TWI636452B - Method and system of voice recognition - Google Patents

Abstract

A speech recognition method and system. The method includes: obtaining a specific type of information text from a predetermined data source; segmenting each obtained information text to obtain a plurality of sentences; and performing word segmentation processing on each sentence to obtain a corresponding word segmentation. Each sentence and the corresponding participle constitute a first mapping corpus; according to each obtained first mapping corpus, a first language model of a preset type is trained, and speech recognition is performed based on the trained first language model. The invention effectively improves the accuracy of speech recognition and effectively reduces the cost of speech recognition.

Description

Speech recognition method and system

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

The language model plays an important role in the speech recognition task. In the existing speech recognition, generally, the labeled dialogue text is used to establish a language model, and the probability of each word is determined by the language model. However, in the prior art, a labeled language is used to build a language model. Because users currently use speech recognition technology in too few scenarios (for example, the more common scenarios are in the areas of voice search and voice control), The type and scope of the corpus that can be collected is too concentrated, which makes this method have the following two disadvantages: one is that it is expensive and expensive to buy; the other is that it is difficult to obtain a sufficient amount of corpus and obtain labeled dialogues The text is difficult, and the timeliness and accuracy of the upgrade cannot be guaranteed, which will affect the training effect and recognition accuracy of the language model, and thus affect the accuracy of speech recognition.

Therefore, how to use the existing corpus resources to effectively improve the accuracy of speech recognition and effectively reduce the cost of speech recognition has become an urgent technical problem.

The main purpose of the present invention is to provide a speech recognition method and system, which aims to effectively improve the accuracy of speech recognition and effectively reduce the cost of speech recognition.

To achieve the above object, the present invention provides a speech recognition method, which includes the following steps: A. obtaining a specific type of information text from a predetermined data source; B. performing sentence segmentation on each obtained information text to obtain several sentences , Performing segmentation processing on each sentence to obtain a corresponding segmentation, and forming a first mapping corpus from each statement and the corresponding segmentation; C. training a first type of first language model based on the obtained first mapping corpus, and based on The trained first language model performs speech recognition.

In an embodiment, the step C is replaced by: training a first language model of a preset type according to each obtained first mapping corpus; and according to each predetermined sample sentence and a second mapping corpus of a corresponding word segmentation Training a second language model of a preset type; mixing the first language model and the second language model trained according to a predetermined model mixing formula to obtain a mixed language model, and based on the obtained mixed language The model performs speech recognition.

In an embodiment, the predetermined model mixing formula is: M = a * M1 + b * M2, where M is a mixed language model, M1 represents a first language model of a preset type, and a represents a preset model. The weight coefficient of M1, M2 represents a second language model of a preset type, and b represents a weight coefficient of a preset model M2.

In an embodiment, the first language model and / or the second language model of the preset type is an n-gram language model, and the training process of the first language model or the second language model of the preset type is as follows: S1. Divide each first mapping corpus or each second mapping corpus into a first scale training set and a second scale verification set; S2, use the training set to train the first language model or the second language S3. Use the validation set to verify the accuracy of the trained first language model or the second language model. If the accuracy is greater than or equal to the preset accuracy, then the training ends, or if the accuracy is less than the preset accuracy , Increase the number of the first mapping corpus or the second mapping corpus and re-execute steps S1, S2, and S3.

In an embodiment, the step of segmenting each segmented sentence includes: when a segmented sentence is selected for segmentation processing, the segmented sentence and a predetermined The word dictionary library performs matching to obtain a first matching result, where the first matching result includes a first number of first phrases and a third number of single words; and the segmented sentence and a predetermined Word dictionary library to match, get To the second matching result, the second matching result includes a second number of second phrases and a fourth number of words; if the first number is equal to the second number and the third number is less than Or equal to the fourth number, the first matching result is used as the word segmentation result of the segmented sentence; if the first number is equal to the second number, and the third number is greater than the first number Four numbers, the second matching result is used as the word segmentation result of the segmented sentence; if the first number is not equal to the second number, and the first number is greater than the second number, then Using the second matching result as the word segmentation result of the segmented sentence; if the first number is not equal to the second number, and the first number is smaller than the second number, the first A matching result is used as the segmentation result of the segmented sentence.

In addition, in order to achieve the above object, the present invention also provides a speech recognition system. The speech recognition system includes: an acquisition module for acquiring a specific type of information text from a predetermined data source; a word segmentation module for acquiring Sentence segmentation of each information text to obtain several sentences, and segmentation processing of each sentence to obtain corresponding segmentation. Each sentence and corresponding segmentation constitute a first mapping corpus; a training recognition module is used to The corpus is mapped, a first language model of a preset type is trained, and speech recognition is performed based on the trained first language model.

In an embodiment, the training recognition module is further configured to: train a first language model of a preset type according to each obtained first mapping corpus; and according to each predetermined sample sentence and a corresponding second word segmentation second Map the corpus to train a second language model of a preset type; mix the first language model and the second language model trained according to a predetermined model mixing formula to obtain a mixed language model, and based on the obtained all language models The mixed language model described above is used for speech recognition.

