A method for Chinese speech enhancement recognition and text correction correction

Abstract

The application belongs to the field of speech and text processing, and particularly relates to a Chinese speech enhancement recognition and text error correction method, which comprises the following steps: preprocessing the audio to be recognized, extracting features through a voiceprint model, and establishing an initial coarse dialect identification model; establishing an initial network model to train the initial coarse dialect identification model to obtain a dialect identification model; determining an error correction candidate word segmentation set based on an N-gram language model; and outputting the text after error correction and correction through a Bigram 2-gram language model and the N-gram language model. The application preprocesses the audio to be recognized, reduces speech recognition interference factors, improves the recognition performance by adopting a GMM-SVM model, has a faster and better training fitting effect by adopting a combination model of a GMM-UBM model and an LSTM model to establish an initial network model, effectively reduces the error rate by processing the text through an N-gram language model and a Bigram 2-gram language model, and optimizes the result of converting the audio to be recognized into text information.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of speech and text processing, in particular to a Chinese speech enhancement recognition and text error correction method. BACKGROUND

[0002] Speech recognition is an interdisciplinary subfield of computational linguistics that combines knowledge and research in linguistics, computer science, and electrical engineering to develop methods and techniques for the recognition and translation of spoken language into text. With the rise of online education and the continuous development of entertainment culture and various media, there is a sharp increase in demand for real-time or post-production Chinese subtitles for online videos on various platforms. Considering the large span, wide geographical coverage, easy confusion, speech speed, accent, and audio distortion of Chinese speech, it is obvious that it is difficult to convert Chinese speech into correct text.

[0003] Traditionally, post-speech subtitle production and text correction are usually done manually, but with the explosion of information and exponential growth of platform data throughput, manual methods are obviously not suitable for the current era. As for machine recognition of speech and conversion of text and error correction, Huawei and Xunfei have made a lot of efforts and attempts and have launched several software applications, but the performance of Chinese speech recognition and text correction in different regional environments is not satisfactory. In addition to errors in speech-to-text technology, errors may occur when users input text through input methods. These errors will affect the readability of the text and are not conducive to human and machine understanding, especially when there are mixed dialectal pronunciations. If these errors are not processed, they will be passed on to subsequent links, which will not be conducive to information transmission and will cause information delay, affecting various social practice activities. SUMMARY

[0004] The purpose of the present application is to provide a Chinese speech enhancement recognition and text error correction method, which solves the problem of possible errors in converting Chinese speech into text and effectively optimizes the results of post-speech text.

[0005] To achieve the above application purposes, the technical solution adopted by the present application is as follows:

[0006] A Chinese speech enhancement recognition and text error correction method, the steps of which are as follows:

[0007] S1: Preprocess the audio to be recognized to obtain preprocessed audio, and establish an initial rough dialect identification model based on a voiceprint model GMM-SVM;

[0008] S2: establish an initial network model and train the initial rough dialect identification model to obtain a dialect identification model; match the preprocessed audio with a standard language library to determine an authentication result according to a matching state, and output corrected audio information, wherein the standard language library comprises a pronunciation rule library, a user common language library, and a syllable knowledge base;

[0009] S3: convert the corrected audio information into speech text, perform word segmentation based on a dictionary knowledge base, and determine a candidate error correction word set through an N-gram language model;

[0010] S4: further simplify the N-gram language model to obtain a Bigram 2-gram language model, and output the speech text after error correction and correction using the N-gram language model and the Bigram 2-gram language model.

[0011] Further as a preferred scheme of the present application, step S1 specifically comprises the following steps:

[0012] S101: utilize an audio synthesis model Tacotron 2.0 to perform noise reduction, repair and completion, and blur enhancement processing on the audio to be recognized to obtain preprocessed audio;

[0013] S102: adopt a voiceprint model GMM-SVM model, and construct a Gaussian hyper vector for each Gaussian component mean value in the GMM as a sample of the SVM;

[0014] S103: on the basis of the voiceprint model, adopt sample voice data to perform machine learning to obtain an initial rough dialect identification model, wherein an input and output function of a prediction neuron of the GMM-SVM is:

