cloud to protect the privacy of conference data. There is no need to transmit conference-related data to the cloud, but only process it on the edge and device. The security of activities can be improved by having a higher level of cybersecurity controls within the organization.

Optionally, the cloud uses a general ASR, which does not need to be changed according to specific application scenarios.

Optionally, at the edge end, the data related to this recognition task, that is, the local reference text, is used to optimize the recognition result.

Optionally, the embodiments of the present application can improve the recognition accuracy rate of speech recognition in a specific scenario. The specific scenario can be a scene with reference materials and speech to be recognized, which is not limited in the embodiments of the present application, that is, it can be Speech recognition based on reference material.

In a speech recognition method provided by an embodiment of the present application, the recognition result of a target recognition object in the cloud is obtained through an edge terminal, and based on the local reference text corresponding to the target recognition object, the recognition result in the cloud is corrected, so as to realize the recognition of the recognition result in the cloud. Optimization to improve the accuracy of speech recognition.

Optionally, modifying the at least one target sequence based on the local reference text corresponding to the target recognition object to obtain an edge end recognition result, including:

Based on the language model corresponding to the target sequence and the local reference text, vocabulary replacement is performed on a candidate sequence in the target sequence; wherein, the language model corresponding to the target sequence is obtained by training based on the local reference text of;

and / or,

Based on named entity recognition NER, lexical replacement is performed on one of the candidate sequences in the target sequence.

Optionally, based on the language model corresponding to the target sequence and the local reference text, vocabulary replacement can be performed on a candidate sequence in the target sequence;

Optionally, the language model corresponding to the target sequence is obtained by training based on the local reference text.

Optionally, the language model can obtain a probability distribution over a sequence of words, and for a given sequence of length m, it can generate a probability P(w_1,w_2,...,w_m) for the entire sequence. In fact, it is to find a way to find a probability distribution, which can represent the probability of occurrence of any sentence or sequence.

Optionally, a local reference text can be used as a sample to train to obtain a corresponding language model.

Optionally, for text content, training can be performed directly to obtain a language model for text sequences (ie word sequences), that is, a word sequence language model.

Optionally, for the speech content, the text content of the reference text can be converted into phonemes for training to obtain a language model for phoneme sequences, that is, a phoneme sequence language model.

Optionally, probability calculation can be performed on a word sequence in the target sequence through a word sequence language model to determine the degree of matching with the reference text, and the words or word combinations with low matching degree can be replaced. Replace with words or phrases that sound similar but more likely in the local reference text.

Optionally, a phoneme sequence in the target sequence can be probabilistically calculated through a phoneme sequence language model to determine the degree of matching with the reference text (in the case that the local reference text is known, the phoneme sequence of the local reference text is known. and unique), and replace some of the phonemes with low matching degree, which can be replaced with phonemes with similar pronunciation but higher occurrence probability in the local reference text.

Optionally, based on named entity recognition NER, vocabulary replacement can be performed on one of the candidate sequences in the target sequence.

Optionally, Named Entity Recognition (NER), also known as "proper name recognition", refers to identifying entities with specific meanings in texts, mainly including names of persons, places, institutions, proper nouns, and the like. It usually consists of two parts: (1) entity boundary identification; (2) determination of entity category (person name, place name, institution name or other).

Optionally, in order to solve the problem of inaccurate recognition of proper nouns, names of institutions or organizations or people and other similar words in speech recognition, NER can be recognized based on named entities, and one of the alternative sequences in the target sequence can be identified. Perform lexical replacement, specifically to replace inaccurately identified names of persons, places, institutions or other proper nouns.

Optionally, a phoneme can be understood as the pinyin of a Chinese character, or it can be understood as a letter pronounced in a word, for example, the phoneme corresponding to the word "name" is "nam".

Optionally, the lexical matching probability of the candidate sequence is the highest among all lexical matching probabilities of the target sequence;

Wherein, for each of the target sequences, the vocabulary matching probability is calculated based on the language model corresponding to the target sequence and the target sequence;

The word matching probability of the target sequence is used to describe the frequency of words in the target sequence appearing in the local reference text.

Optionally, the candidate sequence may have the highest lexical matching probability among all the lexical matching probabilities of the target sequences; that is, the frequency of the vocabulary included in the candidate sequence in the local reference text is the highest among all target sequences. .

Optionally, when determining the candidate sequence in the target sequence, a corresponding language model (LanguageModel, LM) can be used to perform probability calculation on the recognition result obtained from the cloud, and the recognition result with the highest probability is obtained as the candidate sequence, wherein the probability is the vocabulary matching probability. For word sequences, it can be the frequency of the vocabulary in the local reference text, and for the phoneme sequence, it can be the frequency of the phonemes in the phoneme sequence corresponding to the local reference text. , or it can be the frequency that the words corresponding to the phonemes appear in the local reference text respectively.

Optionally, the vocabulary corresponding to a group of phonemes may be one of the most likely words, or may be multiple words. For example, the vocabulary corresponding to a group of phonemes "shan feng" may be "mountain peak" or "mountain wind". .

Optionally, performing vocabulary replacement on an alternative sequence in the target sequence based on the language model corresponding to the target sequence and the local reference text, including:

Based on the language model corresponding to the target sequence, obtain the continuous matching probability corresponding to the continuous word combination in the candidate sequence, wherein the continuous matching probability is used to indicate that the continuous matching probability corresponding to the continuous word combination is in the the frequency of occurrences in the local reference text;

If the continuous matching probability corresponding to any continuous word combination in the candidate sequence is lower than the first preset threshold, the continuous matching probability is replaced by the replacement text in the local reference text the first consecutive word combination of the preset threshold;

Wherein, the phoneme matching degree of the phoneme sequence corresponding to the replacement text and the phoneme sequence of the first continuous vocabulary combination is greater than a second preset threshold, and the replacement text is higher than the first continuous vocabulary combination in the local reference Occurs more frequently in text.

Optionally, based on the language model corresponding to the target sequence, obtain the continuous matching probability corresponding to the continuous vocabulary combination in the candidate sequence;

Optionally, the continuous matching probability may be used to indicate the frequency of occurrence of the continuous word combination corresponding to the continuous matching probability in the local reference text; that is, the higher the continuous matching probability, that is, the corresponding continuous word combination in the local reference text. The higher the frequency in the text.

Optionally, the target sequence may include a phoneme sequence and/or a word sequence;

Taking the target sequence including the phoneme sequence as an example, the corresponding language model may be the phoneme sequence language model, so the edge end can use the phoneme sequence language model to calculate and obtain the continuous matching probability of a candidate sequence in the phoneme sequence.