In an embodiment, the predetermined model mixing formula is: M = a * M1 + b * M2, where M is a mixed language model, M1 represents a first language model of a preset type, and a represents a preset model. Weighting coefficient of M1, M2 represents a preset second language model, and b represents a preset Weighting coefficient of model M2.

In an embodiment, the first language model and / or the second language model of the preset type is an n-gram language model, and the training process of the first language model or the second language model of the preset type is as follows: S1. Divide each first mapping corpus or each second mapping corpus into a first scale training set and a second scale verification set; S2, use the training set to train the first language model or the second language S3. Use the validation set to verify the accuracy of the trained first language model or the second language model. If the accuracy is greater than or equal to the preset accuracy, then the training ends, or if the accuracy is less than the preset accuracy , Increase the number of the first mapping corpus or the second mapping corpus and re-execute steps S1, S2, and S3.

In an embodiment, the word segmentation module is further configured to: when a segmented sentence is selected for word segmentation processing, match the segmented sentence with a predetermined word dictionary library according to a forward maximum matching method, A first matching result is obtained, where the first matching result includes a first number of first phrases and a third number of words; and the segmented sentence is matched with a predetermined word dictionary library according to a reverse maximum matching method, A second matching result is obtained, and the second matching result includes a second number of second phrases and a fourth number of words; if the first number is equal to the second number and the third number is less than Or equal to the fourth number, the first matching result is used as the word segmentation result of the segmented sentence; if the first number is equal to the second number, and the third number is greater than the first number Four numbers, the second matching result is used as the word segmentation result of the segmented sentence; if the first number is not equal to the second number, and the first number is greater than the second number, then Will the first The matching result is used as the word segmentation result of the segmented sentence; if the first number is not equal to the second number, and the first number is less than the second number, the first matching result is used as the The segmentation result of the segmented sentence.

The speech recognition method and system provided by the present invention perform sentence segmentation on a specific type of information text obtained from a predetermined data source, and perform segmentation processing on each segmented sentence to obtain each segmented sentence and the corresponding Segmentation of the first mapping corpus, according to the first A mapping corpus trains a first language model of a preset type, and performs speech recognition based on the trained first language model. Since the corpus resource can be obtained by sentence segmentation and corresponding word segmentation processing of the information text obtained from a plurality of predetermined data sources, and the language model is trained based on the corpus resource, there is no need to obtain labeled dialogue text. Moreover, a sufficient amount of corpus resources can be obtained, and the training effect and recognition accuracy of the language model can be guaranteed, thereby effectively improving the accuracy of speech recognition and effectively reducing the cost of speech recognition.

100‧‧‧ electronic device

10‧‧‧Speech recognition system

11‧‧‧Memory

12‧‧‧ processor

13‧‧‧Display

01‧‧‧Get Module

02‧‧‧Word segmentation module

03‧‧‧Training recognition module

S10‧‧‧step

S20‧‧‧step

S30‧‧‧step

S40‧‧‧step

S50‧‧‧step

S60‧‧‧step

FIG. 1 is a schematic flowchart of a first embodiment of a speech recognition method according to the present invention.

FIG. 2 is a schematic flowchart of a second embodiment of a speech recognition method according to the present invention.

FIG. 3 is a schematic diagram of an operating environment of a voice recognition system 10 according to a preferred embodiment of the present invention.

FIG. 4 is a functional module diagram of a speech recognition system according to an embodiment of the present invention.

In order to make the technical problems, technical solutions and beneficial effects to be more clearly understood by the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

The invention provides a speech recognition method.

Referring to FIG. 1, FIG. 1 is a schematic flowchart of a first embodiment of a speech recognition method according to the present invention.

In the first embodiment, the speech recognition method includes: Step S10, obtaining a specific type of information text from a predetermined data source.

In this embodiment, before training a language model, a specific type of information text (for example, a word) is obtained in real time or from a predetermined number of data sources (for example, Sina Weibo, Baidu Encyclopedia, Wikipedia, Sina News, etc.). Articles and their explanations, news titles, news summaries, Weibo content, etc.). For example, a specific type of information (for example, news headline information, index information, profile information, etc.) can be obtained from a predetermined data source (for example, major news websites, forums, etc.) in real time or periodically through a tool such as a web crawler.

Step S20: Sentence segmentation of each acquired information text to obtain several languages Sentences, perform segmentation processing on each sentence to obtain corresponding segmentation, and each sentence and corresponding segmentation constitute a first mapping corpus.

