Voice transcription processing system

Abstract

The invention belongs to the technical field of voice transcription, and particularly relates to a voice transcription processing system which comprises a multi-channel microphone array module, a preprocessing module, a feature extraction module, a voice recognition and transcription module and a data storage and management module. The multi-channel microphone array module is used for collecting a voice signal of a user and determining a first recorded voice; a microphone array is utilized to specify a user # imgabs0 #, when the user # imgabs1 # speaks, voice signals of other users # imgabs2 # can be automatically shielded, only voice signals of the user # imgabs3 # corresponding to the determined first recorded voice are recorded, and it can be ensured that when a deep learning model recognizes the voice signals and outputs a transliteration text # imgabs4 #, the transliteration text # imgabs5 # can correspond to the user # imgabs6 #.

Description

Technical Field

[0001] The present invention belongs to the technical field of speech transcription, in particular to a speech transcription processing system. Background Art

[0002] In business operations, meetings are very common. In order to ensure the subsequent review of the meeting content, meeting minutes are usually taken. In meeting minutes, manual recording is the most traditional recording method. Due to its timeliness and accuracy issues, it may miss some key points in the meeting, or involuntarily mix in some personal opinions, making the meeting minutes inaccurate, and has gradually been eliminated.

[0003] With the development of intelligent voice in the existing technology, the widely used speech recognition technology can obtain the voice signal of the speaker in the meeting, and use the deep learning model to recognize and transcribe the voice signal, and then obtain the meeting record. Compared with the traditional manual record, the meeting record obtained by analyzing the voice signal using the deep learning model in the existing technology is faster and does not contain personal opinions. The meeting record is more organized and clear; However, in the prior art, due to the large number of participants and the noisy environment, when a microphone is used to obtain the speaker's voice signal according to the existing speech recognition technology, other people and environmental noise may also be included, resulting in the obtained voice signal containing more noise, affecting the effect of speech recognition. At the same time, it will cause the speech recognition and transcribed text to be disordered and inconsistent with the speaker's voice signal, resulting in inaccurate generated meeting records. That is, when multiple speakers speak at the same time in a meeting, it will seriously affect the recognition and transcription effect, and it is impossible to obtain accurate meeting records.

[0004] To this end, the present invention provides a speech transcription processing system. Summary of the invention

[0005] In order to make up for the deficiencies of the prior art, at least one technical problem raised in the background technology is solved.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a speech transcription processing system described in the present invention comprises: Multi-channel microphone array module for collecting user A voice signal, and determining a first recorded voice; A preprocessing module, which preprocesses the collected speech signals; A feature extraction module extracts speech features from the preprocessed speech signal, wherein the speech features include the frequency, energy and voiceprint features of the speech signal; The speech recognition and transcription module recognizes speech signals and transcribes them based on a deep learning model to output transcribed text ; The data storage and management module is used to store and manage the voice signals input into the voice recognition and transcription module and the corresponding transcribed text .

[0007] Preferably, the method for determining the first recorded voice is: Arrange microphone arrays corresponding to multiple users , ; Based on the feature extraction module, the overlapping speech is extracted from the speech features; the overlapping speech is a plurality of users The output speech signal overlaps segments in time and space; Based on the extracted speech features, the strength relationship of multiple speech signals in the overlapping speech is analyzed, and the strength relationship is determined according to the energy size of the speech signal; Separate the speech signal with the largest energy from the overlapping speech and use it as the corresponding microphone array The first recorded speech, and when the energy of the speech signal is lower than the energy threshold, the microphone array No voice signals are recorded; The first recorded speech is enhanced based on spatial filtering technology, and interference signals are suppressed.

[0008] Preferably, the preprocessing of the collected speech signal includes denoising, noise reduction, filtering and frame processing.