[0015] W i (k+i)=W i (k)+λd i (k+1)

[0016] λ>0,i=1,2……N

[0017]

[0018]

[0019]

[0020]

[0021] wherein W is a connection weight value of the prediction neuron, N is the number of input layer units, k is a loop variable, λ is a learning rate, d i is a distance deviation degree of the connection weight value, and x(p) is a true value. The x(n), x(n+1), …, x(m) represent the same set of sample voice data for a predicted value, and the δ p The E represents the difference between the real value and the predicted value. p The E represents the total error of the whole sample space; given a set of initial values w i (0), the calculation is performed until the convergence condition, i.e., the E p is minimized, and the initial coarse dialect identification model is obtained.

[0022] Further, as a preferred scheme of the present application, the step S2 specifically comprises the following steps:

[0023] S201: An initial network model is established by using a combined model of the GMM-UBM model and the LSTM model, and the sample voice data is input into the initial network model for recognition training.

[0024] S202: The sample voice data is labeled with a corresponding dialect type, and the labeled sample voice data is input into the initial network model for deep learning training; in the training process, the model parameters of the initial network model are adjusted until the dialect type output by the initial network model matches the dialect type labeled by the sample voice data, the parameters of the initial coarse dialect identification model are adjusted according to the model parameters of the initial network model at this time, and a dialect identification model of the type of dialect is obtained.

[0025] S203: The preprocessed audio is input into the dialect identification model corresponding to each type of dialect, and the predicted value of each type of dialect corresponding to the preprocessed audio is obtained; the maximum value of each type of dialect predicted value is compared, and the dialect type corresponding to the maximum value is the dialect type of the audio to be identified; the preprocessed audio is matched with a standard language library, the authentication result is determined according to the matching state, and the corrected audio information is output.

[0026] Further, as a preferred scheme of the present application, the dictionary knowledge base in the step S3 comprises a basic word library, a sentence logic library, a context library, and a semantic grammar database; and the step S3 specifically comprises the following contents:

[0027] The audio information corrected in step S2 is converted into speech text, the speech text is preprocessed to obtain a text input string, and initial error correction is performed on the text input string. The initial error correction process specifically comprises: performing word segmentation processing on the text input string according to a segmentation model in the N-gram language model to obtain N1 sequences and N2 sequences; calculating the probability of each keyword in the N1 sequences and the N2 sequences being segmented by the N-gram language model, and then performing secondary word segmentation processing on the text data to be corrected according to the probability to obtain N3 sequences; calculating the word frequency of the keywords in each sentence in the N3 sequences, and judging whether the word frequency reaches a preset word frequency error correction threshold α1, wherein the word frequency is obtained through a basic word library; if the word frequency error correction threshold α1 is reached, it is considered that the keyword does not need to be corrected, and if the word frequency error correction threshold α1 is not reached, the first keyword corresponding to the first keyword is converted into a pinyin feature through the N-gram language model, wherein the first keyword is the keyword with the maximum word frequency; according to the pinyin feature, a recall model is used to recall candidate words corresponding to the first keyword, and other words are selected from the candidate words to correct the first keyword.

[0028] Further, as a preferred scheme of the present application, in step S3, the process of obtaining the error correction candidate segmented word set specifically comprises: the formula for calculating the sentence score probability of each sentence in the N3 sequence in the N-gram language model is as follows:

[0029]

[0030] wherein p(s) represents the sentence score probability, w i represents the weight ratio of the i-th keyword in the sentence, i=1,...,n;

[0031] The error is checked by using the continuous relationship between the words. If the p(s) between w n-1 and w i or the p(s) between w i-1 and w i corresponding to w i-1 corresponding to w i+1 corresponding to w i is correct or not.

[0032] The error correction candidate segmented word set is determined according to the judgment result of whether the keyword is correct or not, the error correction candidate segmented word set is cross-combined to form a plurality of error correction candidate sentences, and the plurality of error correction candidate sentences constitute a confusion set. Considering the existence of multi-phonetic words in speech recognition, the initial confusion set adopts a pinyin confusion set.

