CN109461436B - A method and system for correcting pronunciation errors in speech recognition - Google Patents

Abstract

The invention provides a method and a system for correcting pronunciation errors in speech recognition, wherein the method comprises the following steps: establishing a mapping table between a standard acoustic model and an error acoustic model corresponding to pronounce error-prone characters; acquiring user voice information; recognizing the user voice information, and when the pronunciation-prone wrong character is contained in the voice information, extracting an audio segment corresponding to a word containing the pronunciation-prone wrong character in the user voice information; and when the audio frequency segment is matched with the voice audio frequency matching result in the error acoustic model, prompting a user that the pronunciation is easy to miss and the pronunciation is wrong, and outputting the corresponding voice audio frequency in the standard acoustic model according to the mapping table. The invention prompts and outputs corresponding correct audio when recognizing that the pronunciation of the pronouncing error-prone character of the user is wrong by establishing a mapping table between the standard acoustic model and the error acoustic model.

Description

A method and system for correcting pronunciation errors in speech recognition

technical field

The present invention relates to the technical field of speech recognition, in particular to a method and system for correcting pronunciation errors in speech recognition.

Background technique

With the rapid development of the Internet, people's lives are becoming more and more intelligent. As one of the mainstream communication applications of human-computer interaction in intelligent terminals, voice interaction is increasingly favored by users. The intelligent terminal takes corresponding measures based on the voice input by the user, so the accuracy of the voice input by the user through the terminal terminal seriously affects the feedback made by the intelligent terminal.

There are a large number of polyphonic characters and close-shaped characters in Chinese characters. For some users, it is difficult to distinguish uncommon polyphonic characters and close-shaped characters. The common pronunciation of users is inherently wrong.

In addition, for primary school students, they are still in the process of learning, especially in the case of a small amount of literacy, often mispronounce words containing polyphonic characters or mispronounce words with similar shapes. In this case, the intelligent terminal recognizes the pronunciation. , it will lead to recognition errors, unable to give the correct results to be queried or the corresponding accurate feedback. Therefore, there is a need for a method and system for correcting pronunciation errors in speech recognition to solve the above problems.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method and system for correcting pronunciation errors in speech recognition, so as to realize that by establishing a mapping table between the standard acoustic model and the wrong acoustic model, prompting and outputting the corresponding typos when recognizing that the user's pronunciation is prone to mispronunciation correct audio.

The technical scheme provided by the present invention is as follows:

The present invention provides a method for correcting pronunciation errors of speech recognition, which is characterized in that, comprising:

Establish a mapping table between the standard acoustic model and the wrong acoustic model corresponding to the pronunciation-prone words;

Obtain user voice information;

Identifying the user's voice information, when the voice information contains the pronunciation-prone words, extracting the audio segments corresponding to the pronunciation-prone words in the user's voice information;

When the audio segment matches the speech audio matching result in the wrong acoustic model, the user is prompted to pronounce the typo-prone word incorrectly, and the corresponding speech audio in the standard acoustic model is output according to the mapping table.

Further, before the described mapping table between the standard acoustic model and the erroneous acoustic model corresponding to the easily typo-pronounced word is established, it also includes:

Obtain the easy-to-pronounce words, and generate target words according to the easy-to-pronounce words;

Obtain the speech audio of the target word, and generate the standard acoustic model according to the speech audio of the target word;

Obtain the pronunciation confusing word of described pronunciation typo, replace the pronunciation confusing word in described target word with described pronunciation confusing word to generate confusing word;

Acquire the speech audio of the confused word, and generate the error acoustic model according to the speech audio of the confused word.

Further, it also includes:

When the audio segment matches the speech audio matching result in the standard acoustic model, the user is prompted to pronounce the typo-prone word correctly.

Further, it also includes:

When the audio segment does not match the speech audio matching result in the acoustic model, converting the audio segment into a recognized text, and the acoustic model includes the standard acoustic model and the wrong acoustic model;

If the target word contains the recognized text, determine whether the pronunciation of the audio segment is correct, if so, update the standard acoustic model according to the audio segment; otherwise, update the incorrect acoustic model according to the audio segment, and prompt The user's described pronunciation is prone to typos and mispronunciation, and according to the mapping table, the corresponding speech audio in the standard acoustic model is output;

If the target word does not contain the recognized text, the target word is updated according to the recognized text, and the acoustic model is updated according to the audio segment.

Further, when the target word does not contain the recognition text, updating the target word according to the recognition text, and updating the acoustic model according to the audio clip specifically includes:

When the target word does not contain the recognized text, updating the target word according to the recognized text;

If the audio segment is pronounced correctly, update the standard acoustic model according to the audio segment, update the confused word according to the updated target word, and then update the error acoustic model according to the updated speech audio of the confused word Model;

If the audio segment is pronounced incorrectly, obtain the correct voice audio of the recognized text, update the standard acoustic model according to the correct voice audio, update the confused word according to the updated target word, and then update the confusing word according to the updated target word. The speech audio of the confused words updates the error acoustic model.

The present invention also provides a correction system for speech recognition pronunciation errors, characterized in that it includes:

A mapping table building module is used to establish a mapping table between the standard acoustic model and the wrong acoustic model corresponding to the easily typo-pronounced words;

Obtaining module to obtain user voice information;

An extraction module, identifying the user voice information obtained by the obtaining module, and extracting the audio clips corresponding to the words containing the easily typo-pronounced words in the user voice information when the voice information contains the words that are prone to pronunciation errors;

The processing module, when the audio segment extracted by the extraction module is consistent with the voice audio matching result in the wrong acoustic model in the mapping table establishment module, the user is prompted to pronounce the typo-prone words incorrectly, and according to The mapping table established by the mapping table establishment module outputs the corresponding speech audio in the standard acoustic model.