After obtaining a specific type of each information text from a plurality of predetermined data sources, sentence segmentation can be performed on each of the obtained information texts. For example, each information text can be divided into complete sentences according to punctuation marks. Then, perform segmentation processing on each segmented sentence. For example, segmentation processing can be performed on each segmented sentence by using a string matching segmentation method, such as forward maximum matching, and the character string in a segmented sentence is removed from Word segmentation from left to right; or reverse maximum matching, which splits strings in a segmented sentence from right to left; or shortest path word segmentation, where strings in a segmented sentence require segmentation The number of words is the smallest; or, the two-way maximum matching method, which performs word segmentation matching in both forward and reverse directions. Word segmentation can also be used to perform segmentation processing on each segmented sentence. Word segmentation is a method of machine speech judgment and uses syntactic and semantic information to process ambiguity to segment words. You can also use statistical word segmentation to perform word segmentation on each segmented sentence. From the historical search history of the current user or the historical search history of the general user, according to the statistics of the phrase, it will be counted that some two adjacent words appear more frequently. If there are many, you can use these two adjacent words as phrases to perform segmentation.

After completing the word segmentation processing on the obtained segmented sentences, a first mapping corpus composed of each segmented sentence and the corresponding segmentation can be obtained. By obtaining informational text from multiple predetermined data sources and segmenting the informational text to generate a large number of sentences for word segmentation processing, multiple corpus types can be obtained from a variety of sources, with a wide range and a large number Corpus resources.

In step S30, a first language model of a preset type is trained according to the obtained first mapping corpora, and speech recognition is performed based on the trained first language model.

Based on the first mapping corpus, a first language model of a preset type is trained, and the first language model may be a generative model, an analytical model, an identifying model, and the like. Because the first mapping corpus is obtained from multiple sources, the corpus resources have rich corpus types, a wide range, and a large number. Therefore, the first mapping corpus is used to train the first language model. The training effect is better, and the recognition accuracy of speech recognition based on the trained first language model is higher.

In this embodiment, a specific type of information obtained from a predetermined data source is used. Sentence segmentation of message text and segmentation processing of each segmented sentence to obtain a first mapping corpus of each segmented sentence and corresponding segmentation, and training a first type of first language according to the first mapping corpus Model, and perform speech recognition based on the trained first language model. Since the corpus resource can be obtained by sentence segmentation and corresponding word segmentation processing of the information text obtained from a plurality of predetermined data sources, and the language model is trained based on the corpus resource, there is no need to obtain labeled dialogue text. Moreover, a sufficient amount of corpus resources can be obtained, and the training effect and recognition accuracy of the language model can be guaranteed, thereby effectively improving the accuracy of speech recognition and effectively reducing the cost of speech recognition.

Further, in other embodiments, the above-mentioned step S20 may include: cleaning and denoising each acquired information text. For example, for microblog content, the step of cleaning and denoising includes: deleting user name, id and other information from the microblog content, leaving only the actual content of the microblog; deleting the reposted microblog content, and generally acquired microblogs There is a large number of reposted Weibo content in the content. Repeated reposting of Weibo content will affect the frequency of words. Therefore, the reposted Weibo content must be filtered out. The filtering method is to delete all content that contains "forward" or "http" Weibo content; filtering out special symbols in Weibo content, filtering out all types of symbols in Weibo content; traditional to simplified, there are a large number of traditional characters in Weibo content, using a predetermined Simplified and Traditional correspondence table Convert all traditional characters to simplified characters, and so on.

Sentence segmentation of each information text after cleaning and denoising, for example, use the sentence between two preset types of hyphens “for example, comma, period, exclamation mark, etc.” as one sentence to be segmented, and The segmented sentence is segmented to obtain a mapping corpus of each segmented sentence and a corresponding segmentation (including a phrase and a single word).

As shown in FIG. 2, a second embodiment of the present invention proposes a speech recognition method. Based on the above embodiment, the above step S30 is replaced with: step S40, according to each first mapping corpus obtained, a preset type of training is trained. First language model.

In step S50, a second language model of a preset type is trained according to the second mapping corpus of each predetermined sample sentence and the corresponding segmentation. For example, a number of sample sentences can be determined in advance. For example, a number of sample sentences with the highest frequency or most frequent occurrence can be found from a predetermined data source, and the correct word segmentation (including phrases and words) corresponding to each sample sentence can be determined. according to A second mapping corpus of each predetermined sample sentence and the corresponding word segmentation is used to train a second language model of a preset type.

Step S60: Mix the trained first language model and second language model according to a predetermined model mixing formula to obtain a mixed language model, and perform speech recognition based on the obtained mixed language model. The predetermined model mixing formula may be: M = a * M1 + b * M2, where M is a mixed language model, M1 represents a first language model of a preset type, and a represents a weight coefficient of the preset model M1, M2 represents a second language model of a preset type, and b represents a weight coefficient of the preset model M2.

In this embodiment, on the basis of training the first language model based on the first mapping corpus obtained from multiple data sources, and also according to each predetermined sample sentence and the second mapping corpus of the corresponding word segmentation, The second language model is obtained through training. For example, the predetermined sample sentence may be a preset most commonly used and correct sentence. Therefore, the second language model obtained through training can correctly recognize commonly used speech. The trained first language model and the second language model are mixed according to preset different weight ratios to obtain a mixed language model, and speech recognition is performed based on the obtained mixed language model, which can ensure that the types of speech recognition are rich, The range is wide, and it can ensure the correct recognition of commonly used speech, which further improves the accuracy of speech recognition.