[0033] Further, as a preferred scheme of the present application, step S4 specifically comprises the following contents:

[0034] The formula for calculating the score probability in the bigram 2-gram language model is:

[0035] p(s) = p(w1) p(w2|w1) ··· p(w1|w1···w n-1 )

[0036] The language model is constructed by calculating the maximum likelihood estimate, wherein p(w n |w n-1 ) is calculated by the formula:

[0037] p(w n |w n-1 ) = count(w n ,w n-1 ) / count(w n-1 )

[0038] count(w n-1 ) represents the number of times that w n-1 occurs in the dictionary knowledge base, and w n-1 represents the weight ratio of the n-1th keyword in the sentence;

[0039] The mutual information between two keywords is calculated by the formula:

[0040]

[0041] wherein b i represents the ith keyword in the sentence in the N3 sequence, b i-1 represents the i-1th keyword in the sentence in the N3 sequence, I(b i-1 b i ) is the mutual information between two consecutive keywords, p(b i |b i-1 ) represents the score probability of the keyword b i under the premise that b i-1 occurs, and p(b i ) represents the score probability of the single keyword b i in the sentence, and p(b i-1 ) represents the score probability of the single keyword b i-1 in the sentence;

[0042] If I(b i-1 b i ) satisfies the mutual information threshold α2, it is considered that the keyword b i is reliably connected with the keyword b i-1 , that is, the keyword b i is correct.

[0043] Further, as a preferred scheme of the present application, the step of correcting the speech text in step S4 comprises the following contents:

[0044] S401: Determine the input string X from the text input string. i Is it a single character?

[0045] If it is a single character, then X i Input into the N-gram model and determine X i If the sentence score probability is greater than the word frequency correction threshold α1, and if it is less than or equal to the word frequency correction threshold α1, then X will be... i Mark as an error; if it exceeds the word frequency correction threshold α1, then judge X. i Does it exist in the semantic syntax database? If it does, then X i Output the correct answer; if X does not exist... i Mark as error;

[0046] If X i Not a single word, but X i Calculate X in the input Bigram2 model i The mutual information between keywords is considered. If the mutual information between keywords is less than or equal to the mutual information threshold α2, then X... i The keyword is marked as incorrect; if the mutual information of the keyword is greater than the mutual information threshold α2, then X is judged. i Does the keyword exist in the semantic syntax database? If it does, then X... i Output the correct value; if it does not exist, output X. i The keyword in this text is marked as incorrect.

[0047] S402: Determine if an X is marked as incorrect. i If the keyword in X exists in the error correction candidate word segmentation set, then X will be... i Replace the keyword with a similar group from the pinyin confusion set, and return to step S401; if it does not exist, use the dictionary knowledge base to search for X. i The relevance of this keyword is calculated using the following formula:

[0048] C word (b m ,b n ) = SS(b m ,b n )+CC(b n ,b k )

[0049] Where CC(b) m ,b k ) indicates contextual relevance, SS(b m ,b n ) represents semantic relevance, b m For X i Chinese keywords, bn b is an instance word in the dictionary knowledge base k b is a core word in the text input string

[0050] Calculate CC(b m ,b k ) whether greater than the word relevance threshold α3, if less than or equal to the word relevance threshold α3, b m is discarded as abnormal data, if greater than the word relevance threshold α3, b m is manually corrected and output X i .

[0051] Compared with the prior art, the beneficial effects of the present application are:

[0052] 1. A Chinese speech enhancement recognition method is proposed, which pre-enhances the to-be-recognized audio locally to optimize the recognition effect of the audio;

[0053] 2. The model GMM-SVM is used instead of GMM-UBM to utilize the nonlinear classification ability of the SVM of the kernel function, which greatly improves the recognition performance on the basis of the original;

[0054] 3. The combination model of the GMM-UBM model and the LSTM model is used to establish an initial network model, which has better and faster training fitting effect;

[0055] 4. The N-gram language model and the Bigram 2-gram language model are used in parallel in the text correction module, which greatly plays the maximum advantages of the two and effectively reduces the error rate. BRIEF DESCRIPTION OF DRAWINGS

[0056] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the principles of the present application, and do not constitute a limitation of the present application.