Further, it also includes:

A typo-prone acquisition module to acquire the easily typo-pronounced characters;

A target word generation module, generating target words according to the pronunciation-prone words obtained by the typo-prone acquisition module;

an audio acquisition module, which acquires the voice audio of the target word generated by the target word generation module;

an acoustic model generation module, which generates the standard acoustic model according to the speech audio of the target word acquired by the audio acquisition module;

A confusing word acquisition module, for acquiring the pronunciation confusing words of the easily typo-pronounced words acquired by the typo-prone word acquisition module;

A confusing word generation module, which replaces the pronunciation-prone words in the target words generated by the target word generation module with the pronunciation confusing words obtained by the confusing word acquisition module to generate confusing words;

The audio acquisition module acquires the speech audio of the confused words generated by the confused word generation module;

The acoustic model generation module generates the wrong acoustic model according to the speech audio of the confused word acquired by the audio acquisition module.

Further, it also includes:

The processing module, when the audio segment extracted by the extraction module matches the voice and audio matching results in the standard acoustic model in the mapping table establishment module, prompting the user to pronounce the typo-prone word correctly.

Further, it also includes:

The processing module, when the audio segment extracted by the extraction module does not match the voice audio matching result in the acoustic model in the mapping table establishment module, converts the audio segment into recognized text, and the acoustic models include the standard acoustic model and the error acoustic model;

The control module, if the target word includes the recognition text converted by the processing module, judges whether the pronunciation of the audio clip is correct, if so, updates the standard acoustic model according to the audio clip; otherwise, according to the audio clip Updating the wrong acoustic model, and prompting the user that the pronunciation is prone to typos and mispronunciation, and output the corresponding voice audio in the standard acoustic model according to the mapping table;

The control module, if the target word does not contain the recognition text converted by the processing module, updates the target word according to the recognition text, and updates the acoustic model according to the audio segment.

Further, the control module specifically includes:

A target word updating unit, when the target word does not include the recognition text converted by the processing module, update the target word according to the recognition text;

The control unit, if the pronunciation of the audio segment extracted by the extraction module is correct, then update the standard acoustic model according to the audio segment, update the confusing word according to the target word updated by the target word update unit, and then update the confusing word according to the target word update unit. The updated speech audio of the confused word updates the wrong acoustic model;

The control unit, if the audio segment extracted by the extraction module is mispronounced, obtains the correct voice audio of the recognized text, updates the standard acoustic model according to the correct voice audio, and updates the unit according to the target word The updated target word updates the confused word, and then the error acoustic model is updated according to the updated speech audio of the confused word.

A kind of correction method and system of speech recognition pronunciation error provided by the present invention can bring at least one of the following beneficial effects:

1. In the present invention, a mapping table between the standard acoustic model corresponding to the typo-prone words and the erroneous acoustic model is established to facilitate subsequent matching and recognition of user voice information, and erroneous pronunciation is corrected.

2. In the present invention, when matching user voice information, only audio clips corresponding to words that contain mispronounced words are extracted for matching, which reduces the difficulty and workload of the matching process, speeds up the matching, and improves the matching accuracy.

Description of drawings

The preferred embodiments will be described below in a clear and easy-to-understand manner with reference to the accompanying drawings, and further description will be given of the above-mentioned characteristics, technical features, advantages and implementations of a method and system for correcting pronunciation errors in speech recognition.

Fig. 1 is the flow chart of the first embodiment of the correction method of a kind of speech recognition pronunciation error of the present invention;

Fig. 2 is the flow chart of the second embodiment of the correction method of a kind of speech recognition pronunciation error of the present invention;

Fig. 3 is the flow chart of the 3rd embodiment of the correction method of a kind of speech recognition pronunciation error of the present invention;

Fig. 4 is the structural representation of the 4th embodiment of the correction system of a kind of speech recognition pronunciation error of the present invention;

Fig. 5 is the structural representation of the fifth embodiment of a kind of speech recognition pronunciation error correction system of the present invention;

FIG. 6 is a schematic structural diagram of a sixth embodiment of a system for correcting pronunciation errors in speech recognition according to the present invention.

Description of reference numbers:

1000 Speech Recognition Pronunciation Error Correction System

1100 mapping table establishment module 1200 acquisition module 1300 extraction module

1400 processing module 1500 typo-prone acquisition module 1600 target word generation module

1700 Audio Acquisition Module 1800 Acoustic Model Generation Module 1850 Confused Word Acquisition Module

1900 Confused Word Generation Module

1950 Control Module 1951 Target Word Update Unit 1952 Control Unit

Detailed ways

In order to more clearly describe the embodiments of the present invention or the technical solutions in the prior art, the specific embodiments of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative efforts, and obtain other implementations.

In order to keep the drawings concise, the drawings only schematically show the parts related to the present invention, and they do not represent its actual structure as a product. In addition, in order to make the drawings concise and easy to understand, in some drawings, only one of the components having the same structure or function is schematically shown, or only one of them is marked. As used herein, "one" not only means "only one", but also "more than one".

The first embodiment of the present invention, as shown in Figure 1, is a method for correcting pronunciation errors in speech recognition, including:

S100 establishes a mapping table between the standard acoustic model corresponding to the typo-prone words and the wrong acoustic model.

Specifically, the polyphonic characters and Chinese characters with multiple approximations are used as pronunciation-prone words, the correct pronunciation is used as the standard acoustic model, the polyphonic characters and the approximations are used as the error acoustic model, and then the standard acoustic model and the error acoustic model are established. The mapping table corresponding to each other is convenient to find the corresponding correct pronunciation for correcting the wrong pronunciation later.