[0009] Preferably, the voiceprint feature of the speech signal is used to match the user in the subsequent speaking process. , the method is: Get microphone array the determined first recorded voice; Using voiceprint recognition technology, identifying and obtaining a first voiceprint feature of the first recorded voice; Microphone array The user corresponding to the first voiceprint feature to make a match.

[0010] Preferably, the microphone array The user corresponding to the first voiceprint feature Matching also includes: Based on the matched microphone array With the first voiceprint feature, user ; When the microphone array When a voice signal is recorded, the feature extraction module extracts the Nth voiceprint feature of the voice signal and determines whether the Nth voiceprint feature matches the user. The first voiceprint feature matching: If the Nth voiceprint feature does not match the first voiceprint feature, the microphone array Automatically shield the voice signal; on the contrary, when the Nth voiceprint feature matches the first voiceprint feature, the microphone array The voice signal is recorded.

[0011] Preferably, the method for recognizing speech signals and transcribing them based on a deep learning model is: When the Nth voiceprint feature of the voice signal matches the first voiceprint feature, the voice signal is determined to be the first recorded voice; Microphone Array The speech signal is collected, and features are extracted from the speech signal to obtain speech features; Input the extracted speech features into the trained deep learning model; Analyze and recognize based on deep learning models and output transcribed text ; The deep learning model is based on convolutional neural networks and recurrent neural networks to recognize complex speech feature sequences and transcribe them into text. Output mode.

[0012] Preferably, it also includes a camera control module, the camera control module is used to adjust the camera to automatically turn to the user , and record the video; The adjustment camera automatically turns to the user The method is: Get microphone array Matching users The first voiceprint feature; When the Nth voiceprint feature of the voice signal matches the first voiceprint feature, the camera is controlled to automatically turn to the microphone array Users at location .

[0013] Preferably, the recorded video is used to analyze the user Detect changes in mouth shape, extract mouth shape features, and assist speech recognition and transcription based on mouth shape features; The method for assisting speech recognition and transcription based on lip shape features is: The acquired lip shape features are integrated with the speech features and input into the deep learning model; Optimizing transcription based on deep learning models The accuracy of Transcript generated by speech feature recognition , generate the corresponding lip animation; Determine how well the lip animation matches the lip features: When the lip animation is inconsistent with the lip features, the transcription text is adjusted based on semantic analysis technology , generate optimized text , the optimized text The lip animation must match the lip features; otherwise, no optimization will be performed.

[0014] Preferably, it also includes a user interface module, which provides a user interface, including a login page, a task page, a permission page, and a data record page; The login page is used for users Log in, the task page includes creating real-time tasks and offline tasks, where real-time tasks are real-time speech recognition and transcription tasks, while offline tasks are offline speech recognition and transcription tasks based on input speech signal files; the permission page is used to manage users Permissions, and only users with permissions After logging in through the login page, you can The data record page is used to display the voice records and corresponding transcribed text output by the data storage and management module .

[0015] Preferably, the storage and management of the voice signal input to the voice recognition and transcription module and the corresponding transcribed text The method is: Create a folder according to the project name corresponding to the real-time task or offline task; The speech signal and the corresponding transcribed text The statements in are stored in folders corresponding to the project names in chronological order; Generates logs based on folders and folder creation time, and displays the logs on the data logging page.

[0016] The beneficial effects of the present invention are as follows: The speech transcription processing system of the present invention utilizes multiple microphone arrays Specify User , in the user While speaking, other users can be automatically blocked The voice signal of the first recorded voice is recorded, and only the user corresponding to the determined first recorded voice is recorded. The voice signal is input into the deep learning model to avoid interference with the voice signal, which may cause the recognition and transcription of the text Confusion, according to the microphone array of different channels , and the corresponding records belong to different users The speech signal can ensure that the deep learning model recognizes the speech signal and outputs the transcribed text When the text is transcribed Corresponding users , which is conducive to ensuring the clarity of the meeting minutes and facilitating subsequent review of the meeting minutes. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be further described below in conjunction with the accompanying drawings.