[0057] Figure 1 is the schematic diagram of the method of the present application;

[0058] Figure 2 is the flowchart of the dialect identification model of the present application;

[0059] Figure 3 is the error correction method flowchart of the embodiment provided by the present application;

[0060] Figure 4 is the GMM-SVM model schematic diagram of the embodiment provided by the present application;

[0061] Figure 5 is the audio synthesis model schematic diagram of the embodiment provided by the present application. DETAILED DESCRIPTION

[0062] In order to make the above objectives, characteristics and advantages of the present application more apparent, concrete embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0063] As shown in the drawings, the present application provides a Chinese speech enhancement recognition and text correction method, comprising the following steps: Figure 1

[0064] Step S1: pre-processing the audio to be recognized to obtain pre-processed audio, and establishing an initial rough dialect identification model based on GMM-SVM. In S1, the following steps are specifically included:

[0065] S101: using the audio synthesis model Tacotron2.0 to perform noise reduction, repair and completion, and fuzzy enhancement processing on the audio to be recognized to obtain pre-processed audio, as shown in Figure 5 ;

[0066] S102: using the GMM-SVM voiceprint model, constructing a Gaussian hyper vector for each Gaussian component mean value in GMM as a sample of SVM, as shown in Figure 4 ;

[0067] S103: based on the voiceprint model, using sample speech data for machine learning to obtain an initial rough dialect identification model, wherein the input and output function of the prediction neuron of GMM-SVM is:

[0068] W i (k+i)=W i (k)+λd i (k+1)

[0069] λ>0, i=1, 2…N

[0070]

[0071]

[0072]

[0073]

[0074] Wherein, W is the connection weight of the prediction neuron, N is the number of input layer units, k is the loop variable, λ is the learning rate, d i is the distance deviation of the connection weight, x(p) is the true value, is the predicted value, x(n), x(n+1)…x(m) represent the same group of sample speech data, δ p is the difference between the true value and the predicted value, E p ​Total error representing the whole sample space; given a set of initial values w i (0) Perform calculation until convergence condition is met, i.e. E p The initial rough dialect identification model is obtained when the minimum is reached

[0075] Step S2: As shown in the figure, an initial network model is established and the initial rough dialect identification model is trained to obtain a dialect identification model; the preprocessed audio is matched with a standard language library to determine an authentication result according to a matching state, and corrected audio information is output, wherein the standard language library includes a pronunciation rule library, a user common language library, and a syllable knowledge base. In step S2, the following steps are specifically included: Figure 2

[0076] S201: An initial network model is established by using a combination model of a GMM-UBM model and an LSTM model, sample voice data is input into the initial network model for recognition training, GMM;

[0077] S202: The sample voice data is labeled with a corresponding dialect type, and the labeled sample voice data is input into the initial network model for deep learning training; in the training process, the model parameters of the initial network model are adjusted until the dialect type output by the initial network model matches the dialect type labeled by the sample voice data, the initial rough dialect identification model parameters are adjusted according to the initial network model parameters at this time, and a dialect identification model of the type of dialect is obtained;

[0078] S203: The preprocessed audio is input into the dialect identification model corresponding to each type of dialect to obtain a prediction value of each type of dialect corresponding to the preprocessed audio; the maximum value of each type of dialect prediction value obtained is compared, and the dialect type corresponding to the maximum value is the dialect type of the audio to be recognized; the preprocessed audio is matched with a standard language library to determine an authentication result according to a matching state, and corrected audio information is output.

[0079] Step S3: The corrected audio information is converted into voice text, and word segmentation is performed based on a dictionary knowledge base, and an error correction candidate word segmentation set is determined through an N-gram language model, wherein the dictionary knowledge base includes a basic word library, a sentence logic library, a context library, and a semantic grammar database. Step S3 specifically includes the following content:

[0080] S301: The corrected audio information in step S2 is converted into voice text, the voice text is preprocessed to obtain a text input string, the text input string is initially corrected, the text input string is segmented according to a segmentation model in the N-gram language model to obtain N1 sequence and N2 sequence; the probability of each keyword in the N1 sequence and the N2 sequence being segmented by the N-gram language model is calculated, and the text data to be corrected is further segmented according to the probability to obtain N3 sequence;​