S200 acquires user voice information.

Specifically, the user's voice information is obtained, which may be input by the user in real time. For example, when a user uses an intelligent terminal to read aloud, voice correction needs to be performed, or a primary school student needs to check the learning achievement of the primary school student when learning new vocabulary through the intelligent terminal. It may also be to realize the recorded audio, such as checking whether the pronunciation in the audio recorded by the students is accurate, etc. When the system detects a wrong pronunciation, it will prompt for correction, so the user can choose whether the system or a single application will enable the function of prompting and correcting the wrong pronunciation.

S300 identifies the user's voice information, and when the voice information contains the easily typo-pronounced words, extracting audio segments corresponding to the words containing the easily-pronounced typo in the user voice information.

Specifically, the obtained user voice information is first converted into text information, and then it is identified whether the text information contains any one or more of the above-mentioned words that are prone to mispronunciation. Part-of-speech, marking words containing easily typo-pronounced words, and finally extracting one or more audio clips corresponding to the marked words from the user's speech information according to the marking result.

S400, when the audio segment matches the voice and audio matching results in the wrong acoustic model, prompt the user that the pronunciation of the typo-prone word is wrong, and output the corresponding voice and audio in the standard acoustic model according to the mapping table .

Specifically, according to the typo-prone words corresponding to the extracted audio clips, the corresponding speech audios of the typo-prone words in the standard acoustic model and the error acoustic model are found, and then the extracted audio clips and the found speech audios are matched one by one.

If the voice audio that matches the matching result belongs to the voice audio in the wrong acoustic model, it means that the part of the typo-prone word in the user's voice information is mispronounced, so the user is prompted to pronounce the typo-prone word incorrectly.

Since the same typo-prone word in the user's voice information may appear multiple times, but not every place is mispronounced, it is necessary to clearly explain the typo-prone word in combination with the context of the typo-prone word in the user's voice information when prompting a mispronunciation specific location.

In addition, through the mapping table between the standard acoustic model and the wrong acoustic model, the correct speech audio corresponding to the word formed by the typo-prone word at the position is found, and output to the user, so that the user can correct the pronunciation.

In this embodiment, after acquiring the user's voice information, firstly identify the typo-prone words contained in the user's voice information, and then extract the audio segment, standard acoustic model and error acoustic model corresponding to the words containing the typo-prone words in the user's voice information to match the speech audio in . Only valid audio clips are extracted for matching, which reduces the length of speech that needs to be matched, reduces the requirements for the system matching capability, speeds up the matching speed, and increases the matching accuracy.

The second embodiment of the present invention is an optimized embodiment of the above-mentioned first embodiment, as shown in FIG. 2 , including:

S010 obtains the pronunciation-prone words, and generates a target word according to the pronunciation-prone words.

Specifically, the polyphonic words and Chinese characters with multiple shapes are regarded as easy-to-pronounce characters, and the target words are generated according to the easy-to-pronounce characters. Near words, and the pronunciation of the typo in the target word is easy to confuse, such as Xin Xin Xuezi, download, finale, etc. Or the pronunciation of some words does not affect the daily communication of users, but the pronunciation is actually wrong, such as stubbornness, frame, etc., should also be corrected, especially when primary school students just start learning new words, they should start from the most Form correct pronunciation from the beginning to avoid later correction.

S020 acquires the speech audio of the target word, and generates the standard acoustic model according to the speech audio of the target word.

Specifically, the voice audio of each target word is obtained, wherein each typo may correspond to one or more voice audios, and each target word is obtained from the audio of standard Mandarin pronunciation, and if the user's pronunciation needs to be corrected later , the system outputs the audio of the standard Mandarin pronunciation first.

However, due to the influence of the user's gender, age, region, dialect accent and accent, etc., there may be cases where the user's pronunciation is correct but cannot match the standard Mandarin pronunciation audio. Therefore, in addition to obtaining the standard Mandarin pronunciation audio for each target word, you should also try to obtain the audio. The audio of the correct pronunciation of people of different ages and different regions, for the correct pronunciation but not standard, the user can choose whether to prompt and correct it later.

The standard acoustic model is generated according to the speech and audio of the target word. The standard acoustic model is classified according to the words that are prone to typos. When there are many target words generated by a word that is prone to pronunciation errors, or there are many speech and audio corresponding to a certain target word, you can choose to Under the classification of easy-to-pronounce words, it is classified again according to the target words.

S030 obtains the pronunciation confusing words of the pronunciation-prone words, and replaces the pronunciation-prone words in the target words with the pronunciation confusing words to generate confusing words.

Specifically, obtain a pronunciation confusing word that is prone to mispronunciation, and the pronunciation confusing word is a mispronunciation-prone word with other pronunciations or a similar word to the pronunciation-prone word, and replace the easily mispronounced word in the target word with a pronunciation confusing word to generate a confusing word, and also It is to combine the typo-prone words of other pronunciations with the target words, or combine the typos-prone words and the target words to generate confusing words.

Confused words are words corresponding to the voices that the user may mispronounce when reading the target words. From the composition point of view, the confused words do not conform to the rules of word composition. audio.

S040 acquires the speech audio of the confused word, and generates the error acoustic model according to the speech audio of the confused word.

Specifically, the voice audio of each confused word is obtained. Similarly, due to the influence of the user's gender, age, region, dialect accent, and tone, etc., there may be multiple voice audios corresponding to the same confused word. Generate an error acoustic model based on the speech audio of the confused word, and the error acoustic model is classified according to the corresponding pronunciation prone words. Select Classify again by confusing words under Mispronounceable words.