Further, in other embodiments, the training process of the preset type of the first language model or the second language model is as follows: A. Divide each first mapping corpus or each second mapping corpus into a first proportion (For example, 70%) training set and a second proportion (for example, 30%) validation set; B. using the training set to train the first language model or the second language model; C. using the validation set Verify the accuracy of the trained first language model or second language model. If the accuracy is greater than or equal to the preset accuracy, then the training ends, or if the accuracy is less than the preset accuracy, add the first mapping corpus or Steps A, B, and C are performed again for the number of second mapping corpora until the accuracy rate of the trained first language model or second language model is greater than or equal to a preset accuracy rate.

Further, in other embodiments, the first language model of the preset type And / or the second language model is an n-gram language model. The n-gram language model is a language model commonly used in large vocabulary continuous speech recognition. For Chinese, it is called the Chinese Language Model (CLM, Chinese Language Model). The Chinese language model uses the collocation information between adjacent words in the context. When it is necessary to convert consecutive non-spaced pinyin, strokes, or numbers representing letters or strokes into Chinese character strings (that is, sentences), it can calculate the Sentences, thus realizing automatic conversion to Chinese characters, avoiding the problem of recoding of many Chinese characters corresponding to the same pinyin (or stroke string, number string). n-gram is a statistical language model used to predict the nth item based on the first (n-1) items. At the application level, these items can be phonemes (speech recognition applications), characters (input method applications), words (word segmentation applications), or base pairs (gene information), and n-gram models can be generated from large-scale text or audio corpora.

The n-gram language model is based on the assumption that the occurrence of the nth word is only related to the first n-1 words and not related to any other words. The probability of the entire sentence is the product of the probability of each word appearing. It can be obtained by directly counting the number of simultaneous occurrences of n words from the mapping corpus. For a sentence T, assuming T is composed of the word sequences W1, W2, ..., Wn, then the probability of occurrence of sentence T is P (T) = P (W1W2 ... Wn) = P (W1) P (W2 | W1) P (W3 | W1W2) ... P (Wn | W1W2 ... Wn-1). In this embodiment, in order to solve the n-gram with a probability of occurrence 0, in the training of the first language model and / or the second language model, this embodiment adopts a maximum likelihood estimation method, that is, P (Wn | W1W2 ... Wn-1) = C (W1W2 ... Wn) / C (W1W2 ... Wn-1) That is, during the training of the language model, the number of occurrences of the sequence W1W2 ... Wn and The number of times W1W2 ... Wn-1 appears can calculate the probability of the nth word to determine the probability of the corresponding word and realize speech recognition.

Further, in other embodiments, the step of segmenting each segmented sentence in the above step S20 may include: according to a forward maximum matching method, a character string to be processed in each segmented sentence and a predetermined word A dictionary library (for example, the word dictionary library can be a general word dictionary library or an expandable learning word dictionary library) to match to obtain the first matching result; according to the reverse maximum matching method, each sentence that is divided into minutes is waited for. Processed strings and predetermined word dictionary libraries (for example, the word dictionary library can be a general-purpose word dictionary library, or Is an extensible learning dictionary dictionary) to obtain a second matching result. The first matching result includes a first number of first phrases, the second matching result includes a second number of second phrases, and the first matching result includes a third number of words. , The second matching result includes a fourth number of words.

If the first number is equal to the second number, and the third number is less than or equal to the fourth number, output the first matching result (including a phrase and a word) corresponding to the segmented sentence ; If the first number is equal to the second number, and the third number is greater than the fourth number, output the second matching result (including a phrase and a word) corresponding to the segmented sentence; If the first number is not equal to the second number, and the first number is greater than the second number, outputting the second matching result (including a phrase and a single word) corresponding to the segmented sentence; If the first number is not equal to the second number, and the first number is smaller than the second number, the first matching result (including a phrase and a single word) corresponding to the segmented sentence is output.

In this embodiment, a two-way matching method is used to perform word segmentation on each obtained segmented sentence, and forward and reverse simultaneous segmentation matching is performed to analyze the stickiness of the combined content in the strings to be processed in each segmented sentence. Under the circumstances, the phrase is more likely to represent the core point of view information, that is, the core point of view information can be more expressed through the phrase. Therefore, the word segmentation matching with forward and reverse to find the word segmentation result with fewer words and more phrases is used as the word segmentation result of the segmented sentence, thereby improving the accuracy of the word segmentation, and thus ensuring the training effect of the language model. And recognition accuracy.

The invention further provides a speech recognition system. Please refer to FIG. 3, which is a schematic diagram of an operating environment of a voice recognition system 10 according to a preferred embodiment of the present invention.