Taking the word sequence including the word sequence as an example, the corresponding language model may be the word sequence language model, so the edge terminal can use the word sequence language model to calculate and obtain the continuous matching probability of a candidate sequence in the word sequence.

Optionally, the vocabulary included in the candidate sequence has the highest frequency in the local reference text among all target sequences.

Optionally, after calculating and obtaining the continuous matching probability of all or part of the continuous word combinations in the candidate sequence, it can be determined that the continuous matching probability corresponding to the partial continuous word combination is lower than the first preset threshold, then it can be considered that the part Consecutive word combinations appear very rarely or not even in the local reference text, and there are words or continuous words with similar pronunciation in the local reference text, and they appear more frequently in the local reference text, then the pronunciation can be considered similar And words or phrases that appear more frequently are replacement texts, and some continuous word combinations whose continuous matching probability is lower than the first preset threshold are replaced.

Optionally, the phoneme matching degree of the phoneme sequence corresponding to the replacement text and the phoneme sequence of the first continuous vocabulary combination is greater than a second preset threshold (that is, the pronunciation is similar), and the replacement text is higher than the first continuous vocabulary combination. Occurs more frequently in the local reference text. For example, if the replacement text is "Zhang San", and the first continuous word combination is "Zhang San", the pronunciations of the two are similar, but "Zhang San" appears more frequently in the local reference text, the alternative sequence can be The first consecutive word combination "Zhang San" is replaced by "Zhang San".

Optionally, the first preset threshold may be a preset positive number less than 1, the second preset threshold may be a positive number less than 1, the phoneme sequence corresponding to the replacement text and the phoneme sequence of the first continuous vocabulary combination. The phoneme matching degree is greater than the second preset threshold, indicating that in the reference text, the pronunciation of the replacement text is similar to the pronunciation of the first continuous word combination. The accuracy of speech recognition can be improved by replacing consecutive word combinations in the candidate sequence with replacement text that is similar in pronunciation and has a higher probability of appearing in the reference text.

Optionally, the target sequence includes:

phoneme sequences, and/or word sequences;

Optionally, a phoneme can be understood as the pinyin of a Chinese character, or it can be understood as a letter pronounced in a word, for example, the phoneme corresponding to the word "name" is "nam".

Optionally, the vocabulary corresponding to a group of phonemes may be one of the most likely words, or may be multiple words. For example, the vocabulary corresponding to a group of phonemes "shan feng" may be "mountain peak" or "mountain wind". .

Correspondingly, the language model corresponding to the phoneme sequence includes a phoneme sequence language model;

The language model corresponding to the word sequence includes a word sequence language model.

Optionally, the target sequence may be a phoneme sequence, a word sequence, or both a phoneme sequence and a word sequence.

Correspondingly, the language model corresponding to the phoneme sequence includes a phoneme sequence language model; the language model corresponding to the word sequence includes a word sequence language model.

Optionally, the phoneme sequence language model and the word sequence language model are obtained by training based on the local reference text.

Alternatively, a phoneme sequence language model or word sequence language model can obtain a probability distribution over a phoneme sequence or word sequence, and for a given sequence of length m, it can generate a probability P(w_1,w_2, ..., w_m). In fact, it is to find a way to find a probability distribution, which can represent the probability of occurrence of any sentence or sequence.

Specifically, the recognition result obtained from the cloud may be an n-best (top N best) phoneme sequence recognition result, an n-best (top N best) word sequence recognition result, or an n-best (top N best) word sequence recognition result. Top N best) phoneme sequence recognition results and n-best (top N best) word sequence recognition results. Wherein, the phoneme sequence is a sequence composed of the smallest phonetic units, the word sequence may be a sequence composed of words or words, and N may be a natural number greater than 0.

Optionally, if the target sequence includes a word sequence, performing word replacement on an alternative sequence in the target sequence based on the language model corresponding to the target sequence and the local reference text, including:

Based on the word sequence language model, vocabulary replacement is performed on a candidate sequence in the word sequence to obtain a first word sequence.

Optionally, the word sequence includes N word sequences, such as the phonetic pronunciation of "shanfeng" contains three word sequences of "mountain peak, mountain wind, fan wind", and the above-mentioned three word sequences are subjected to probability calculation by the word sequence language model. , if the occurrence probability of "mountain peak" in the local reference text is the highest, then "mountain peak" is obtained as the candidate sequence.

Optionally, if the target sequence includes a phoneme sequence, performing lexical replacement on an alternative sequence in the target sequence based on the language model corresponding to the target sequence and the local reference text, including:

Based on the phoneme sequence language model, vocabulary replacement is performed on a candidate sequence in the phoneme sequence to obtain a first phoneme sequence;

Based on the word sequence language model, vocabulary replacement is performed on a candidate sequence in the at least one word sequence corresponding to the first phoneme sequence.

Optionally, based on the phoneme sequence language model, vocabulary replacement is performed on a candidate sequence in the phoneme sequence to obtain a first phoneme sequence.

For example, the target recognition result is pronounced "zhangsan", and the phoneme sequence obtained from the cloud can be "zhangsan, zhangshan, zhangsang", and the above phoneme sequence is subjected to probability calculation through the phoneme language model, and the phoneme zhang and phoneme san are obtained respectively in the description The probability of the local reference text is the highest, then the phoneme sequence "zhangsan" is selected as the candidate sequence. If the probability of occurrence is higher, "zhuangsan" is used as the first phoneme sequence.

For another example, the target recognition result is pronounced "wo de ming zi shi zhangsan", and the phoneme sequence obtained from the cloud can be "wo de ming zi shi zhangsan", "wo de ming zi shi zhangshan", "wo deming zi shi zhangsang" ”, perform probability calculation on the above phoneme sequence through the phoneme language model, and obtain that the phoneme sequence “wo de ming zi shi zhangsan” has the highest probability in the local reference text, then select the phoneme sequence “wode ming zi shi zhangsan” as the candidate sequence , if the "zhangsan" in the alternative sequence "wo de ming zi shi zhangsan" has a higher probability of occurrence in the local reference text than the alternative sequence "zhangsan" in the local reference text, then "zhuangsan" Substitute the phoneme "zhangsan" to obtain the first phoneme sequence "wo deming zi shi zhuangsan".

Optionally, based on the word sequence language model, vocabulary replacement is performed on a candidate sequence in the at least one word sequence corresponding to the first phoneme sequence. The first phoneme sequence can correspond to multiple word sequences, for example, the phoneme sequence "wo demingzi shi zhangsan" can correspond to "my name is Zhang San", "my name is Zhang San", "my name is Zhang San", etc. word sequence. Among them, "my name is Zhang San" has the highest frequency of occurrence in the local reference text, and "my name is Zhang San" is selected as an alternative sequence in the at least one word sequence corresponding to the first phoneme sequence.