S100 establishes a mapping table between the standard acoustic model corresponding to the typo-prone words and the wrong acoustic model.

S200 acquires user voice information.

S300 identifies the user's voice information, and when the voice information contains the easily typo-pronounced words, extracting audio segments corresponding to the words containing the easily-pronounced typo in the user voice information.

S400, when the audio segment matches the voice and audio matching results in the wrong acoustic model, prompt the user that the pronunciation of the typo-prone word is wrong, and output the corresponding voice and audio in the standard acoustic model according to the mapping table .

S500, when the audio segment matches the speech audio matching result in the standard acoustic model, prompt the user to pronounce the typo-prone word correctly.

Specifically, according to the typo-prone words corresponding to the extracted audio clips, the corresponding speech audios of the typo-prone words in the standard acoustic model and the error acoustic model are found, and then the extracted audio clips and the found speech audios are matched one by one.

If the matching voice audio belongs to the standard acoustic model, it means that the part of the typo-prone words in the user's voice information is pronounced correctly, but it is necessary to further judge whether the matching voice audio is standard Mandarin voice audio, and if so It indicates that the user's pronunciation is not only correct but also standard. If not, the user is prompted that the pronunciation of the typo-prone word is correct but not standard, and the user is asked whether to output standard Mandarin pronunciation audio.

In this embodiment, a standard acoustic model for the correct pronunciation of the typo-prone characters and a corresponding erroneous acoustic model for the mispronunciation of a combination of a polyphonic word that is prone to mispronunciation and a form-approximation word are respectively generated, and two acoustic models are established. The mapping table of the corresponding relationship between the models is used for subsequent identification of the user's voice information and then corresponding matching, which can accurately identify and correct the mispronunciation in the user's voice information.

The third embodiment of the present invention is an optimized embodiment of the above-mentioned first embodiment, as shown in FIG. 3 , including:

S100 establishes a mapping table between the standard acoustic model corresponding to the typo-prone words and the wrong acoustic model.

S200 acquires user voice information.

S300 identifies the user's voice information, and when the voice information contains the easily typo-pronounced words, extracting audio segments corresponding to the words containing the easily-pronounced typo in the user voice information.

S400, when the audio segment matches the voice and audio matching results in the wrong acoustic model, prompt the user that the pronunciation of the typo-prone word is wrong, and output the corresponding voice and audio in the standard acoustic model according to the mapping table .

S600, when the audio segment does not match the speech audio matching result in the acoustic model, convert the audio segment into recognized text, and the acoustic model includes the standard acoustic model and the wrong acoustic model.

Specifically, the acoustic model is found according to the typo-prone words corresponding to the extracted audio clips, that is, the standard acoustic model and the wrong acoustic model. match.

If the matching results of the audio segment and the speech audio in the acoustic model do not match, it means that the pronunciation corresponding to the audio segment is not included in the acoustic model. Therefore, the target word, confusion word and acoustic model should be updated according to the audio segment. . Recognize the audio segment converts the audio segment to recognized text.

S610, if the target word contains the recognized text, determine whether the pronunciation of the audio segment is correct, and if so, update the standard acoustic model according to the audio segment; otherwise, update the incorrect acoustic model according to the audio segment, and The user is prompted that the pronunciation of the typo-prone word is incorrect, and the corresponding speech audio in the standard acoustic model is output according to the mapping table.

Specifically, the recognition text is recognized, and the recognized text is matched with the target word. If the matching result shows that the recognized text matches a target word, it is determined whether the pronunciation of the audio clip is correct. If it is correct, the audio clip may be due to Accents are not included in the standard acoustic model for reasons such as rare accents. Therefore, the standard acoustic model is updated according to the audio clip, and then the user is prompted that the pronunciation is prone to typos, but the pronunciation is correct but not standard, whether it is necessary to output standard Mandarin pronunciation audio.

If the pronunciation of the audio clip is incorrect, it means that the audio clip is a wrong pronunciation of a target word that is not included in the wrong acoustic model. Therefore, the wrong acoustic model is updated according to the audio clip, and then the user is prompted that the pronunciation is prone to typos and the wrong pronunciation is output. The corresponding standard Mandarin pronunciation audio is corrected for the user.

S620 If the target word does not contain the recognized text, update the target word according to the recognized text, and update the acoustic model according to the audio segment.

Specifically, the recognized text is recognized, and the recognized text is matched with the target word. If the matching result shows that the recognized text does not match all the target words, it means that the recognized text has not been included as the target word. Therefore, according to the recognized text Update the target words and then update the acoustic model based on the audio clip.

In step S620, if the target word does not contain the recognition text, updating the target word according to the recognition text, and updating the acoustic model according to the audio segment specifically includes:

S621 When the target word does not contain the recognized text, update the target word according to the recognized text.

S622 If the pronunciation of the audio clip is correct, update the standard acoustic model according to the audio clip, update the confused word according to the updated target word, and then update the error according to the voice audio of the updated confused word acoustic model.

Specifically, after the target word is updated according to the recognized text, the pronunciation of the audio segment is judged, and if the pronunciation is correct, the standard acoustic model is updated according to the audio segment, and other pronunciation audios of the recognized text can be obtained to update the standard acoustic model subsequently.

Then, the pronunciation-prone words in the updated target words are replaced with pronunciation-confused words, and the confused words are correspondingly updated, the speech audio of the updated confused words is re-acquired, and the error acoustic model is updated according to the re-acquired speech audio.

S623 If the audio segment is pronounced incorrectly, obtain the correct voice audio of the recognized text, update the standard acoustic model according to the correct voice audio, update the confused word according to the updated target word, and then update the After the speech audio of the confused word updates the error acoustic model.