In this embodiment, the voice recognition system 10 is installed and runs in the electronic device 100. The electronic device 100 may include, but is not limited to, a memory 11, a processor 12, and a display 13. FIG. 3 only shows the electronic device 100 having components 11-13, but it should be understood that it is not required to implement all the illustrated components, and more or fewer components may be implemented instead.

The storage 11 may be an internal storage unit of the electronic device 100 in some embodiments, such as a hard disk or a memory of the electronic device 100. The storage 11 may also be an external storage device of the electronic device 100 in other embodiments, such as a plug-in hard disk and a Smart Media Card (SMC) provided on the electronic device 100. Secure Digital (SD) card, Flash Card, etc. Further, the storage 11 may further include both an internal storage unit of the electronic device 100 and an external storage device. The storage 11 is configured to store application software installed on the electronic device 100 and various types of data, such as codes of the speech recognition system 10. The storage 11 may also be used to temporarily store data that has been output or is to be output.

In some embodiments, the processor 12 may be a central processing unit (CPU), a microprocessor, or other data processing chip, and is configured to run the program code or process data stored in the storage 11. For example, the speech recognition system 10 is executed.

In some embodiments, the display 13 may be an LED display, a liquid crystal display, a touch-type liquid crystal display, an OLED (Organic Light-Emitting Diode) organic touch device, or the like. The display 13 is used to display information processed in the electronic device 100 and to display a visual user interface, such as a menu interface for voice recognition, a result of voice recognition, and the like. The components 11-13 of the electronic device 100 communicate with each other through a system bus.

Please refer to FIG. 4, which is a functional module diagram of a speech recognition system 10 according to a preferred embodiment of the present invention. In this embodiment, the speech recognition system 10 may be divided into one or more modules, and the one or more modules are stored in the storage 11 and are processed by one or more processors. (This embodiment is executed by the processor 12) to complete the present invention. For example, in FIG. 4, the speech recognition system 10 can be divided into an acquisition module 01, a tokenizer module 02, and a training recognition module 03. The module referred to in the present invention refers to a series of computer program instruction segments capable of performing specific functions, and is more suitable for describing the execution process of the speech recognition system 10 in the electronic device 100 than programs. The following description will specifically introduce the functions of the acquisition module 01, word segmentation module 02, and training recognition module 03.

Acquisition module 01, which is used to obtain a specific type of information from a predetermined data source Message text.

In this embodiment, before training a language model, a specific type of information text (for example, a word) is obtained in real time or from a predetermined number of data sources (for example, Sina Weibo, Baidu Encyclopedia, Wikipedia, Sina News, etc.). Articles and their explanations, news titles, news summaries, Weibo content, etc.). For example, a specific type of information (for example, news headline information, index information, profile information, etc.) can be obtained from a predetermined data source (for example, major news websites, forums, etc.) in real time or periodically through a tool such as a web crawler.

The word segmentation module 02 is used for sentence segmentation of each obtained information text to obtain several sentences, and word segmentation processing for each sentence to obtain corresponding word segmentation, and each sentence and the corresponding word segmentation form a first mapping corpus.

After obtaining a specific type of each information text from a plurality of predetermined data sources, sentence segmentation can be performed on each of the obtained information texts. For example, each information text can be divided into complete sentences according to punctuation marks. Then, perform segmentation processing on each segmented sentence. For example, segmentation processing can be performed on each segmented sentence by using a string matching segmentation method, such as forward maximum matching, and the character string in a segmented sentence is removed from Word segmentation from left to right; or reverse maximum matching, which splits strings in a segmented sentence from right to left; or shortest path word segmentation, where strings in a segmented sentence require segmentation The number of words is the smallest; or, the two-way maximum matching method, which performs word segmentation matching in both forward and reverse directions. Word segmentation can also be used to perform segmentation processing on each segmented sentence. Word segmentation is a method of machine speech judgment and uses syntactic and semantic information to process ambiguity to segment words. You can also use statistical word segmentation to perform word segmentation on each segmented sentence. From the historical search history of the current user or the historical search history of the general user, according to the statistics of the phrase, it will be counted that some two adjacent words appear more frequently. If there are many, you can use these two adjacent words as phrases to perform segmentation.

After completing the word segmentation processing on the obtained segmented sentences, a first mapping corpus composed of each segmented sentence and the corresponding segmentation can be obtained. By obtaining informational text from multiple predetermined data sources and segmenting the informational text to generate a large number of sentences for word segmentation processing, multiple corpus types can be obtained from a variety of sources, with a wide range and a large number Corpus resources.

The training recognition module 03 is used for training according to each first mapping corpus obtained. Practice a first language model of a preset type and perform speech recognition based on the trained first language model.

Based on the first mapping corpus, a first language model of a preset type is trained, and the first language model may be a generative model, an analytical model, an identifying model, and the like. Because the first mapping corpus is obtained from multiple sources, the corpus resources have rich corpus types, a wide range, and a large number. Therefore, the first mapping corpus is used to train the first language model. The training effect is better, and the recognition accuracy of speech recognition based on the trained first language model is higher.