[0018] Figure 1 is a flow chart of the present invention; Figure 2 It is a matching diagram between the microphone array and the user in the present invention. DETAILED DESCRIPTION

[0019] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0020] like Figure 1 As shown, a speech transcription processing system according to an embodiment of the present invention comprises a multi-channel microphone array module, a preprocessing module, a feature extraction module, a speech recognition and transcription module, and a data storage and management module; the multi-channel microphone array module is used to collect user The preprocessing module is used to preprocess the collected voice signal; the feature extraction module is used to extract voice features from the preprocessed voice signal, wherein the voice features include the frequency, energy and voiceprint features of the voice signal; the voice recognition and transcription module recognizes the voice signal and transcribes it based on the deep learning model to output the transcribed text ; The data storage and management module is used to store and manage the voice signals input into the voice recognition and transcription module and the corresponding transcribed text .

[0021] In the prior art, due to the large number of participants and the noisy environment, according to the existing speech recognition technology, when a microphone is used to obtain the speaker's voice signal, other people and environmental noise may also be included, resulting in the obtained voice signal containing more noise, affecting the effect of speech recognition. At the same time, it will cause the speech recognition and transcribed text to be disordered and inconsistent with the speaker's voice signal, resulting in inaccurate generated meeting records. That is, when multiple speakers speak at the same time in a meeting, it will seriously affect the recognition and transcription effect, and it is impossible to obtain accurate meeting records.

[0022] Based on the above, in one embodiment of the present invention, a multi-channel microphone array module is used to set up a plurality of user-based Microphone array , using multiple microphone arrays Capture from different users The speech signal is enhanced by spatial filtering technology. signals from other directions, thereby effectively reducing the impact of background noise and improving user The quality of the speech signal, where the microphone array The first recorded voice determined may be a plurality of microphone arrays Specify User , in the user While speaking, other users can be automatically blocked The voice signal of the first recorded voice is recorded, and only the user corresponding to the determined first recorded voice is recorded. The voice signal is input into the deep learning model to avoid interference with the voice signal, which may cause the recognition and transcription of the text Confusion, according to the microphone array of different channels , and the corresponding records belong to different users The speech signal can ensure that the deep learning model recognizes the speech signal and outputs the transcribed text When the text is transcribed Corresponding users , which is conducive to ensuring the clarity of the meeting minutes and facilitating subsequent review of the meeting minutes.

[0023] In one embodiment, a method for determining the first recorded voice is: Arrange microphone arrays corresponding to multiple users , ; Based on the feature extraction module, the overlapping speech is extracted from the speech features; the overlapping speech is a plurality of users The output speech signal overlaps segments in time and space; Based on the extracted speech features, the strength relationship of multiple speech signals in the overlapping speech is analyzed, and the strength relationship is determined according to the energy size of the speech signal; Separate the speech signal with the largest energy from the overlapping speech and use it as the corresponding microphone array The first recorded speech, and when the energy of the speech signal is lower than the energy threshold, the microphone array No voice signals are recorded; The first recorded speech is enhanced based on spatial filtering technology, and interference signals are suppressed.