Specifically, after the target word is updated according to the recognized text, the pronunciation of the audio segment is judged, and if the pronunciation is incorrect, the correct voice audio corresponding to the recognized text is obtained, including the standard Mandarin pronunciation audio and the pronunciation of other dialects, accents, etc. Audio, update the standard acoustic model according to the acquired correct speech audio.

Then, the pronunciation-prone words in the updated target words are replaced with pronunciation-confused words, and the confused words are correspondingly updated, the speech audio of the updated confused words is re-acquired, and the error acoustic model is updated according to the re-acquired speech audio.

In this embodiment, when the extracted audio segment does not match the voice-audio matching result in the acoustic model, the audio segment is converted into recognized text, and then the recognized text and audio segment are recognized, and different situations are classified and processed, which is fast and accurate. The user's pronunciation is judged and processed accordingly.

The fourth embodiment of the present invention, as shown in FIG. 4 , is a system 1000 for correcting pronunciation errors in speech recognition, including:

The mapping table establishment module 1100 establishes a mapping table between the standard acoustic model and the erroneous acoustic model corresponding to the easily typo-pronounced words.

Specifically, the mapping table building module 1100 uses polyphonic words and Chinese characters with multiple approximations as pronunciation-prone words, takes correct pronunciation as a standard acoustic model, and uses polyphonic characters and approximations as an error acoustic model, and then establishes a standard acoustic model The corresponding mapping table between the model and the wrong acoustic model, so as to find the corresponding correct pronunciation for correcting the wrong pronunciation later.

The obtaining module 1200 obtains user voice information.

Specifically, the acquisition module 1200 acquires user voice information, which may be input by the user in real time. For example, when a user uses an intelligent terminal to read aloud, he needs to perform voice correction, or when a primary school student learns new vocabulary through a smart terminal, it is necessary to perform a correction on the primary school students' learning results. test. It may also be to realize the recorded audio, such as checking whether the pronunciation in the audio recorded by the students is accurate, etc. When the system detects a wrong pronunciation, it will prompt for correction, so the user can choose whether the system or a single application will enable the function of prompting and correcting the wrong pronunciation.

The extracting module 1300 identifies the user's voice information obtained by the obtaining module 1200, and when the voice information contains the typo-prone words, extracts the audio corresponding to the words that contain the typo-prone words in the user's voice information Fragment.

Specifically, the extraction module 1300 first converts the acquired user voice information into text information, and then identifies whether the text information contains any one or more of the above-mentioned words that are prone to mispronunciation, and if so, analyzes the components of sentences in the text information through word segmentation technology and part-of-speech of words, mark words that contain easily mispronounced words, and finally extract one or more audio clips corresponding to the marked words from the user's voice information according to the marking result.

The processing module 1400, when the audio segment extracted by the extracting module 1300 is consistent with the voice and audio matching results in the wrong acoustic model in the mapping table building module 1100, prompting the user that the pronunciation is prone to typos and the pronunciation is wrong. , and output the corresponding speech audio in the standard acoustic model according to the mapping table established by the mapping table establishment module 1100 .

Specifically, the processing module 1400 finds the corresponding voice audio of the typo-prone word in the standard acoustic model and the error acoustic model according to the typo-prone words corresponding to the extracted audio clips, and then performs the extracted audio clips and the found voice audios one by one. match. If the voice audio that matches the matching result belongs to the voice audio in the wrong acoustic model, it means that the part of the typo-prone word in the user's voice information is mispronounced, so the user is prompted to pronounce the typo-prone word incorrectly.

Since the same typo-prone word in the user's voice information may appear multiple times, but not every place is pronounced incorrectly, when the processing module 1400 prompts a mispronunciation, it is necessary to clearly explain the typo-prone word in the user's voice information in conjunction with the context of the pronunciation-prone word in the user's voice information. Pronounce the exact location of the typo.

In addition, through the mapping table between the standard acoustic model and the wrong acoustic model, the correct speech audio corresponding to the word formed by the typo-prone word at the position is found, and output to the user, so that the user can correct the pronunciation.

In this embodiment, after acquiring the user's voice information, firstly identify the typo-prone words contained in the user's voice information, and then extract the audio segment, standard acoustic model and error acoustic model corresponding to the words containing the typo-prone words in the user's voice information to match the speech audio in . Only valid audio clips are extracted for matching, which reduces the length of speech that needs to be matched, reduces the requirements for the system matching capability, speeds up the matching speed, and increases the matching accuracy.

The fifth embodiment of the present invention is an optimized embodiment of the above-mentioned fourth embodiment, as shown in FIG. 5 , including:

The typo-prone character acquisition module 1500 acquires the pronunciation-prone characters.

The target word generating module 1600 generates a target word according to the pronunciation-prone words obtained by the typo-prone obtaining module 1500 .

Specifically, the typo-prone acquisition module 1500 regards polyphonic words and Chinese characters with a plurality of approximations as pronunciation-prone characters, and the target word generation module 1600 generates target words according to the pronunciation-prone characters, and the target words are words that include pronunciation-prone characters. The typo-prone words are polyphonic words or have multiple approximations, and the pronunciation of the typo-prone words in the target words is easy to confuse, such as Xinxinzizi, download, finale, etc. Or the pronunciation of some words does not affect the daily communication of users, but the pronunciation is actually wrong, such as stubbornness, frame, etc., should also be corrected, especially when primary school students just start learning new words, they should start from the most Form the correct pronunciation at the beginning to avoid later correction.

The audio acquisition module 1700 acquires the speech audio of the target word generated by the target word generation module 1600 .