Optionally, if the target sequence includes a phoneme sequence and a word sequence, based on the language model corresponding to the target sequence and the local reference text, perform vocabulary replacement on an alternative sequence in the target sequence, and also include:

If the fourth matching probability corresponding to the first word sequence is greater than the fourth matching probability corresponding to the second word sequence, then determining that the first word sequence is the edge recognition result;

If the fourth matching probability corresponding to the second word sequence is greater than the fourth matching probability corresponding to the first word sequence, then determining that the second word sequence is the edge recognition result;

Wherein, the first word sequence is obtained after performing vocabulary replacement on a candidate sequence in at least one word sequence corresponding to the first phoneme sequence;

Optionally, the first phoneme sequence may correspond to multiple word sequences, and one of the word sequences is selected as a candidate sequence, and the word matching probability of the candidate sequence is the highest among the word matching probabilities of all the target sequences. For example, the vocabulary corresponding to a group of phonemes "shan feng" can be "mountain peak" or "mountain wind". If "mountain peak" has the highest occurrence probability in the local reference text, "mountain peak" is selected as the phoneme "shan feng". Alternative sequence.

Optionally, if the speech pronunciation of the target recognition object is "wo de ming zi shi zhang san", the word sequence includes "my name is Zhang San, my name is Zhang San, and my name is Zhang San", Among the above three word sequences, "my name is Zhang San" has the highest occurrence probability in the local reference text, and "my name is Zhang San" is selected as the candidate sequence. The next word "Zhang San" of "yes" in the alternative sequence has a low probability of occurrence through the calculation in the language model, then a word or word with a similar pronunciation and a higher probability of occurrence in the reference text can be selected for replacement, such as If "Zhang Shan" appears many times in the reference text, replace "Zhang San" with "Zhang Shan" to get the first word sequence of "My name is Zhang Shan".

The second word sequence is obtained by performing lexical replacement on a candidate sequence in the word sequence in the target sequence based on the word sequence language model;

The fourth matching probability corresponding to the first word sequence is used to describe the degree of matching between the first word sequence and the local reference text;

The fourth matching probability corresponding to the second word sequence is used to describe the degree of matching between the second word sequence and the local reference text.

Optionally, based on the named entity recognition NER, vocabulary replacement is performed on a candidate sequence in the target sequence to obtain the edge recognition result, including:

Based on the NER vocabulary, replace the vocabulary in a candidate sequence in the target sequence by the first replacement vocabulary in the NER vocabulary to obtain the edge recognition result;

Wherein, the phoneme matching probability of the phoneme sequence corresponding to the replacement vocabulary and the phoneme sequence of the vocabulary in the candidate sequence is greater than a third preset threshold.

Optionally, a NER vocabulary is generated by local reference text, and the edge end replaces the vocabulary in an alternative sequence in the target sequence by the first replacement vocabulary in the NER vocabulary, and obtains the edge end recognition result, The first replacement word is a word whose phoneme matching probability of the phoneme sequence of the corresponding phoneme sequence and the word in the candidate sequence is greater than the third preset threshold, that is, the first replacement word and the word in the candidate sequence. The corresponding words have similar pronunciations. The third preset threshold is a positive number less than 1, and the phoneme matching probability between the phoneme sequence corresponding to the replacement vocabulary and the phoneme sequence of the vocabulary in the candidate sequence is greater than the third preset threshold, indicating that in the NER vocabulary The replacement words sound similar to words in the alternative sequence.

Optionally, the method further includes:

Identifying the local reference text based on the NER, and generating an NER vocabulary;

Based on a dictionary and/or a character-to-phoneme (Grapheme-to-Phoneme, G2P) technology, the phoneme corresponding to the NER vocabulary is obtained.

Optionally, based on the NER vocabulary, the first replacement vocabulary in the NER vocabulary is used to replace the vocabulary in a candidate sequence in the target sequence, and the edge recognition result is obtained, as described in the target sequence. The word in one of the candidate sequences in the sequence is "open ai", and the word "OpenAI" corresponding to the pronunciation of "open ai" in the NER vocabulary has a higher matching probability, then the word in the candidate sequence " "open ai" is replaced by "OpenAI", where "OpenAI" is an artificial intelligence non-profit organization jointly established by many Silicon Valley tycoons.

Specifically, the dictionary is requested from the cloud and downloaded to the edge.

Named Entity Recognition (NER, Named Entity Recognition) technology is used to identify person names, place names, organization names, proper nouns, etc. in the text content to generate NER vocabulary.

Optionally, the local reference text includes:

Professional information about the event to which the local reference text belongs, participant information, event background information, and event content.

Optionally, the local reference text may include conference speech, screen projection ppt of the conference, names, positions, background identities of participants, background information of the conference, and the like.

Optionally, in this application, the combination of the cloud and the edge terminal is used to improve the accuracy of speech recognition, and the local reference text may not be uploaded to the cloud. The data optimizes the ASR recognition results returned from the cloud to protect the privacy of conference data. There is no need to transmit conference-related data to the cloud, but only process it on the edge and device. The security of activities can be improved by having a higher level of cybersecurity controls within the organization.

Optionally, the participant's name, position, title, etc. are added to the NER vocabulary, so as to improve the accuracy of the NER.

Specifically, if there may be uncommon words in the information such as the names and positions of the participants, the cloud recognition result for the uncommon words may be inaccurate. In this application, the participant's name, position, title, etc. are added to the NER vocabulary, and the cloud's recognition result of the rare word can be replaced by the rare word in the NER vocabulary to obtain a more accurate recognition result.

Optionally, if the local reference text includes materials such as conference speech, screen projection ppt, etc., text recognition technology is used to automatically perform text recognition to recognize the text content in the image. If the local reference text is an electronic version of the material, word recognition is not required.

Optionally, the local reference text further includes conference-related background information such as the participant's name and job introduction.

At the edge, the speech recognition results returned by the cloud are optimized through materials such as conference speeches and screencast ppts.

Optionally, the pronunciation corresponding to the vocabulary in the NER vocabulary can be obtained by using the downloaded dictionary, the pronunciation corresponding to the vocabulary in the NER vocabulary can also be obtained by using the character-to-phoneme G2P technology, or based on the downloaded dictionary, using the character-to-phoneme conversion. G2P technology obtains the pronunciation corresponding to the vocabulary in the NER vocabulary.

Optionally, when the vocabulary in the NER vocabulary does not have a corresponding pronunciation in the downloaded dictionary, G2P technology can be used to convert the vocabulary in the NER vocabulary to obtain the NER vocabulary using the downloaded dictionary as a reference model. Pronunciation of vocabulary.

