A system for extracting key information from speech

Through the combination of voice acquisition, preprocessing, extraction and matching modules, the problem of slow voice processing speed and low matching accuracy of voice interaction system in noisy environments is solved, and efficient and accurate voice information processing and matching is achieved, suitable for multiple environment and speech samples.

CN117935862BActive Publication Date: 2024-11-15浙江鑫淼通讯有限责任公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410267039.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-08
Publication Date
2024-11-15
Estimated Expiration
2044-03-08

AI Technical Summary

Technical Problem

The existing voice interaction system has slow voice processing speed and low speech matching accuracy in noisy environments, making it difficult to accurately identify user instructions and provide personalized services.

Method used

A voice key information extraction system is designed, including voice acquisition module, preprocessing module, key information extraction module, key information matching module and buffer determination module. Through microphone array acquisition, audio quality evaluation, segmentation processing, feature extraction, information Integration, matching judgment and semantic calibration to realize automatic collection, processing and matching of voice information.

Benefits of technology

It improves the efficiency and accuracy of voice processing, reduces labor costs, can quickly and accurately process and match voice information, ensures the integrity and matching accuracy of information, and is suitable for various environments and voice samples.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117935862B_ABST
    Figure CN117935862B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of speech extraction technology, and in particular to a speech key information extraction system, which includes: a speech acquisition module, a preprocessing module, a key information extraction module, a key information matching module and a buffer determination module; by filtering and quality evaluation of the collected speech, and segmenting and extracting feature information of the collected audio to obtain a number of key information fragments, integrating the obtained key information fragments to obtain the final key information, and semantically calibrating the final key information to obtain the key information to be matched, matching and adjusting the key information to be matched to determine the target key information in the standard information library. The present invention realizes the automatic acquisition, processing, extraction and matching of speech key information, can quickly and accurately process speech, extract information and match, and improves the speed and accuracy of speech processing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of speech extraction, and in particular to a speech key information extraction system. Background Art

[0002] In noisy stations, when the voice assistant receives the user's voice command, the large amount of human voices, vehicle noise and environmental noise have a negative impact on the voice recognition performance of the voice assistant, resulting in the voice assistant being unable to accurately recognize the user's instructions or requests, causing the voice assistant to misunderstand the user's intentions or be unable to correctly execute the user's requests. In a crowded station environment, multiple people may speak at the same time, resulting in speech overlap, making it difficult for the voice assistant to distinguish the voice commands of different users, and thus unable to provide personalized services. The noisy background environment makes it difficult for the voice assistant to understand the user's semantic intentions, increasing the difficulty of using the voice assistant.

[0003] Chinese patent application publication number CN116524932A discloses an intelligent voice interaction system and method based on artificial intelligence, comprising: a login module (1), used to provide a node for user information registration and identity authentication, and provide access rights after identity authentication; a voice recognition module (2), used to receive user input voice, perform word segmentation, classification, part-of-speech tagging and lexical analysis processing; an information extraction module (3), used to obtain key words and sentences in the voice information, mark them, and output them as target features; a training module (4), used to serve as an authentication node for the initial user, obtain the user's voiceprint features through the voice recognition module (2), guide the user to output voice through preset words and sentences, analyze the word usage and grammatical language habits, and allocate independent types of database nodes to the user account; an association matching module (5), used to collect the user's voiceprint features according to the current user The user's language habits are indexed in the database to find the appropriate answer to the question; a data judgment module (6) is used to analyze whether the current user's output voice is valid; a guidance module (7) is used to associate with the key words in the current voice conversion words and sentences when the data judgment module (6) determines that the voice is invalid, and edit the guidance voice instructions based on the association data; a voice library module (8) is used as an initial language template database to receive the input question answer, and its language system includes: corpus, rules and knowledge; a data imitation module (9) is used to imitate and transform the initial question answer associated with the association matching module (5) based on the wording and grammatical language habits obtained by the training module (4), and output the imitation data; a voice feedback module (10) is used to convert the final imitation data into the corresponding voice bar for playback.

[0004] Current voice interaction systems have slow voice processing speeds and low voice matching accuracy. Summary of the invention

[0005] To this end, the purpose of the present invention is to provide a speech key information extraction system for overcoming the problems of slow speech processing speed and low speech matching accuracy of current speech interaction systems.