Specifically, the audio acquisition module 1700 acquires the audio of each target word, wherein each typo may correspond to one or more audios, and each target word acquires the audio of the standard Mandarin pronunciation. When the pronunciation is corrected, the audio output by the system is preferably the standard Mandarin pronunciation.

However, due to the influence of the user's gender, age, region, dialect accent and accent, etc., there may be cases where the user's pronunciation is correct but cannot match the standard Mandarin pronunciation audio. Therefore, in addition to obtaining the standard Mandarin pronunciation audio for each target word, you should also try to obtain the audio. The audio of the correct pronunciation of people of different ages and different regions, for the correct pronunciation but not standard, the user can choose whether to prompt and correct it later.

The acoustic model generation module 1800 generates the standard acoustic model according to the speech audio of the target word acquired by the audio acquisition module 1700 .

Specifically, the acoustic model generation module 1800 generates a standard acoustic model according to the speech audio of the target word, and the standard acoustic model is classified according to the words that are prone to mispronunciation. There are many voices and audios, and you can choose to classify them again according to the target words under the classification of easy-to-pronounce words.

The confusing word obtaining module 1850 obtains the pronunciation confusing word of the easily typo-pronounced character obtained by the typo-prone character obtaining module 1500 .

The confusing word generating module 1900 replaces the pronunciation-prone words in the target words generated by the target word generating module 1600 with the pronunciation confusing words obtained by the confusing word obtaining module 1850 to generate a confusing word.

Specifically, the confusing word obtaining module 1850 obtains a pronunciation confusing word that is prone to mispronunciation, and the pronunciation confusing word is a mispronunciation-prone word with other pronunciations or a shape-approximation word of the mispronounced word. The confusing words are generated by replacing the pronunciation confusing words, that is, combining the mispronounced words of other pronunciations with the target words, or combining the typos-prone words and the target words to generate the confusing words.

Confused words are words corresponding to the voices that the user may mispronounce when reading the target words. From the composition point of view, the confused words do not conform to the rules of word composition. audio.

The audio acquisition module 1700 acquires the speech audio of the confused words generated by the confused word generation module 1900 .

The acoustic model generation module 1800 generates the wrong acoustic model according to the speech audio of the confused word acquired by the audio acquisition module 1700 .

Specifically, the audio acquisition module 1700 acquires the voice audio of each confused word. Similarly, due to the influence of the user's gender, age, region, dialect accent, and tone, there may be multiple voice audios corresponding to the same confused word. The acoustic model generation module 1800 generates an incorrect acoustic model according to the speech audio of the confusing word, and the incorrect acoustic model is classified according to the corresponding pronunciation prone words. There are many audios, and you can choose to classify them again according to confusing words under the classification of easy-to-pronounce words.

The mapping table establishment module 1100 establishes a mapping table between the standard acoustic model and the wrong acoustic model corresponding to the pronunciation-prone words according to the standard acoustic model and the wrong acoustic model generated by the acoustic model generation module 1800 .

The obtaining module 1200 obtains user voice information.

The extracting module 1300 identifies the user's voice information obtained by the obtaining module 1200, and when the voice information contains the typo-prone words, extracts the audio corresponding to the words that contain the typo-prone words in the user's voice information Fragment.

The processing module 1400, when the audio segment extracted by the extracting module 1300 is consistent with the voice and audio matching results in the wrong acoustic model in the mapping table building module 1100, prompting the user that the pronunciation is prone to typos and the pronunciation is wrong. , and output the corresponding speech audio in the standard acoustic model according to the mapping table established by the mapping table establishment module 1100 .

The processing module 1400, when the audio segment extracted by the extraction module 1300 matches the voice and audio matching results in the standard acoustic model in the mapping table establishment module 1100, prompts the user that the pronunciation is prone to typos. Pronounced correctly.

Specifically, the processing module 1400 finds the corresponding voice audio of the typo-prone word in the standard acoustic model and the error acoustic model according to the typo-prone words corresponding to the extracted audio clips, and then performs the extracted audio clips and the found voice audios one by one. match.

If the matching speech audio of the processing module 1400 is the speech audio that belongs to the standard acoustic model, it means that the pronunciation of the part of the typo-prone words in the user's speech information is correct, but it needs to be further judged whether the matching speech audio is the standard Mandarin pronunciation. Audio, if it indicates that the user's pronunciation is not only correct but also standard, if not, it prompts the user that the pronunciation of the typo is correct but not standard, and asks the user whether to output standard Mandarin pronunciation audio.

In this embodiment, a standard acoustic model for the correct pronunciation of the typo-prone characters and a corresponding erroneous acoustic model for the mispronunciation of a combination of a polyphonic word that is prone to mispronunciation and a form-approximation word are respectively generated, and two acoustic models are established. The mapping table of the corresponding relationship between the models is used for subsequent identification of the user's voice information and then corresponding matching, which can accurately identify and correct the mispronunciation in the user's voice information.

The sixth embodiment of the present invention is an optimized embodiment of the above-mentioned fourth embodiment, as shown in FIG. 6 , including:

The mapping table establishment module 1100 establishes a mapping table between the standard acoustic model and the erroneous acoustic model corresponding to the easily typo-pronounced words.

The obtaining module 1200 obtains user voice information.

The extracting module 1300 identifies the user's voice information obtained by the obtaining module 1200, and when the voice information contains the typo-prone words, extracts the audio corresponding to the words that contain the typo-prone words in the user's voice information Fragment.

The processing module 1400, when the audio segment extracted by the extracting module 1300 is consistent with the voice and audio matching results in the wrong acoustic model in the mapping table building module 1100, prompting the user that the pronunciation is prone to typos and the pronunciation is wrong. , and output the corresponding speech audio in the standard acoustic model according to the mapping table established by the mapping table establishment module 1100 .