[0024] Based on the above, during the meeting recording process, when the user When speaking, if there are other users Speaking simultaneously may cause users Microphone array Also other users The speech signals that overlap in time and space are included, which causes the user The voice signal is chaotic and causes users The text recognized and transcribed by the speech signal In one embodiment of the present invention, the feature extraction module extracts the overlapping speech features and extracts the strength relationship of multiple speech signal energies. The strength relationship is determined according to the energy size of the speech signal to obtain the microphone array The recorded voice signal with the highest energy is used as the first recorded voice. Based on this, the microphone array can automatically shield or suppress other voice signals and interference signals different from the first recorded voice, preventing the user from When speaking, the generated speech signal is disturbed, which leads to the recognition and transcription of text. It is worth noting that an energy threshold is also set, which is used to determine whether the user's voice signal needs to be recorded. It is foreseeable that in some cases, the user The voice signal does not need to be recorded, such as with other users Therefore, in this case, the user The energy of the speech signal must be different from the energy of normal speech, so it is also necessary to set an energy threshold to shield the speech signal generated in this case and prevent a large number of redundant sentences from existing in the meeting minutes; it is worth noting that the first recorded speech meets the microphone array With users That is, when the user When speaking as the target speaker, other users Microphone array All microphones are closed, and no voice signals are recorded; Specifically, existing users ,user and users The conference is composed of When speaking, users and users Also speaking, in order to avoid the microphone array In the included target speaker, that is, user When the voice signal is received, the user and users The voice signal is recorded into the microphone array The voice signal with the largest energy will be separated from the overlapping voices as the first recorded voice, that is, the user The voice signal is the first recorded voice, and the user's Voice signal, shielding the user and users Based on the speech signal, the microphone array Included users After the speech signal is input into the deep learning model, the corresponding Transcript of , to avoid transcribing the text when consulting the meeting minutes later User appears in and users The text recognized and output by the voice signal causes a messy situation.

[0025] In one embodiment, the preprocessing of the collected speech signal includes denoising, noise reduction, filtering and frame processing.

[0026] In one embodiment, the voiceprint feature of the speech signal is used to match the user in the subsequent speaking process. , the method is: Get microphone array the determined first recorded voice; Using voiceprint recognition technology, identifying and obtaining a first voiceprint feature of the first recorded voice; Microphone array The user corresponding to the first voiceprint feature to make a match.

[0027] It is worth noting that the first voiceprint feature of the first recorded voice is determined to facilitate the subsequent microphone array For users The voice signal is recorded, that is, when the user When speaking, the user can be identified based on the first voiceprint feature. Matching, so as to use the corresponding microphone array To be included; Specifically, existing users ,user and users The conference is composed of Corresponding microphone array After the first recorded voice is determined, the voiceprint feature of the first recorded voice can be extracted and recorded and used to match the user , which is also the microphone array With users To match, when the user When speaking later, the microphone array By acquiring voice signals and identifying and extracting voiceprint features in voice signals based on voiceprint recognition technology, users can be quickly recorded. There is no need to compare the energy of the speech signal, which improves the user The recognition speed of speech signals and the accuracy of transcription.

[0028] In one embodiment, if Figure 2 As shown, the microphone array The user corresponding to the first voiceprint feature Matching also includes: Based on the matched microphone array With the first voiceprint feature, user ; When the microphone array When a voice signal is recorded, the feature extraction module extracts the Nth voiceprint feature of the voice signal and determines whether the Nth voiceprint feature matches the user. The first voiceprint feature matching: If the Nth voiceprint feature does not match the first voiceprint feature, the microphone array Automatically shield the voice signal; on the contrary, when the Nth voiceprint feature matches the first voiceprint feature, the microphone array The voice signal is recorded.

[0029] Based on the above, in the microphone array With the first voiceprint feature, user After matching, the first voiceprint feature of the first recorded voice is obtained to match the user Matching to improve the corresponding microphone array For Users In one embodiment of the present invention, when the microphone array When a voice signal is recorded, the voiceprint feature of the voice signal is extracted and recorded as the Nth voiceprint feature. The Nth voiceprint feature is then compared with the microphone array. The first voiceprint feature recorded is matched. If the match is successful, the microphone array The voice signal will be recorded. Otherwise, if the match fails, the microphone array The recorded voice signal will be blocked to improve the user Corresponding microphone array The recognition speed of speech signals; It is worth noting that users The Nth voiceprint feature in the speech signal generated in the subsequent speech and the microphone array The first voiceprint feature recorded matches, so the microphone array Identify the Nth voiceprint feature as the first voiceprint feature, that is, a successful match, microphone array Will include users The voice signal generated during the subsequent speech is the first recorded voice. Speak at the same time, then the user The N+1th voiceprint feature corresponding to the speech signal and the microphone array If the first voiceprint feature recorded does not match, the microphone array Users will not be included The speech signal generated at the same time.