By using the cloud-side-terminal collaboration method in this embodiment, the computing resources of the cloud-side-terminal are fully utilized, so that the tasks of cloud processing are reduced, the cloud network delay is reduced, and the corresponding speed of speech recognition is improved.

Optionally, after obtaining the edge end identification result, the method further includes:

Based on the professional information of the activity to which the local reference text belongs, the relevant information of the local reference text is saved to the server.

Specifically, take the subject content of the activity to which the reference text belongs as a classification, and save the relevant information of the local reference text to the server. Information about the AI's local reference text is saved to the server's AI category. In this way, relevant text data can be continuously accumulated within the organization, and the recognition accuracy in this field can be continuously improved.

With the voice recognition orientation provided by the embodiment of the present application, the edge terminal can optimize the recognition result, and it is not necessary to transmit the conference-related data to the cloud, but only process it at the edge terminal, which can protect the privacy of the conference data. Implementing higher-level network security controls can improve meeting security. The application scenarios of this application are not limited to conferences, but can also be lectures, lectures and other activities. The cloud uses general speech recognition technology and does not need to be changed according to specific application scenarios.

In a speech recognition method provided by an embodiment of the present application, the recognition result of a target recognition object in the cloud is obtained through an edge terminal, and based on the local reference text corresponding to the target recognition object, the recognition result in the cloud is corrected, so as to realize the recognition of the recognition result in the cloud. Optimization to improve the accuracy of speech recognition.

FIG. 3 is the second schematic flowchart of the speech recognition method proposed by the embodiment of the present application. As shown in FIG. 3 , the method includes the following steps:

Step 301, the device terminal obtains a cloud identification result or an edge identification result, wherein the cloud identification result includes a target sequence obtained by the cloud identifying the target identification object;

Optionally, the recognition result of the edge terminal is the recognition result obtained after the edge terminal performs vocabulary replacement on the cloud recognition result, that is, optimization.

Step 302 , modifying the one target sequence based on the local vocabulary of the device to obtain the identification result of the device.

Optionally, if the device end obtains the recognition result of the edge end, the device end may continue to optimize the speech recognition result returned by the edge end.

In a speech recognition method provided by an embodiment of the present application, the recognition result of a target recognition object in the cloud is obtained through an edge terminal, and based on the local reference text corresponding to the target recognition object, the recognition result in the cloud is corrected, so as to realize the recognition of the recognition result in the cloud. Optimization to improve the accuracy of speech recognition.

Optionally, the local vocabulary based on the device side, modifying the one target sequence to obtain the device side identification result, including:

Based on the second replacement vocabulary in the local vocabulary, replace the vocabulary corresponding to the second replacement vocabulary in the target sequence, and obtain the device-side recognition result;

Optionally, the second replacement vocabulary is a vocabulary in the local vocabulary that is similar in pronunciation to the vocabulary in the target sequence and has a higher occurrence probability.

For example, the target sequence is "My name is Zhang San", wherein the words "Zhang San" with similar pronunciation in the local vocabulary are "Zhang San" and "Zhang San", wherein the vocabulary "Zhang San" " has a higher occurrence probability than the word "Zhang San" in the target sequence, then "Zhang San" is the second replacement word in the local vocabulary.

Wherein, the phoneme matching probability of the phoneme sequence of the vocabulary corresponding to the second replacement vocabulary and the phoneme sequence of the second replacement vocabulary is greater than a fourth preset threshold.

Specifically, the fourth preset threshold is a positive number less than 1, and the phoneme matching probability of the phoneme sequence of the vocabulary corresponding to the second replacement vocabulary and the phoneme sequence of the second replacement vocabulary is greater than the fourth preset threshold. It means that the pronunciation of the word in the local vocabulary is similar to the pronunciation of the word in the target sequence corresponding to the word in the local vocabulary.

Optionally, the vocabulary includes: a preset error-prone vocabulary and/or a real-time NER vocabulary.

Among them, the preset error-prone word list includes words that are easy to identify and wrong and words that appear frequently in this meeting but are not common.

Optionally, the method further includes: acquiring real-time NER vocabulary in the display information corresponding to the target recognition object in real time based on the NER technology.

Optionally, the real-time NER vocabulary in the display information corresponding to the target recognition object is acquired in real time.

Specifically, based on the NER technology, the information on the screen is captured in real time, and the NER vocabulary on the current screen is obtained. However, in the present application, it is not limited to the information on the screen, and may also be the information of the real-time captured entity text.

Optionally, the vocabulary includes: a preset error-prone vocabulary and/or a real-time NER vocabulary.

Optionally, use a syntax analysis technology to analyze the recognition results obtained by the device on the cloud or on the edge, and replace words such as subject and object in the vocabulary with words with similar pronunciation, thereby further improving the recognition accuracy.

In the speech recognition orientation provided by the embodiment of the present application, the recognition result is optimized by the device, so that it is not necessary to transmit the conference-related data to the cloud, but only processed on the device, which can protect the privacy of the conference data. Implementing higher-level network security controls can improve meeting security. In this application, it is not limited to meetings, but also activities such as lectures and lectures.

Optionally, the cloud recognition result of the target recognition object is obtained through the edge terminal, and based on the local reference text corresponding to the target recognition object, the cloud recognition result is corrected, so as to realize the optimization of the cloud recognition result and improve the accuracy of speech recognition. sex.

Fig. 4 is the third schematic flow chart of the speech recognition method provided by the embodiment of the present application. As shown in Fig. 4, the flow is specifically:

Step 401, obtain the n-best recognition result of the phoneme sequence and the n-best recognition result of the word sequence from the cloud;

Step 402, requesting a dictionary from the cloud, and downloading the dictionary to the edge;

Step 403, generate the LM at the edge, generate the NER vocabulary, obtain the pronunciation corresponding to the vocabulary in the NER vocabulary, use the LM to select the recognition result returned from the cloud, and use the LM and the NER vocabulary to optimize the recognition result;

Step 404 , construct a vocabulary list on the device side, obtain the NER vocabulary on the current screen, and further optimize the recognition result.

FIG. 5 is one of the schematic structural diagrams of the speech recognition device provided by the embodiment of the present application. As shown in FIG. 5 , the speech recognition device includes a memory, a transceiver, and a processor:

a memory for storing a computer program; a transceiver for sending and receiving data under the control of the processor; a processor for reading the computer program in the memory and performing the following operations:

The edge terminal obtains a cloud recognition result, where the cloud recognition result includes at least one target sequence obtained by the cloud for target recognition object recognition;

Based on the local reference text corresponding to the target recognition object, the at least one target sequence is modified to obtain an edge end recognition result.