The processing module 1400, when the audio segment extracted by the extraction module 1300 does not match the voice audio matching result in the acoustic model in the mapping table establishment module 1100, converts the audio segment into a recognized text, The acoustic models include the standard acoustic model and the false acoustic model.

Specifically, the processing module 1400 finds the acoustic model according to the typo-prone words corresponding to the extracted audio clips, that is, the voice audio of the typo-prone words corresponding to the standard acoustic model and the error acoustic model, and then compares the extracted audio clips and the found speech sounds. The audio is matched one by one.

If the matching results of the audio segment and the speech audio in the acoustic model do not match, it means that the pronunciation corresponding to the audio segment is not included in the acoustic model. Therefore, the target word, confusion word and acoustic model should be updated according to the audio segment. . Recognize the audio segment converts the audio segment to recognized text.

The control module 1950, if the target word includes the recognition text converted by the processing module 1400, judges whether the pronunciation of the audio clip is correct, and if so, updates the standard acoustic model according to the audio clip; otherwise, according to the The audio segment updates the erroneous acoustic model, and prompts the user that the typo-prone word is mispronounced, and outputs the corresponding speech audio in the standard acoustic model according to the mapping table.

Specifically, the recognition text is recognized, and the recognized text is matched with the target word. If the matching result shows that the recognized text matches a target word, it is determined whether the pronunciation of the audio clip is correct. If it is correct, the audio clip may be due to Accents are not included in the standard acoustic model for reasons such as rare accents. Therefore, the standard acoustic model is updated according to the audio clip, and then the user is prompted that the pronunciation is prone to typos, but the pronunciation is correct but not standard, whether it is necessary to output standard Mandarin pronunciation audio.

If the pronunciation of the audio clip is incorrect, it means that the audio clip is a wrong pronunciation of a target word that is not included in the wrong acoustic model. Therefore, the wrong acoustic model is updated according to the audio clip, and then the user is prompted that the pronunciation is prone to typos and the wrong pronunciation is output. The corresponding standard Mandarin pronunciation audio is corrected for the user.

The control module 1950, if the target word does not contain the recognized text transformed by the processing module 1400, updates the target word according to the recognized text, and updates the acoustic model according to the audio clip.

Specifically, the recognized text is recognized, and the recognized text is matched with the target word. If the matching result shows that the recognized text does not match all the target words, it means that the recognized text has not been included as the target word. Therefore, according to the recognized text Update the target words and then update the acoustic model based on the audio clip.

The control module 1950 specifically includes:

The target word updating unit 1951, when the target word does not contain the recognized text transformed by the processing module 1400, updates the target word according to the recognized text.

The control unit 1952, if the pronunciation of the audio segment extracted by the extraction module 1300 is correct, updates the standard acoustic model according to the audio segment, and updates the confusing word according to the target word updated by the target word update unit 1951 , and then update the false acoustic model according to the updated speech audio of the confused word.

Specifically, after updating the target word according to the recognized text, the control unit 1952 judges the pronunciation of the audio segment, and if the pronunciation is correct, then updates the standard acoustic model according to the audio segment, and subsequently obtains other pronunciations of the recognized text to update the standard acoustic audio. Model.

Then, the pronunciation-prone words in the updated target words are replaced with pronunciation-confused words, and the confused words are correspondingly updated, the speech audio of the updated confused words is re-acquired, and the error acoustic model is updated according to the re-acquired speech audio.

The control unit 1952, if the audio segment extracted by the extraction module 1300 is mispronounced, obtains the correct voice audio of the recognized text, updates the standard acoustic model according to the correct voice audio, and updates the standard acoustic model according to the target word The updating unit 1951 updates the confused word with the updated target word, and then updates the erroneous acoustic model according to the speech audio of the updated confused word.

Specifically, after updating the target word according to the recognized text, the control unit 1952 judges the pronunciation of the audio segment, and if the pronunciation is incorrect, then obtains the correct speech audio corresponding to the recognized text, including standard Mandarin pronunciation audio and other dialects, accents and other pronunciation audio, and update the standard acoustic model according to the acquired correct speech audio.

Then, the pronunciation-prone words in the updated target words are replaced with pronunciation-confused words, and the confused words are correspondingly updated, the speech audio of the updated confused words is re-acquired, and the error acoustic model is updated according to the re-acquired speech audio.

In this embodiment, when the extracted audio segment does not match the voice-audio matching result in the acoustic model, the audio segment is converted into recognized text, and then the recognized text and audio segment are recognized, and different situations are classified and processed, which is fast and accurate. The user's pronunciation is judged and processed accordingly.

It should be noted that the above embodiments can be freely combined as required. The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements and modifications can be made. It should be regarded as the protection scope of the present invention.

Claims (8)

1. A method for correcting pronunciation errors in speech recognition, comprising:

establishing a mapping table between a standard acoustic model and an error acoustic model corresponding to pronounce error-prone characters;

acquiring user voice information;

recognizing the user voice information, and when the pronunciation-prone wrong character is contained in the voice information, extracting an audio segment corresponding to a word containing the pronunciation-prone wrong character in the user voice information;

when the audio frequency fragment is matched with the voice audio frequency matching result in the error acoustic model, prompting a user that the pronunciation is easy to miss, and outputting the corresponding voice audio frequency in the standard acoustic model according to the mapping table;

wherein, before establishing the mapping table between the standard acoustic model and the error acoustic model corresponding to the pronunciation error-prone word, the method further comprises: acquiring the pronouncing error-prone characters, and generating target words according to the pronouncing error-prone characters;

when the audio segment does not accord with the voice audio matching result in the acoustic model, converting the audio segment into recognition text, wherein the acoustic model comprises the standard acoustic model and the error acoustic model;

if the target words contain the recognition texts, judging whether the pronunciation of the audio clips is correct, and if so, updating the standard acoustic model according to the audio clips; otherwise, updating the error acoustic model according to the audio fragment;

and if the target word does not contain the identification text, updating the target word according to the identification text, and updating the acoustic model according to the audio fragment.