[0030] In one embodiment, the method for recognizing speech signals and transcribing them based on a deep learning model is: When the Nth voiceprint feature of the voice signal matches the first voiceprint feature, the voice signal is determined to be the first recorded voice; Microphone Array The speech signal is collected, and features are extracted from the speech signal to obtain speech features; Input the extracted speech features into the trained deep learning model; Analyze and recognize based on deep learning models and output transcribed text ; The deep learning model is based on convolutional neural networks and recurrent neural networks to recognize complex speech feature sequences and transcribe them into text. Output mode.

[0031] In one embodiment, a camera control module is further included, wherein the camera control module is used to adjust the camera to automatically turn to the user , and record the video; The adjustment camera automatically turns to the user The method is: Get microphone array Matching users The first voiceprint feature; When the Nth voiceprint feature of the voice signal matches the first voiceprint feature, the camera is controlled to automatically turn to the microphone array Users at location .

[0032] In microphone array Collected by users After receiving the speech signal, the speech signal will be input into the deep learning model to obtain the transcribed text corresponding to the speech signal. However, in actual application, due to the deep learning model based speech recognition and output transcription text There may still be errors, that is, individual words may be wrong, so the transcription needs to be corrected. To correct the deviation, in order to be able to The lip shape of the user is automatically assisted by the deep learning model to correct the transcribed text generated by speech recognition. , which can improve text quality and reduce error rate. In one embodiment of the present invention, when the microphone array With users After the match is successful, get the user The first voiceprint feature, when the subsequent user When speaking, by extracting the user Voice signals can automatically distinguish users position, thereby adjusting the camera to face the user First, the camera can record the progress of the meeting and generate digital meeting records based on voice transcription. Second, the user The first voiceprint feature of the voice signal when speaking is interconnected with the camera rotation and can be used by the user Control the camera to automatically turn to the user when speaking , so as to clearly record the user The mouth shape of the voice recognition is used to correct the transcribed text generated by the speech recognition. , reducing the transcription The error rate.

[0033] In one embodiment, the recorded video is used to analyze the user Detect changes in mouth shape, extract mouth shape features, and assist speech recognition and transcription based on mouth shape features; The method for assisting speech recognition and transcription based on lip shape features is: The acquired lip shape features are integrated with the speech features and input into the deep learning model; Optimizing transcription based on deep learning models The accuracy of Transcript generated by speech feature recognition , generate the corresponding lip animation; Determine how well the lip animation matches the lip features: When the lip animation is inconsistent with the lip features, the transcription text is adjusted based on semantic analysis technology , generate optimized text , the optimized text The lip animation must match the lip features; otherwise, no optimization will be performed.

[0034] Based on the above, use the camera to record the user The video of the speech will be used for subsequent user Recognize the voice signal and output the transcribed text When using The video of the speech is used to generate the transcribed text Make corrections to avoid generating transcribed text Wrong words appear in the text, improve the output transcription text In one embodiment of the present invention, when the user When speaking, the microphone array The voice signal will be recorded, and the Nth voiceprint feature of the voice signal will be matched with the first voiceprint feature, that is, the user When the match is successful, the camera will automatically turn to the corresponding user , it is conceivable that if the user When speaking, the microphone array The first voiceprint feature recorded is consistent with the user After the Nth voiceprint feature in the speech signal generated by the speaker is matched successfully, the camera will automatically turn to the microphone array , so that the microphone array Users at location Conduct video recording; Based on the recorded video, extract the user The mouth shape changes and the mouth shape features are extracted. The transcribed text generated by speech signal recognition When some wrong words appear, the corresponding lip animation can be generated and then communicated with the user. When the lip shape animation is inconsistent with the lip shape features, the transcribed text can be adjusted using semantic analysis technology based on the content expressed by the lip shape features. , as optimized text On the contrary, when the lip animation is consistent with the lip features, the transcription text is not correct. Make optimizations.