The transceiver 502 is used to receive and transmit data under the control of the processor 503 .

5, the bus architecture may include any number of interconnected buses and bridges, specifically one or more processors represented by processor 503 and various circuits of memory represented by memory 501 are linked together. The bus architecture may also link together various other circuits, such as peripherals, voltage regulators, and power management circuits, which are well known in the art and, therefore, will not be described further herein. The bus interface provides the interface. Transceiver 502 may be a number of elements, including transmitters and receivers, providing means for communicating with various other devices over transmission media including wireless channels, wired channels, fiber optic cables, and the like Transmission medium. For different user equipments, the user interface 504 may also be an interface capable of externally connecting a desired device, and the connected devices include but are not limited to a keypad, a display, a speaker, a microphone, a joystick, and the like.

The processor 503 is responsible for managing the bus architecture and general processing, and the memory 501 may store data used by the processor 503 in performing operations.

Optionally, the processor 503 may be a CPU (central processor), an ASIC (Application Specific Integrated Circuit, application specific integrated circuit), an FPGA (Field-Programmable Gate Array, field programmable gate array) or a CPLD (Complex Programmable Logic Device, Complex Programmable Logic Device), the processor can also use a multi-core architecture.

Optionally, in one embodiment, modifying the at least one target sequence based on the local reference text corresponding to the target recognition object to obtain an edge recognition result, including:

Based on the language model corresponding to the target sequence and the local reference text, vocabulary replacement is performed on a candidate sequence in the target sequence; wherein, the language model corresponding to the target sequence is obtained by training based on the local reference text and/or, performing lexical replacement on one of the candidate sequences in the target sequence based on named entity recognition NER.

Optionally, in one embodiment, the lexical matching probability of the candidate sequence is the highest among all lexical matching probabilities of the target sequence;

Wherein, for each of the target sequences, the vocabulary matching probability is calculated based on the language model corresponding to the target sequence and the target sequence;

The word matching probability of the target sequence is used to describe the frequency of words in the target sequence appearing in the local reference text.

Optionally, in one embodiment, performing lexical replacement on a candidate sequence in the target sequence based on the language model corresponding to the target sequence and the local reference text, including:

Based on the language model corresponding to the target sequence, obtain the continuous matching probability corresponding to the continuous word combination in the candidate sequence, wherein the continuous matching probability is used to indicate that the continuous matching probability corresponding to the continuous word combination is in the the frequency of occurrences in the local reference text;

If the continuous matching probability corresponding to any continuous word combination in the candidate sequence is lower than the first preset threshold, the continuous matching probability is replaced by the replacement text in the local reference text the first consecutive word combination of the preset threshold;

Wherein, the phoneme matching degree of the phoneme sequence corresponding to the replacement text and the phoneme sequence of the first continuous vocabulary combination is greater than a second preset threshold, and the replacement text is higher than the first continuous vocabulary combination in the local reference Occurs more frequently in text.

Optionally, in one embodiment, the target sequence includes:

phoneme sequence, and/or word sequence; correspondingly, the language model corresponding to the phoneme sequence includes a phoneme sequence language model; the language model corresponding to the word sequence includes a word sequence language model.

Optionally, in one embodiment, if the target sequence includes a phoneme sequence, the vocabulary is performed on an alternative sequence in the target sequence based on the language model corresponding to the target sequence and the local reference text. Replacement, including:

Based on the phoneme sequence language model, vocabulary replacement is performed on a candidate sequence in the phoneme sequence to obtain a first phoneme sequence;

Based on the word sequence language model, vocabulary replacement is performed on a candidate sequence in the at least one word sequence corresponding to the first phoneme sequence.

Optionally, in one embodiment, if the target sequence further includes a word sequence, the target sequence is based on the language model corresponding to the target sequence and the local reference text, and an alternative sequence in the target sequence is performed. Vocabulary replacement, also including:

If the fourth matching probability corresponding to the first word sequence is greater than the fourth matching probability corresponding to the second word sequence, then determining that the first word sequence is the edge recognition result;

If the fourth matching probability corresponding to the second word sequence is greater than the fourth matching probability corresponding to the first word sequence, then determining that the second word sequence is the edge recognition result;

Wherein, the first word sequence is obtained after performing vocabulary replacement on a candidate sequence in at least one word sequence corresponding to the first phoneme sequence;

The second word sequence is obtained by performing lexical replacement on a candidate sequence in the word sequence in the target sequence based on the word sequence language model;

The fourth matching probability corresponding to the first word sequence is used to describe the degree of matching between the first word sequence and the local reference text;

The fourth matching probability corresponding to the second word sequence is used to describe the degree of matching between the second word sequence and the local reference text.

Optionally, in one embodiment, the NER-based named entity recognition, performing vocabulary replacement on a candidate sequence in the target sequence, and obtaining the edge recognition result, including:

Based on the NER vocabulary, replace the vocabulary in a candidate sequence in the target sequence by the first replacement vocabulary in the NER vocabulary to obtain the edge recognition result;

Wherein, the phoneme matching probability of the phoneme sequence corresponding to the replacement vocabulary and the phoneme sequence of the vocabulary in the candidate sequence is greater than a third preset threshold.

Optionally, in one embodiment, the operation further includes:

Identifying the local reference text based on the NER, and generating an NER vocabulary;

Based on the dictionary and/or the character-to-phoneme G2P technology, the phonemes corresponding to the NER vocabulary are obtained. Optionally, in one embodiment, the local reference text includes: professional information of the activity to which the local reference text belongs, participant information, activity background information, and activity content.

Optionally, in one embodiment, after obtaining the edge identification result, the operation further includes: saving the relevant information of the local reference text to the server based on the professional information of the activity to which the local reference text belongs.

A speech recognition device provided by an embodiment of the present application obtains the recognition result of the target recognition object in the cloud through the edge terminal, and corrects the recognition result in the cloud based on the local reference text corresponding to the target recognition object, so as to realize the recognition of the recognition result in the cloud. Optimization to improve the accuracy of speech recognition.

The processor is configured to execute any one of the methods provided in the embodiments of the present application according to the obtained executable instructions by invoking the computer program stored in the memory. The processor and memory may also be physically separated.

It should be noted here that the above-mentioned device provided by the embodiment of the present invention can realize all the method steps realized by the above-mentioned method embodiment, and can achieve the same technical effect, and the same as the method embodiment in this embodiment is not repeated here. The parts and beneficial effects will be described in detail.