2. The method for correcting pronunciation errors in speech recognition according to claim 1, wherein the step of establishing the mapping table between the standard acoustic model and the incorrect acoustic model corresponding to the pronunciation-prone word further comprises:

acquiring the voice audio of the target word, and generating the standard acoustic model according to the voice audio of the target word;

acquiring pronunciation confusion words of the pronunciation easily-wrong words, and replacing the pronunciation easily-wrong words in the target words with the pronunciation confusion words to generate confusion words;

and acquiring the voice audio of the confusing words, and generating the error acoustic model according to the voice audio of the confusing words.

3. The method for correcting pronunciation errors in speech recognition according to claim 1 or 2, further comprising:

and when the audio frequency segment is matched with the voice audio frequency matching result in the standard acoustic model, prompting the user that the pronunciation of the error-prone character is correct.

4. The method according to claim 2, wherein the updating the target word according to the recognized text and the acoustic model according to the audio piece when the target word does not include the recognized text specifically includes:

when the target word does not contain the recognition text, updating the target word according to the recognition text;

if the pronunciation of the audio clip is correct, updating the standard acoustic model according to the audio clip, updating the confusion terms according to the updated target terms, and then updating the error acoustic model according to the voice audio of the updated confusion terms;

and if the audio fragment is mispronounced, acquiring correct voice audio of the recognized text, updating the standard acoustic model according to the correct voice audio, updating the confusion terms according to the updated target terms, and updating the wrong acoustic model according to the updated voice audio of the confusion terms.

5. A system for correcting pronunciation errors in speech recognition, comprising:

the mapping table establishing module is used for establishing a mapping table between a standard acoustic model and an error acoustic model corresponding to the pronounce error-prone character;

the acquisition module acquires user voice information;

the extraction module is used for identifying the user voice information acquired by the acquisition module, and extracting an audio clip corresponding to a word containing the pronouncing error-prone character in the user voice information when the pronunciation error-prone character is contained in the voice information;

the processing module prompts a user that the pronunciation is easy to miss and the pronunciation is wrong when the audio frequency segment extracted by the extraction module is matched with the voice audio frequency matching result in the wrong acoustic model in the mapping table establishing module, and outputs the corresponding voice audio frequency in the standard acoustic model according to the mapping table established by the mapping table establishing module;

the error-prone character acquisition module is used for acquiring the pronunciation error-prone characters before the mapping table establishment module establishes a mapping table between a standard acoustic model and an error acoustic model corresponding to the pronunciation error-prone characters;

the target word generation module is used for generating a target word according to the pronunciation error-prone characters acquired by the error-prone character acquisition module;

the processing module is used for converting the audio clip into an identification text when the audio clip extracted by the extraction module is not consistent with a voice audio matching result in an acoustic model in the mapping table establishing module, wherein the acoustic model comprises the standard acoustic model and the error acoustic model;

the control module is used for judging whether the pronunciation of the audio clip is correct or not if the target word contains the recognition text converted by the processing module, and updating the standard acoustic model according to the audio clip if the pronunciation of the audio clip is correct; otherwise, updating the error acoustic model according to the audio fragment;

and if the target word does not contain the identification text converted by the processing module, the control module updates the target word according to the identification text and updates the acoustic model according to the audio segment.

6. The system for correcting pronunciation errors in speech recognition according to claim 5, further comprising:

the audio acquisition module is used for acquiring the voice audio of the target words generated by the target word generation module;

the acoustic model generating module is used for generating the standard acoustic model according to the voice audio of the target word acquired by the audio acquiring module;

the confusing character acquisition module is used for acquiring the pronunciation confusing characters of the pronunciation confusing characters acquired by the confusing character acquisition module;

a confusion word generation module, which replaces the pronunciation-prone wrong character in the target word generated by the target word generation module with the pronunciation-prone wrong character acquired by the confusion word acquisition module to generate a confusion word;

the audio acquisition module is used for acquiring the voice audio of the confusion word generated by the confusion word generation module;

the acoustic model generating module is used for generating the error acoustic model according to the voice audio of the confusing words acquired by the audio acquiring module.

7. The system for correcting pronunciation errors in speech recognition according to claim 5 or 6, further comprising:

and the processing module prompts a user that the pronunciation of the error-prone character is correct when the audio frequency segment extracted by the extraction module is matched with the voice audio frequency matching result in the standard acoustic model in the mapping table establishing module.

8. The system for correcting pronunciation errors in speech recognition according to claim 6, wherein the control module specifically comprises:

a target word updating unit, which updates the target word according to the recognition text when the target word does not contain the recognition text converted by the processing module;

the control unit is used for updating the standard acoustic model according to the audio segment if the pronunciation of the audio segment extracted by the extraction module is correct, updating the confusion word according to the target word updated by the target word updating unit, and then updating the wrong acoustic model according to the voice audio of the updated confusion word;

the control unit is used for acquiring correct voice audio of the recognized text if the audio segment extracted by the extraction module is in wrong pronunciation, updating the standard acoustic model according to the correct voice audio, updating the confusion word according to the target word updated by the target word updating unit, and then updating the wrong acoustic model according to the updated voice audio of the confusion word.

Images