[0081] S302: Calculate the word frequency of each keyword in the N3 sequence, and determine whether the word frequency reaches a preset word frequency error correction threshold a1, wherein the word frequency is obtained through a basic word library; if the word frequency error correction threshold a1 is reached, it is considered that the keyword does not need to be corrected; if the word frequency error correction threshold a1 is not reached, the first keyword corresponding to the word frequency is converted into a pinyin feature through an N-gram language model, wherein the first keyword is the keyword with the largest word frequency; according to the pinyin feature, a recall model is used to recall candidate words corresponding to the first keyword, and other words are selected from the candidate words to correct the first keyword.

[0082] S303: Obtain a correction candidate segmentation set, and the specific process is as follows: the formula for calculating the sentence score probability of each sentence in the N3 sequence in the N-gram language model is:

[0083]

[0084] Wherein, p(s) represents the sentence score probability, w i represents the weight ratio of the i-th keyword in the sentence, i=1,...,n;

[0085] Use the continuous relationship between words to check errors, if p(s) between w n-1 and w i or p(s) between w i-1 and w i satisfies the word frequency error correction threshold a1, it can be determined that w i-1 corresponding keyword and w i+1 corresponding keyword, w i corresponding keyword is connected, and then whether w n-1 corresponding keyword is wrong is determined; according to the judgment result of whether the keyword is wrong, a correction candidate segmentation set is determined, and the correction candidate segmentation set is cross combined to form a plurality of correction candidate sentences, and the plurality of correction candidate sentences constitute a confusion set; considering the existence of homophonic characters in speech recognition, the initial confusion set adopts a pinyin confusion set.

[0086] Step S4: Simplify the N-gram language model to obtain a Bigram 2-gram language model, and output after correcting and correcting the speech text using the N-gram language model and the Bigram 2-gram language model.

[0087] The standard for the Bigram 2-gram language model to determine whether the text is correct is: the formula for calculating the sentence score probability in the Bigram 2-gram language model is:

[0088] p(s)=p(w1)p(w2|w1)···p(w1|w1···w n-1 )

[0089] A language model is constructed by calculating the maximum likelihood estimate, where p(w n |w n-1 The calculation formula is:

[0090] p(w n |w n-1 ) = count(w n ,w n-1 ) / count(w n-1 )

[0091] count(w n-1 ) indicates w n-1 The number of times it appears in the dictionary knowledge base, w n-1 This represents the weight ratio of the (n-1)th keyword in the sentence; the mutual information between two keywords is calculated using the following formula:

[0092]

[0093] Among them, b i Let b represent the i-th keyword in the sentence of the N3 sequence. i-1 I(b) represents the (i-1)th keyword in the sentence of the N3 sequence. i-1 b i p(b) represents the mutual information between two consecutive keywords. i |b i-1 ) indicates the keyword b i In b i-1 The probability of scoring given that the event occurs, p(b) i ) represents a single keyword b i The probability of scoring in the sentence, p(b) i-1 ) represents a single keyword b i-1 The probability of scoring in a sentence;

[0094] If I(b) i-1 b i If the mutual information threshold α2 is satisfied, then keyword b is considered to be... i With keyword b i-1 Reliable continuation between them, i.e., keyword b i correct.

[0095] like Figure 3 As shown, the specific steps for text correction in step S4 are as follows:

[0096] S401: Determine the input string X from the text input string. i Is it a single character?

[0097] If it is a single character, then X i Input into the N-gram model and determine X iis greater than the word frequency error correction threshold α1, if less than or equal to the word frequency error correction threshold α1, X i is marked as an error; if greater than the word frequency error correction threshold α1, whether X i exists in the semantic grammar database, if exists, X i is correct and output, if X i does not exist, X i is marked as an error;

[0098] If X i is not a single word, X i is input into the Bigram2 model to calculate the inter-word mutual information of the key word in X i , if the key word mutual information is less than or equal to the mutual information threshold α2, X i in which the key word is marked as an error; if the key word mutual information is greater than the mutual information threshold α2, whether X i in which the key word exists in the semantic grammar database, if exists, X i is correct and output, if X i does not exist, X i in which the key word is marked as an error;