[0035] In one embodiment, a user interface module is further included to provide a user interface including a login page, a task page, a permission page, and a data record page; The login page is used for users Log in, the task page includes creating real-time tasks and offline tasks, where real-time tasks are real-time speech recognition and transcription tasks, while offline tasks are offline speech recognition and transcription tasks based on input speech signal files; the permission page is used to manage users Permissions, and only users with permissions After logging in through the login page, you can The data record page is used to display the voice records and corresponding transcribed text output by the data storage and management module .

[0036] In one embodiment, the storage and management of the voice signal input to the voice recognition and transcription module and the corresponding transcribed text The method is: Create a folder according to the project name corresponding to the real-time task or offline task; The speech signal and the corresponding transcribed text The statements in are stored in folders corresponding to the project names in chronological order; Generates logs based on folders and folder creation time, and displays the logs on the data logging page.

[0037] Based on the above, at the beginning of the meeting, the microphone array is prioritized. The first recorded voice corresponding to the first recorded voice is then used to implement the microphone array according to the first voiceprint feature of the first recorded voice. With users The match is made after identifying the user After the Nth voiceprint feature in the voice signal matches the first voiceprint feature, the voice signal is input into the deep learning model and the output transcription text In the Mark the corresponding text as a label and arrange the users in chronological order The speech signal is recognized and transcribed. When the meeting is over, the recognized and transcribed sentences are transcribed in chronological order and stored in the folder corresponding to the project name together with the speech signal. Then a log is generated according to the time when the folder was created. When you check the meeting records later, you can get the meeting time through the creation time, and check the users in the meeting through the transcribed text. 's speech.

[0038] Working principle: Using multi-channel microphone array module, setting based on multiple users Microphone array , using multiple microphone arrays Capture from different users The speech signal is enhanced by spatial filtering technology. signals from other directions, thereby effectively reducing the impact of background noise and improving user The quality of the speech signal, where the microphone array The first recorded voice determined may be a plurality of microphone arrays Specify User , in the user While speaking, other users can be automatically blocked The voice signal of the first recorded voice is recorded, and only the user corresponding to the determined first recorded voice is recorded. The voice signal is input into the deep learning model to avoid interference with the voice signal, which may cause the recognition and transcription of the text Confusion, according to the microphone array of different channels , and the corresponding records belong to different users The speech signal can ensure that the deep learning model recognizes the speech signal and outputs the transcribed text When the text is transcribed Corresponding users , which is conducive to ensuring the clarity of the meeting minutes and facilitating subsequent review of the meeting minutes.

[0039] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A speech transcription processing system, characterized in that: include: Multi-channel microphone array module for collecting user A voice signal, and determining a first recorded voice; A preprocessing module, which preprocesses the collected speech signals; A feature extraction module extracts speech features from the preprocessed speech signal, wherein the speech features include the frequency, energy and voiceprint features of the speech signal; The speech recognition and transcription module recognizes speech signals and transcribes them based on a deep learning model to output transcribed text ; The data storage and management module is used to store and manage the voice signals input into the voice recognition and transcription module and the corresponding transcribed text .

2. A speech transcription processing system according to claim 1, characterized in that: The method for determining the first recorded voice is: Arrangement corresponds to multiple users Microphone array , ; Extract speech features from overlapping speech based on a feature extraction module; The overlapping voices are for multiple users The output speech signal overlaps segments in time and space; Based on the extracted speech features, the strength relationship of multiple speech signals in the overlapping speech is analyzed, and the strength relationship is determined according to the energy size of the speech signal; Separate the speech signal with the largest energy from the overlapping speech and use it as the corresponding microphone array The first recorded speech, and when the energy of the speech signal is lower than the energy threshold, the microphone array No voice signals are recorded; The first recorded speech is enhanced based on spatial filtering technology, and interference signals are suppressed.