In this embodiment, sentence segmentation is performed on a specific type of information text obtained from a predetermined data source, and segmentation processing is performed on each segmented sentence to obtain a first mapping corpus of each segmented sentence and a corresponding segmentation. Training a first language model of a preset type according to the first mapping corpus, and performing speech recognition based on the trained first language model. Since the corpus resource can be obtained by sentence segmentation and corresponding word segmentation processing of the information text obtained from a plurality of predetermined data sources, and the language model is trained based on the corpus resource, there is no need to obtain labeled dialogue text. Moreover, a sufficient amount of corpus resources can be obtained, and the training effect and recognition accuracy of the language model can be guaranteed, thereby effectively improving the accuracy of speech recognition and effectively reducing the cost of speech recognition.

Further, in other embodiments, the word segmentation module 02 is further configured to clean and denoise each acquired information text. For example, for microblog content, the step of cleaning and denoising includes: deleting user name, id and other information from the microblog content, leaving only the actual content of the microblog; deleting the reposted microblog content, and generally acquired microblogs There is a large number of reposted Weibo content in the content. Repeated reposting of Weibo content will affect the frequency of words. Therefore, the reposted Weibo content must be filtered out. The filtering method is to delete all content that contains "forward" or "http" Weibo content; filtering out special symbols in Weibo content, filtering out all types of symbols in Weibo content; traditional to simplified, there are a large number of traditional characters in Weibo content, using a predetermined Simplified and Traditional correspondence table Convert all traditional characters to simplified characters, and so on.

Sentence segmentation of each information text after cleaning and denoising, for example, use the sentence between two preset types of hyphens “for example, comma, period, exclamation mark, etc.” as one sentence to be segmented, and The segmented sentence is segmented to obtain each segmented language. A mapping corpus of sentences and corresponding participles (including phrases and words).

Further, in other embodiments, the training recognition module 03 is further configured to train a first language model of a preset type according to each of the obtained first mapping corpora.

According to the second mapping corpus of each predetermined sample sentence and the corresponding word segmentation, a second language model of a preset type is trained. For example, a number of sample sentences can be determined in advance. For example, a number of sample sentences with the highest frequency or most frequent occurrence can be found from a predetermined data source, and the correct word segmentation (including phrases and words) corresponding to each sample sentence can be determined. According to the second mapping corpus of each predetermined sample sentence and the corresponding word segmentation, a second language model of a preset type is trained.

The trained first language model and the second language model are mixed according to a predetermined model mixing formula to obtain a mixed language model, and speech recognition is performed based on the obtained mixed language model. The predetermined model mixing formula may be: M = a * M1 + b * M2, where M is a mixed language model, M1 represents a first language model of a preset type, and a represents a weight coefficient of the preset model M1, M2 represents a second language model of a preset type, and b represents a weight coefficient of the preset model M2.

In this embodiment, on the basis of training the first language model based on the first mapping corpus obtained from multiple data sources, and also according to each predetermined sample sentence and the second mapping corpus of the corresponding word segmentation, The second language model is obtained through training. For example, the predetermined sample sentence may be a preset most commonly used and correct sentence. Therefore, the second language model obtained through training can correctly recognize commonly used speech. The trained first language model and the second language model are mixed according to preset different weight ratios to obtain a mixed language model, and speech recognition is performed based on the obtained mixed language model, which can ensure that the types of speech recognition are rich, The range is wide, and it can ensure the correct recognition of commonly used speech, which further improves the accuracy of speech recognition.

Further, in other embodiments, the training process of the preset type of the first language model or the second language model is as follows: A. Divide each first mapping corpus or each second mapping corpus into a first proportion (E.g., 70%) training set and a second proportion (e.g., 30%) validation set; B. Use the training set to train the first language model or the second language model; C. Use the validation set to verify the accuracy of the trained first language model or the second language model, if the accuracy rate is greater than or equal to the If the accuracy rate is set, the training is ended, or if the accuracy rate is less than the preset accuracy rate, the number of the first mapping corpus or the second mapping corpus is increased and steps A, B, and C are performed again until the first The accuracy rate of a language model or a second language model is greater than or equal to a preset accuracy rate.

Further, in other embodiments, the first language model and / or the second language model of the preset type is an n-gram language model. The n-gram language model is a language model commonly used in large vocabulary continuous speech recognition. For Chinese, it is called the Chinese Language Model (CLM, Chinese Language Model). The Chinese language model uses the collocation information between adjacent words in the context. When it is necessary to convert consecutive non-spaced pinyin, strokes, or numbers representing letters or strokes into Chinese character strings (that is, sentences), it can calculate the Sentences, thus realizing automatic conversion to Chinese characters, avoiding the problem of recoding of many Chinese characters corresponding to the same pinyin (or stroke string, number string). n-gram is a statistical language model used to predict the nth item based on the first (n-1) items. At the application level, these items can be phonemes (speech recognition applications), characters (input method applications), words (word segmentation applications), or base pairs (gene information), and n-gram models can be generated from large-scale text or audio corpora.