FIG. 6 is the second structural schematic diagram of the speech recognition device provided by the embodiment of the present application. As shown in FIG. 6 , the speech recognition device includes a memory, a transceiver, and a processor:

a memory for storing a computer program; a transceiver for sending and receiving data under the control of the processor; a processor for reading the computer program in the memory and performing the following operations:

The device end obtains the cloud identification result or the edge identification result, wherein the cloud identification result includes a target sequence obtained by the cloud identifying the target identification object;

Based on the local vocabulary on the device side, the one target sequence is modified to obtain the device-side recognition result.

The transceiver 602 is used to receive and transmit data under the control of the processor 603 .

Wherein, in FIG. 6 , the bus architecture may include any number of interconnected buses and bridges, specifically one or more processors represented by the processor 603 and various circuits of the memory represented by the memory 601 are linked together. The bus architecture may also link together various other circuits, such as peripherals, voltage regulators, and power management circuits, which are well known in the art and, therefore, will not be described further herein. The bus interface provides the interface. Transceiver 602 may be a number of elements, including transmitters and receivers, providing means for communicating with various other devices over transmission media including wireless channels, wired channels, fiber optic cables, and the like Transmission medium. For different user equipments, the user interface 604 may also be an interface capable of externally connecting the required equipment, and the connected equipment includes but is not limited to a keypad, a display, a speaker, a microphone, a joystick, and the like.

The processor 603 is responsible for managing the bus architecture and general processing, and the memory 601 may store data used by the processor 603 in performing operations.

Optionally, the processor 603 may be a CPU (central processor), an ASIC (Application Specific Integrated Circuit, application specific integrated circuit), an FPGA (Field-Programmable Gate Array, field programmable gate array) or a CPLD (Complex Programmable Logic Device, Complex Programmable Logic Device), the processor can also use a multi-core architecture.

Optionally, the local vocabulary based on the device side, modifying the one target sequence to obtain the device side identification result, including:

Based on the second replacement vocabulary in the local vocabulary, replace the vocabulary corresponding to the second replacement vocabulary in the target sequence, and obtain the device-side recognition result;

Wherein, the phoneme matching probability of the phoneme sequence of the vocabulary corresponding to the second replacement vocabulary and the phoneme sequence of the second replacement vocabulary is greater than a fourth preset threshold.

Optionally, the vocabulary includes: a preset error-prone vocabulary and/or a real-time NER vocabulary.

Optionally, the operation further includes: acquiring real-time NER vocabulary in the display information corresponding to the target recognition object in real time based on the NER technology.

A speech recognition device provided by an embodiment of the present application obtains the recognition result of the target recognition object in the cloud through the edge terminal, and corrects the recognition result in the cloud based on the local reference text corresponding to the target recognition object, so as to realize the recognition of the recognition result in the cloud. Optimization to improve the accuracy of speech recognition.

The processor is configured to execute any one of the methods provided in the embodiments of the present application according to the obtained executable instructions by invoking the computer program stored in the memory. The processor and memory may also be physically separated.

It should be noted here that the above-mentioned device provided by the embodiment of the present invention can realize all the method steps realized by the above-mentioned method embodiment, and can achieve the same technical effect, and the same as the method embodiment in this embodiment is not repeated here. The parts and beneficial effects will be described in detail.

FIG. 7 is a third schematic structural diagram of a speech recognition device provided by an embodiment of the present application. As shown in FIG. 7 , the speech recognition device includes a first acquisition unit 710 and a first correction unit 720, wherein:

The first obtaining unit 710 is used for the edge terminal to obtain a cloud recognition result, where the cloud recognition result includes at least one target sequence obtained by the cloud for target recognition object recognition;

The first modifying unit 720 is configured to modify the at least one target sequence based on the local reference text corresponding to the target recognition object to obtain an edge recognition result.

Optionally, the speech recognition apparatus obtains the cloud recognition result through the first obtaining unit 710, and then may modify the at least one target sequence through the first modifying unit 720 based on the local reference text corresponding to the target recognition object to obtain an edge. end recognition result.

A speech recognition device provided by an embodiment of the present application obtains the recognition result of the target recognition object in the cloud through the edge terminal, and corrects the recognition result in the cloud based on the local reference text corresponding to the target recognition object, so as to realize the recognition of the recognition result in the cloud. Optimization to improve the accuracy of speech recognition.

It should be noted that the division of units in the embodiments of the present application is illustrative, and is only a logical function division, and other division methods may be used in actual implementation. In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware, or may be implemented in the form of software functional units.

If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in a processor-readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of software products in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, and the computer software products are stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor (processor) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other media that can store program codes .

Optionally, the first correction unit 720 is used for:

Based on the language model corresponding to the target sequence and the local reference text, vocabulary replacement is performed on a candidate sequence in the target sequence; wherein, the language model corresponding to the target sequence is obtained by training based on the local reference text of;

and / or,

Based on named entity recognition NER, lexical replacement is performed on one of the candidate sequences in the target sequence.

Optionally, the lexical matching probability of the candidate sequence is the highest among all lexical matching probabilities of the target sequence;

Wherein, for each of the target sequences, the vocabulary matching probability is calculated based on the language model corresponding to the target sequence and the target sequence;

The word matching probability of the target sequence is used to describe the frequency of words in the target sequence appearing in the local reference text.

Optionally, the first modification unit 720 is configured to: based on the language model corresponding to the target sequence, obtain the continuous matching probability corresponding to the continuous word combination in the candidate sequence, wherein the continuous matching probability is used to represent the the frequency that the continuous word combination corresponding to the continuous matching probability appears in the local reference text;

If the continuous matching probability corresponding to any continuous word combination in the candidate sequence is lower than the first preset threshold, the continuous matching probability is replaced by the replacement text in the local reference text the first consecutive word combination of the preset threshold;

Wherein, the phoneme matching degree of the phoneme sequence corresponding to the replacement text and the phoneme sequence of the first continuous vocabulary combination is greater than a second preset threshold, and the replacement text is higher than the first continuous vocabulary combination in the local reference Occurs more frequently in text.

Optionally, the target sequence includes:

phoneme sequences, and/or word sequences;

Correspondingly, the language model corresponding to the phoneme sequence includes a phoneme sequence language model;

The language model corresponding to the word sequence includes a word sequence language model.

Optionally, if the target sequence includes a phoneme sequence, the first modification unit 720 is configured to: based on the phoneme sequence language model, perform vocabulary replacement on an alternative sequence in the phoneme sequence to obtain a first phoneme sequence;

Based on the word sequence language model, vocabulary replacement is performed on a candidate sequence in the at least one word sequence corresponding to the first phoneme sequence.