[0099] S402: Determine whether the key word in X i marked as an error exists in the error correction candidate segmentation set, if exists, X word in which the key word is replaced with the similar group in the pinyin confusion set, and returns to step S401; if not, X m in which the key word is calculated by using the word correlation degree, and the word correlation degree calculation formula is as follows:

[0100] C n (b m ,b n )=SS(b n ,b k )

[0101] Wherein CC(b m ,b k ) represents the context correlation degree, SS(b m ,b n ) represents the semantic correlation degree, b m is the key word in X i , b n is the instance word in the dictionary knowledge base, and b k is the core word of the text input string;

[0102] Calculate whether CC(b m ,b k ) is greater than the word correlation degree threshold α3, if less than or equal to the word correlation degree threshold α3, bm As abnormal data rejection, if greater than the word relevance threshold a3, b m Output X after manual correction i .

[0103] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can make modifications to the technical solutions described in the foregoing embodiments, or make equivalent replacements to part of the technical features, any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for enhancing Chinese speech recognition and correcting text errors, characterized in that, Includes the following steps: S1: Preprocess the audio to be recognized to obtain the preprocessed audio, and establish an initial coarse dialect identification model based on GMM-SVM; S2: Establish an initial network model and train the initial coarse dialect identification model to obtain a dialect identification model; match the preprocessed audio with a standard linguistic database, determine the authentication result based on the matching status, and output the corrected audio information, wherein the standard linguistic database includes a pronunciation rule database, a user common language database, and a syllable knowledge database. S3: Convert the corrected audio information into speech text, perform word segmentation based on the dictionary knowledge base, and determine the set of candidate word segments for error correction through the N-gram language model; S4: The N-gram language model is further simplified to obtain the Bigram2 language model. The N-gram language model and the Bigram2 language model are used to correct errors in the speech text before outputting it. Step S1 specifically includes the following steps: S101: The audio to be identified is processed by noise reduction, repair and completion, and blur enhancement using the audio synthesis model Tacotron2.0 to obtain the preprocessed audio; S102: Using the GMM-SVM model, a Gaussian hypervector is constructed from the mean of each Gaussian component in the GMM as a sample for the SVM; S103: Based on the GMM-SVM model, machine learning is performed using sample speech data to obtain an initial coarse dialect identification model. The input-output function of the prediction neuron in the GMM-SVM model is: ； in, To predict the connection weights of neurons, The number of input layer units, For loop variable, For learning rate, The distance deviation of the connection weights. For the true value, For predicted values, This represents the same set of sample speech data. The difference between the actual value and the predicted value. Represents the total error across the entire sample space; given a set of initial values Perform calculations until the convergence condition is met, i.e. When the minimum value is reached, the initial rough dialect identification model is obtained; Step S2 specifically includes the following steps: S201: An initial network model is established using a combination of GMM-UBM and LSTM models, and sample speech data is input into the initial network model for recognition training. S202: Label the dialect type corresponding to the sample speech data, input the labeled sample speech data into the initial network model for deep learning training; during the training process, adjust the model parameters of the initial network model until the dialect type output by the initial network model matches the dialect type labeled in the sample speech data, and adjust the parameters of the initial coarse dialect identification model according to the parameters of the initial network model at this time to obtain the dialect identification model of that type of dialect; S203: Input the preprocessed audio into the dialect identification model corresponding to each type of dialect to obtain the predicted value of each type of dialect corresponding to the preprocessed audio; compare the maximum value of each type of dialect prediction value, and the dialect type corresponding to the maximum value is the dialect type of the audio to be identified; match the preprocessed audio with the standard lexicon, determine the authentication result according to the matching status, and output the corrected audio information.