The n-gram language model is based on the assumption that the occurrence of the nth word is only related to the first n-1 words and not related to any other words. The probability of the entire sentence is the product of the probability of each word appearing. It can be obtained by directly counting the number of simultaneous occurrences of n words from the mapping corpus. For a sentence T, assuming T is composed of the word sequences W1, W2, ..., Wn, then the probability of occurrence of sentence T is P (T) = P (W1W2 ... Wn) = P (W1) P (W2 | W1) P (W3 | W1W2) ... P (Wn | W1W2 ... Wn-1). In this embodiment, in order to solve the n-gram with a probability of occurrence 0, in the training of the first language model and / or the second language model, this embodiment adopts a maximum likelihood estimation method, that is, P (Wn | W1W2 ... Wn-1) = C (W1W2 ... Wn) / C (W1W2 ... Wn-1) That is, during the training of the language model, the number of occurrences of the sequence W1W2 ... Wn and The number of occurrences of W1W2 ... Wn-1, you can calculate the occurrence probability of the nth word to determine the Corresponds to the probability of words to achieve speech recognition.

Further, in other embodiments, the above-mentioned word segmentation module 02 is further configured to: according to the forward maximum matching method, the character string to be processed in each segmented sentence and a predetermined word dictionary library (for example, the word dictionary library) (Can be a general word dictionary library or an expandable learning word dictionary library) to perform matching to obtain a first matching result; According to the reverse maximum matching method, the character string to be processed in each sentence and the predetermined word dictionary library (for example, the word dictionary library can be a general word dictionary library or an expandable learning word dictionary library) Perform matching to get the second matching result. The first matching result includes a first number of first phrases, the second matching result includes a second number of second phrases, and the first matching result includes a third number of words. , The second matching result includes a fourth number of words.

If the first number is equal to the second number, and the third number is less than or equal to the fourth number, output the first matching result (including a phrase and a word) corresponding to the segmented sentence ; If the first number is equal to the second number, and the third number is greater than the fourth number, output the second matching result (including a phrase and a word) corresponding to the segmented sentence; If the first number is not equal to the second number, and the first number is greater than the second number, outputting the second matching result (including a phrase and a single word) corresponding to the segmented sentence; If the first number is not equal to the second number, and the first number is smaller than the second number, the first matching result (including a phrase and a single word) corresponding to the segmented sentence is output.

In this embodiment, a two-way matching method is used to perform word segmentation on each obtained segmented sentence, and forward and reverse simultaneous segmentation matching is performed to analyze the stickiness of the combined content in the strings to be processed in each segmented sentence. Under the circumstances, the phrase is more likely to represent the core point of view information, that is, the core point of view information can be more expressed through the phrase. Therefore, the word segmentation matching with forward and reverse to find the word segmentation results with fewer words and more phrases is used as the word segmentation result of the segmented sentence, thereby improving the accuracy of the word segmentation and thus ensuring the language model. Training effect and recognition accuracy.

It should be noted that, in this article, the terms "including", "including" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, It also includes other elements not explicitly listed, or elements inherent to such a process, method, article, or device. Without more restrictions, an element limited by the sentence "including a ..." does not exclude that there are other identical elements in the process, method, article, or device that includes the element.

Through the description of the foregoing implementation manners, those skilled in the art can clearly understand that the methods in the above embodiments can be implemented by means of software plus a necessary universal hardware platform, and of course, can also be implemented by hardware, but in many cases The former is a better implementation. Based on such an understanding, the technical solution of the present invention, in essence, or a part that contributes to the existing technology, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk). ) Includes several instructions for causing a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in the embodiments of the present invention.

The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, and thus do not limit the patent scope of the present invention. The sequence numbers of the foregoing embodiments of the present invention are only for description, and do not represent the superiority or inferiority of the embodiments. In addition, although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than here.

Those skilled in the art can implement the present invention in various modifications without departing from the scope and essence of the present invention. For example, the features of one embodiment can be used in another embodiment to obtain another embodiment. Any modification, equivalent replacement and improvement made within the technical concept of the present invention shall fall within the patent scope of the present invention.