Optionally, if the target sequence further includes a word sequence, the first modification unit 720 is configured to: if the fourth matching probability corresponding to the first word sequence is greater than the fourth matching probability corresponding to the second word sequence, determine the first matching probability. The word sequence is the edge recognition result;

If the fourth matching probability corresponding to the second word sequence is greater than the fourth matching probability corresponding to the first word sequence, then determining that the second word sequence is the edge recognition result;

Wherein, the first word sequence is obtained after performing vocabulary replacement on a candidate sequence in at least one word sequence corresponding to the first phoneme sequence;

The second word sequence is obtained by performing lexical replacement on a candidate sequence in the word sequence in the target sequence based on the word sequence language model;

The fourth matching probability corresponding to the first word sequence is used to describe the degree of matching between the first word sequence and the local reference text;

The fourth matching probability corresponding to the second word sequence is used to describe the degree of matching between the second word sequence and the local reference text.

Optionally, the first correction unit 720 is configured to: based on the NER vocabulary, replace the vocabulary in a candidate sequence in the target sequence with the first replacement vocabulary in the NER vocabulary to obtain the edge. end identification result;

Wherein, the phoneme matching probability of the phoneme sequence corresponding to the replacement vocabulary and the phoneme sequence of the vocabulary in the candidate sequence is greater than a third preset threshold.

Optionally, also include:

a first generating unit, configured to identify the local reference text based on the NER, and generate a NER vocabulary;

The third obtaining unit is configured to obtain the phonemes corresponding to the NER vocabulary based on the dictionary and/or the character-to-phoneme G2P technology.

Optionally, the local reference text includes:

Professional information about the event to which the local reference text belongs, participant information, event background information, and event content.

Optionally, also include:

After the edge identification result is obtained, the relevant information of the local reference text is stored in the server based on the professional information of the activity to which the local reference text belongs.

FIG. 8 is the fourth schematic structural diagram of the speech recognition device provided by the embodiment of the present application. As shown in FIG. 8 , the speech recognition device includes a second acquisition unit 810 and a second correction unit 820, wherein:

The second obtaining unit 810 is used for the device end to obtain the cloud identification result or the edge identification result, wherein the cloud identification result includes a target sequence obtained by the cloud identifying the target identification object;

The second modification unit 820 is configured to modify the one target sequence based on the local vocabulary of the device to obtain the identification result of the device.

Optionally, the speech recognition apparatus obtains the cloud recognition result or the edge recognition result through the second obtaining unit 810, and then, based on the local vocabulary of the device end, can correct the one target sequence through the second modifying unit 820 to obtain the device end. Identify the results.

A speech recognition device provided by an embodiment of the present application obtains the recognition result of the target recognition object in the cloud through the edge terminal, and corrects the recognition result in the cloud based on the local reference text corresponding to the target recognition object, so as to realize the recognition of the recognition result in the cloud. Optimization to improve the accuracy of speech recognition.

Optionally, the second correction unit 820 is used for:

Based on the second replacement vocabulary in the local vocabulary, replace the vocabulary corresponding to the second replacement vocabulary in the target sequence, and obtain the device-side recognition result;

Wherein, the phoneme matching probability of the phoneme sequence of the vocabulary corresponding to the second replacement vocabulary and the phoneme sequence of the second replacement vocabulary is greater than a fourth preset threshold.

Optionally, the vocabulary includes: a preset error-prone vocabulary and/or a real-time NER vocabulary.

Optionally, also include:

The fourth acquisition module is configured to acquire real-time NER vocabulary in the display information corresponding to the target recognition object in real time based on the NER technology.

It should be noted that the division of units in the embodiments of the present application is illustrative, and is only a logical function division, and other division methods may be used in actual implementation. In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware, or may be implemented in the form of software functional units.

If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in a processor-readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of software products in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, and the computer software products are stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor (processor) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other media that can store program codes .

It should be noted here that the above-mentioned device provided by the embodiment of the present invention can realize all the method steps realized by the above-mentioned method embodiment, and can achieve the same technical effect, and the same as the method embodiment in this embodiment is not repeated here. The parts and beneficial effects will be described in detail.

On the other hand, an embodiment of the present application further provides a processor-readable storage medium, where a computer program is stored in the processor-readable storage medium, and the computer program is used to enable the processor to execute the methods, including:

The edge terminal obtains a cloud recognition result, where the cloud recognition result includes at least one target sequence obtained by the cloud for target recognition object recognition;

Based on the local reference text corresponding to the target recognition object, the at least one target sequence is modified to obtain an edge end recognition result.

or

The device end obtains the cloud identification result or the edge identification result, wherein the cloud identification result includes a target sequence obtained by the cloud identifying the target identification object;

Based on the local vocabulary on the device side, the one target sequence is modified to obtain the device-side recognition result.

The processor-readable storage medium can be any available medium or data storage device that can be accessed by a processor, including, but not limited to, magnetic storage (eg, floppy disk, hard disk, magnetic tape, magneto-optical disk (MO), etc.), optical storage (eg, CD, DVD, BD, HVD, etc.), and semiconductor memory (eg, ROM, EPROM, EEPROM, non-volatile memory (NANDFLASH), solid-state disk (SSD)), and the like.

As will be appreciated by those skilled in the art, the embodiments of the present application may be provided as a method, a system, or a computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media having computer-usable program code embodied therein, including but not limited to disk storage, optical storage, and the like.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It will be understood that each flow and/or block in the flowcharts and/or block diagrams, and combinations of flows and/or blocks in the flowcharts and/or block diagrams, can be implemented by computer-executable instructions. These computer-executable instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

These processor-executable instructions may also be stored in a processor-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the processor-readable memory result in the manufacture of means including the instructions product, the instruction means implements the functions specified in the flow or flow of the flowchart and/or the block or blocks of the block diagram.

These processor-executable instructions can also be loaded onto a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process that Execution of the instructions provides steps for implementing the functions specified in the flowchart or blocks and/or the block or blocks of the block diagrams.

Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims (20)

1. A speech recognition method, comprising:

the method comprises the steps that an edge end obtains a cloud end identification result, wherein the cloud end identification result comprises at least one target sequence which is obtained by the cloud end aiming at target identification objects;

and correcting the at least one target sequence based on the local reference text corresponding to the target recognition object to obtain an edge end recognition result.

2. The speech recognition method according to claim 1, wherein the modifying the at least one target sequence based on the local reference text corresponding to the target recognition object to obtain an edge recognition result comprises:

performing vocabulary replacement on one alternative sequence in the target sequence based on the language model corresponding to the target sequence and the local reference text; the language model corresponding to the target sequence is obtained based on the local reference text training;

and/or the presence of a gas in the gas,

performing lexical replacement on one of the alternative sequences in the target sequence based on Named Entity Recognition (NER).