2. The Chinese speech enhancement recognition and text error correction method according to claim 1, characterized in that, The dictionary knowledge base in step S3 includes a basic word database, a sentence logic database, a context database, and a semantic grammar database; step S3 specifically includes the following: The corrected audio information in step S2 is converted into speech text. The speech text is preprocessed to obtain a text input string. Initial error correction is performed on the text input string. Specifically, the initial error correction process involves segmenting the text input string into words according to the segmentation model in the N-gram language model. sequence sum Sequence; Calculate the sequence sum The probability of each keyword in the sequence being segmented by the N-gram language model is used to perform secondary word segmentation on the text data to be corrected, based on the probability. Sequence; Calculate the The word frequency of keywords in each sentence of the sequence is used to determine whether the word frequency reaches a preset word frequency correction threshold. The word frequency is obtained through a basic word lexicon; if the word frequency error correction threshold is reached... If the keyword frequency does not reach the specified error correction threshold, then it is considered that no error correction is needed. The corresponding first keyword is converted into pinyin features through the N-gram language model, where the first keyword is the keyword with the highest word frequency; based on the pinyin features, a recall model is used to recall candidate words corresponding to the first keyword, and other words are selected from the candidate words to correct the first keyword.

3. The Chinese speech enhancement recognition and text error correction method according to claim 2, characterized in that, In step S3, the process of obtaining the error correction candidate word segmentation set specifically includes: the N-gram speech model... The formula for the score probability of each sentence in the sequence is: ； in, This represents the probability of a sentence score. Indicates the first word in the sentence The weight ratio of each keyword, ; Error detection is performed using the continuity between words. and Between or and Between Satisfying the term frequency error correction threshold That is enough to determine Corresponding keywords and Corresponding keywords The corresponding keyword sequence is then used to determine... Are the corresponding keywords incorrect? Based on the judgment result of whether the keyword is incorrect, a set of candidate word segments for error correction is determined. The set of candidate word segments for error correction is cross-combined to form several candidate sentences for error correction. The several candidate sentences for error correction constitute a confusion set. Considering that there are polyphonic characters in speech recognition, the initial confusion set adopts the Pinyin confusion set.

4. The Chinese speech enhancement recognition and text error correction method according to claim 3, characterized in that, Step S4 specifically includes the following: The formula for calculating sentence score probabilities using the Bigram2 meta-language model is: ； A language model is constructed by calculating the maximum likelihood estimate, where The calculation formula is: ； express The number of times it appears in the dictionary knowledge base Indicates the first word in the sentence The weight ratio of each keyword; The formula for calculating the mutual information between two keywords is as follows: ； in, express Sentence number in the sequence One keyword, express Sentence number in the sequence One keyword, The mutual information of two consecutive keywords. Keywords The probability of scoring, Keywords The probability of scoring in a sentence. Keywords The probability of scoring in a sentence; like Meet mutual information threshold Then the keywords are considered With keywords Reliable connection between them, i.e., keywords correct.

5. The Chinese speech enhancement recognition and text error correction method according to claim 4, characterized in that, In step S4, the step of correcting errors in the spoken text includes the following: S401: Determine the input string to be searched in the text input string. Is it a single character? If it is a single character, Input into the N-gram model and determine Does the sentence score probability exceed the word frequency error correction threshold? If it is less than or equal to the word frequency error correction threshold ,Will Mark as an error; if it exceeds the word frequency correction threshold. Then judge Does it exist in the semantic syntax database? If it does, then Output if correct, otherwise output if not. Mark as error; like Not a single word, but a general The calculation is performed by inputting the Bigram2 model. The mutual information between keywords is determined by the mutual information threshold. ,Will The keyword in this text is marked as incorrect. If the keyword mutual information is greater than the mutual information threshold Then judge Does the keyword exist in the semantic grammar database? If it does, then... Output if correct, otherwise output The keyword in this text is marked as incorrect. S402: Determine if the mark is incorrect Does the keyword in the text exist in the error correction candidate word segmentation set? If it does, then... Replace the keyword with a similar group from the Pinyin confusion set, and return to step S401; if it does not exist, use the dictionary knowledge base to... The relevance of this keyword is calculated using the following formula: ； in Indicates the relevance of the context. Indicates semantic relevance. for Chinese keywords, For example words in the dictionary knowledge base, The core words in the text input string; calculate Is it greater than the word relevance threshold? If it is less than or equal to the word relevance threshold ,Will Data that is considered abnormal and is discarded if it exceeds the aforementioned word relevance threshold. ,right Output after manual error correction .

Images