3. A speech transcription processing system according to claim 2, characterized in that: The preprocessing of the collected speech signal includes denoising, noise reduction, filtering and frame processing.

4. A speech transcription processing system according to claim 3, characterized in that: The voiceprint features of the speech signal are used to match the user in the subsequent speaking process. , the method is: Get microphone array the determined first recorded voice; Using voiceprint recognition technology, identifying and obtaining a first voiceprint feature of the first recorded voice; Microphone array The user corresponding to the first voiceprint feature to make a match.

5. A speech transcription processing system according to claim 4, characterized in that: The microphone array The user corresponding to the first voiceprint feature Matching also includes: Based on the matched microphone array With the first voiceprint feature, user ; When the microphone array When a voice signal is recorded, the feature extraction module extracts the Nth voiceprint feature of the voice signal and determines whether the Nth voiceprint feature matches the user. The first voiceprint feature matching: If the Nth voiceprint feature does not match the first voiceprint feature, the microphone array Automatically shield the voice signal; on the contrary, when the Nth voiceprint feature matches the first voiceprint feature, the microphone array The voice signal is recorded.

6. A speech transcription processing system according to claim 5, characterized in that: The method for recognizing speech signals and transcribing based on a deep learning model is: When the Nth voiceprint feature of the voice signal matches the first voiceprint feature, the voice signal is determined to be the first recorded voice; Microphone Array The speech signal is collected, and features are extracted from the speech signal to obtain speech features; Input the extracted speech features into the trained deep learning model; Analyze and recognize based on deep learning models and output transcribed text ; The deep learning model is based on convolutional neural networks and recurrent neural networks to recognize complex speech feature sequences and transcribe them into text. Output mode.

7. A speech transcription processing system according to claim 6, characterized in that: The camera control module is used to adjust the camera to automatically turn to the user. , and record the video; The adjustment camera automatically turns to the user The method is: Get microphone array Matching users The first voiceprint feature; When the Nth voiceprint feature of the voice signal matches the first voiceprint feature, the camera is controlled to automatically turn to the microphone array Users at location .

8. A speech transcription processing system according to claim 7, characterized in that: The recorded video is used to analyze the user Detect changes in mouth shape, extract mouth shape features, and assist speech recognition and transcription based on mouth shape features; The method for assisting speech recognition and transcription based on lip shape features is: The acquired lip shape features are integrated with the speech features and input into the deep learning model; Optimizing transcription based on deep learning models The accuracy of Transcript generated by speech feature recognition , generate the corresponding lip animation; Determine how well the lip animation matches the lip features: When the lip animation is inconsistent with the lip features, the transcription text is adjusted based on semantic analysis technology , generate optimized text , the optimized text The lip animation must match the lip features; otherwise, no optimization will be performed.

9. A speech transcription processing system according to claim 8, characterized in that: It also includes a user interface module, which provides a user interface, including a login page, a task page, a permission page, and a data record page; The login page is used for users Log in, the task page includes creating real-time tasks and offline tasks, where real-time tasks are real-time speech recognition and transcription tasks, while offline tasks are offline speech recognition and transcription tasks based on input speech signal files; the permission page is used to manage users Permissions, and only users with permissions After logging in through the login page, you can The data record page is used to display the voice records and corresponding transcribed text output by the data storage and management module .

10. A speech transcription processing system according to claim 9, characterized in that: The storage and management of the voice signals input into the voice recognition and transcription module and the corresponding transcribed text The method is: Create a folder according to the project name corresponding to the real-time task or offline task; The speech signal and the corresponding transcribed text The statements in are stored in folders corresponding to the project names in chronological order; Generates logs based on folders and folder creation time, and displays the logs on the data logging page.

Images