3. The speech recognition method of claim 2 wherein the vocabulary match probabilities of the alternative sequences are highest among the vocabulary match probabilities of all of the target sequences;

wherein, for each target sequence, the vocabulary matching probability is calculated based on the language model corresponding to the target sequence and the target sequence;

the vocabulary matching probability of the target sequence is used for describing the frequency of occurrence of the vocabularies in the target sequence in the local reference text.

4. The speech recognition method of claim 2, wherein the vocabulary replacement of the candidate sequence in the target sequence based on the language model corresponding to the target sequence and the local reference text comprises:

obtaining continuous matching probabilities corresponding to continuous word combinations in the alternative sequences based on the language model corresponding to the target sequence, wherein the continuous matching probabilities are used for representing the frequency of the continuous word combinations corresponding to the continuous matching probabilities appearing in the local reference text;

if the continuous matching probability corresponding to any continuous vocabulary combination in the alternative sequence is lower than a first preset threshold value, replacing the first continuous vocabulary combination with the continuous matching probability lower than the first preset threshold value through a replacement text in the local reference text;

the phoneme matching degree of the phoneme sequence corresponding to the replacement text and the phoneme sequence of the first continuous vocabulary combination is larger than a second preset threshold, and the occurrence frequency of the replacement text in the local reference text is higher than that of the first continuous vocabulary combination.

5. The speech recognition method of any one of claims 2 to 4, wherein the target sequence comprises:

a sequence of phonemes, and/or a sequence of words;

correspondingly, the language model corresponding to the phoneme sequence comprises a phoneme sequence language model;

the language model corresponding to the word sequence comprises a word sequence language model.

6. The speech recognition method of claim 5, wherein if the target sequence comprises a phoneme sequence, the vocabulary replacement of an alternative sequence in the target sequence based on the language model corresponding to the target sequence and the local reference text comprises:

performing vocabulary replacement on one alternative sequence in the phoneme sequences based on the phoneme sequence language model to obtain a first phoneme sequence;

and performing vocabulary replacement on one alternative sequence in at least one word sequence corresponding to the first phoneme sequence based on the word sequence language model.

7. The speech recognition method of claim 6, wherein if the target sequence further comprises a word sequence, the vocabulary replacement is performed on one alternative sequence in the target sequence based on the language model corresponding to the target sequence and the local reference text, and further comprising:

if the fourth matching probability corresponding to the first word sequence is greater than the fourth matching probability corresponding to the second word sequence, determining that the first word sequence is the edge end identification result;

if the fourth matching probability corresponding to the second word sequence is greater than the fourth matching probability corresponding to the first word sequence, determining that the second word sequence is the edge end identification result;

the first word sequence is obtained after vocabulary replacement is carried out on one alternative sequence in at least one word sequence corresponding to the first phoneme sequence;

the second word sequence is obtained after vocabulary replacement is carried out on one alternative sequence in the word sequences in the target sequence based on the word sequence language model;

the fourth matching probability corresponding to the first word sequence is used for describing the matching degree of the first word sequence and the local reference text;

and the fourth matching probability corresponding to the second word sequence is used for describing the matching degree of the second word sequence and the local reference text.

8. The speech recognition method according to claim 2, wherein the performing vocabulary replacement on one candidate sequence in the target sequence based on the named entity recognition NER to obtain the edge recognition result comprises:

replacing the vocabulary in one alternative sequence in the target sequence by the first replacement vocabulary in the NER vocabulary based on the NER vocabulary to obtain the edge end recognition result;

and the phoneme matching probability of the phoneme sequence corresponding to the alternative vocabulary and the phoneme sequence of the vocabulary in the alternative sequence is greater than a third preset threshold value.

9. The speech recognition method of claim 8, further comprising:

identifying the local reference text based on the NER to generate an NER word list;

and acquiring phonemes corresponding to the NER vocabulary based on a dictionary and/or character-to-phoneme G2P technology.

10. The speech recognition method of any one of claims 1 to 4 or 6 to 9, wherein the local reference text comprises:

professional information, participant information, activity background information, and activity content of an activity to which the local reference text belongs.

11. The speech recognition method according to any one of claims 1 to 4 or 6 to 9, wherein after obtaining the edge end recognition result, the method further comprises:

and storing the related information of the local reference text into a server based on the professional information of the activity to which the local reference text belongs.

12. A speech recognition method, comprising:

the method comprises the steps that a device side obtains a cloud end identification result or an edge end identification result, wherein the cloud end identification result comprises a target sequence obtained by identifying a target identification object by a cloud end;

and correcting the target sequence based on the local word list of the equipment end to obtain an equipment end identification result.

13. The speech recognition method of claim 12, wherein the modifying the target sequence based on the local vocabulary of the device side to obtain a device side recognition result comprises:

replacing the vocabulary corresponding to the second replacement vocabulary in the target sequence based on the second replacement vocabulary in the local vocabulary to obtain the equipment terminal recognition result;

and the phoneme matching probability of the phoneme sequence of the vocabulary corresponding to the second alternative vocabulary and the phoneme sequence of the second alternative vocabulary is greater than a fourth preset threshold value.

14. The speech recognition method of claim 12, wherein the vocabulary comprises: a pre-set error prone vocabulary and/or a real-time NER vocabulary.

15. The speech recognition method of claim 14, wherein the method further comprises: and acquiring real-time NER vocabularies in the display information corresponding to the target recognition object in real time based on the NER technology.

16. A speech recognition device comprising a memory, a transceiver, a processor:

a memory for storing a computer program; a transceiver for transceiving data under the control of the processor; a processor for reading the computer program in the memory and performing the steps of the speech recognition method according to any of claims 1-11.

17. A speech recognition apparatus comprising a memory, a transceiver, a processor:

a memory for storing a computer program; a transceiver for transceiving data under control of the processor; a processor for reading the computer program in the memory and performing the steps of the speech recognition method according to any of claims 12-15.

18. A speech recognition apparatus, comprising:

the device comprises a first acquisition unit, a second acquisition unit and a third acquisition unit, wherein the first acquisition unit is used for an edge end to acquire a cloud end identification result, and the cloud end identification result comprises at least one target sequence which is acquired by the cloud end aiming at a target identification object;

and the first correcting unit is used for correcting the at least one target sequence based on the local reference text corresponding to the target recognition object to obtain an edge end recognition result.

19. A speech recognition apparatus, comprising:

the second acquisition unit is used for acquiring a cloud end identification result or an edge end identification result by the equipment end, wherein the cloud end identification result comprises a target sequence obtained by identifying a target identification object by a cloud end;

and the second correcting unit is used for correcting the target sequence based on the local word list of the equipment end to obtain an equipment end identification result.

20. A processor-readable storage medium, characterized in that the processor-readable storage medium stores a computer program for causing a processor to perform the method of any one of claims 1 to 15.

Images