Claims (10)

A speech recognition method, which includes the following steps: A. Obtain a specific type of information text from a predetermined data source; the specific type includes an entry and its explanation, a news title, a news summary, and / or Weibo content; B, Sentence segmentation of each obtained information text according to punctuation marks to obtain several sentences, and segmentation processing of each sentence to obtain corresponding segmentation, and each sentence and corresponding segmentation constitute a first mapping corpus; and C. A mapping corpus, a first language model of a preset type is trained, and speech recognition is performed based on the trained first language model. The speech recognition method according to claim 1, wherein the step C is replaced by: training a first language model of a preset type according to each obtained first mapping corpus; and according to each predetermined sample sentence and corresponding Training the second language corpus of the word segmentation of a preset type; and training the first language model and the second language model according to a predetermined model mixing formula to obtain a mixed language model, Speech recognition is performed based on the obtained mixed language model. The speech recognition method according to item 2 of the claim, wherein the predetermined model mixing formula is: M = a * M1 + b * M2, where M is a mixed language model, and M1 represents the first of the preset types. Language model, a represents the weight coefficient of the preset model M1, M2 represents the second language model of the preset type, and b represents the weight coefficient of the preset model M2. The speech recognition method according to claim 2 or 3, wherein the first language model of the preset type and / or the second language model is an n-gram language model, and the first language model of the preset type Or the second language model training process is as follows: S1, each first mapping corpus or each second mapping corpus is divided into a first proportion of training set and a second proportion of validation set; S2, training using the training set The first language model or the second language model; S3, using the verification set to verify the accuracy of the trained first language model or the second language model, and if the accuracy rate is greater than or equal to a preset accuracy rate, the training ends, Alternatively, if the accuracy rate is less than the preset accuracy rate, increase the number of the first mapping corpus or the second mapping corpus and perform steps S1, S2, and S3 again. The speech recognition method according to claim 1, wherein the step of performing segmentation processing on each segmented sentence includes: when a segmented sentence is selected for segmentation processing, the segmentation is performed according to the forward maximum matching method. The segmented sentence is matched with a predetermined word dictionary library to obtain a first matching result. The first matching result includes a first number of first phrases and a third number of words. The segmented sentence is matched with a predetermined word dictionary library to obtain a second matching result, where the second matching result includes a second number of second phrases and a fourth number of words; if the first number matches If the second number is equal, and the third number is less than or equal to the fourth number, the first matching result is used as a segmentation result of the segmented sentence; if the first number is equal to the first number If the two numbers are equal, and the third number is greater than the fourth number, the second matching result is used as the word segmentation result of the segmented sentence; if the first number is different from the second number And the first number is greater than the second number, the second matching result is used as the word segmentation result of the segmented sentence; if the first number is not equal to the second number, and If the first number is smaller than the second number, the first matching result is used as a segmentation result of the segmented sentence. A speech recognition system includes: an acquisition module for acquiring a specific type of information text from a predetermined data source; the specific type includes an entry and its explanation, a news title, a news summary, and / or Weibo content; word segmentation A module for segmenting each obtained information text according to punctuation marks to obtain several sentences, and performing segmentation processing on each sentence to obtain corresponding segmentations, and each sentence and the corresponding segmentation constitute a first mapping corpus; training a recognition module A group for training a first language model of a preset type according to each obtained first mapping corpus, and performing speech recognition based on the trained first language model. The speech recognition system according to claim 6, wherein the training recognition module is further configured to: train a first language model of a preset type according to each obtained first mapping corpus; and according to each predetermined sample The second mapping corpus of the sentence and the corresponding word segmentation is used to train a second language model of a preset type; according to a predetermined model mixing formula, the first language model and the second language model trained are mixed to obtain a mixture A language model, and perform speech recognition based on the obtained mixed language model. The speech recognition system according to claim 7, wherein the predetermined model mixing formula is: M = a * M1 + b * M2, where M is a mixed language model, and M1 represents the first of the preset types. Language model, a represents the weight coefficient of the preset model M1, M2 represents the second language model of the preset type, and b represents the weight coefficient of the preset model M2. The speech recognition system according to claim 7 or 8, wherein the first language model and / or the second language model of the preset type is an n-gram language model, and the first language model of the preset type Or the second language model training process is as follows: S1, each first mapping corpus or each second mapping corpus is divided into a first proportion of training set and a second proportion of validation set; S2, training using the training set The first language model or the second language model; S3, using the verification set to verify the accuracy of the trained first language model or the second language model, and if the accuracy rate is greater than or equal to a preset accuracy rate, the training ends, Alternatively, if the accuracy rate is less than the preset accuracy rate, increase the number of the first mapping corpus or the second mapping corpus and perform steps S1, S2, and S3 again. The speech recognition system according to item 6, wherein the segmentation module is further configured to: when a segmented sentence is selected for segmentation processing, the segmented sentence is compared with The determined word dictionary library is matched to obtain a first matching result, where the first matching result includes a first number of first phrases and a third number of single words; the segmented sentence and the The determined word dictionary library is matched to obtain a second matching result, where the second matching result includes a second number of second phrases and a fourth number of single words; if the first number is equal to the second number , And the third number is less than or equal to the fourth number, the first matching result is used as the segmentation result of the segmented sentence; if the first number is equal to the second number, and If the third quantity is greater than the fourth quantity, the second matching result is used as the segmentation result of the segmented sentence; if the first quantity is not equal to the second quantity, and the first quantity Greater than If the second quantity is described, the second matching result is used as the segmentation result of the segmented sentence; if the first quantity is not equal to the second quantity, and the first quantity is smaller than the second quantity , Then use the first matching result as the segmentation result of the segmented sentence.