[0006] To achieve the above object, the present invention provides a speech key information extraction system, comprising:

[0007] A voice collection module, used to collect a first audio and determine an actual audio quality evaluation value of the first audio;

[0008] A preprocessing module, configured to preprocess the first audio to obtain a second audio; and segment the second audio to obtain a plurality of initial audio segments;

[0009] A key information extraction module, which is used to identify the second audio and extract key information according to the identification result; it includes: a second audio segmentation processing unit, a feature extraction unit and an information integration unit; the second audio segmentation processing unit is used to segment the second audio according to the actual segmentation duration to obtain a plurality of actual audio segments; the feature extraction unit is used to extract feature information in each actual audio segment to obtain a plurality of key information segments; the information integration unit is used to integrate the key information segments to obtain the final key information;

[0010] A key information matching module, which is used to match and correct the final key information to obtain a number of actual key information and key information to be matched; it includes: a matching judgment unit and a semantic calibration unit; the matching judgment unit is used to match the final key information with a number of standard key information in the standard information library to determine a number of actual key information that meets the standard matching sensitivity, and determine the sensitivity level of the final key information; the semantic calibration unit is used to perform semantic calibration on the final key information according to the difference in the sensitivity level to obtain the key information to be matched;

[0011] A buffer determination module is used to determine target key information in each actual key information according to the key information to be matched.

[0012] Furthermore, the voice acquisition module includes: a microphone array acquisition unit, a first audio filtering unit and an audio quality assessment unit; the microphone array acquisition unit is used to collect initial audio in an external environment; the first audio filtering unit is used to identify the voice activity part in the initial audio, filter out non-human voice audio to obtain the first audio; the audio quality assessment unit is used to mark the first audio to obtain different types of marked audio and determine the actual audio quality assessment value according to the marking result.

[0013] Furthermore, the preprocessing module includes: a second audio filtering unit, a first audio segmentation processing unit and an audio segmentation correction unit; the second audio filtering unit is used to determine whether audio re-collection is required according to the actual audio quality evaluation value, and to obtain the second audio according to the standard human voice spectrum diagram and each marked audio in the first audio; the first audio segmentation processing unit is used to segment the second audio according to a preset fixed segmentation time length to obtain a plurality of initial audio segments; the audio segmentation correction unit is used to correct the fixed segmentation time length according to the number of execution verbs in each initial audio segment to obtain the actual segmentation time length.

[0014] Furthermore, the first audio filtering unit is used to identify the voice activity part to obtain the audio frequencies of different sound sources in the external environment;

[0015] For any sound source, if the audio frequency corresponding to the sound source presents a complex waveform feature, the first audio filtering unit determines that the sound source is a human voice and retains the sound source;

[0016] If the audio frequency corresponding to the sound source presents a single waveform feature, the first audio filtering unit determines that the sound source is non-human voice and deletes the sound source.

[0017] Further, the audio quality assessment unit marks the repeated audio segments, the blank audio segments and the blank audio segments in the first audio as: a type I mark, a type II mark and a type III mark in sequence;

[0018] The audio quality evaluation unit determines the actual audio quality evaluation value according to the number of class I marks, the number of class II marks, and the number of class III marks.

[0019] Further, the second audio filtering unit deletes, skips or repairs the first audio according to each mark type to obtain a first processed audio;

[0020] The second audio filtering unit determines the single-character human voice on the first processed audio; compares the voiceprint spectrogram corresponding to each human voice audio in the first processed audio with the standard human voice spectrogram;

[0021] For any voiceprint spectrum, if the voiceprint spectrum is consistent with the standard human voice spectrum Figure 1 If the voiceprint spectrum is consistent with the human voice audio, the second audio filtering unit determines that the human voice audio corresponding to the voiceprint spectrum is the second audio.

[0022] Further, for any initial audio segment, the audio segmentation correction unit determines a correction method for the fixed segmentation duration according to the number of the execution verbs in the initial audio segment;

[0023] If the number of the execution verbs is greater than the audio segmentation evaluation value, the audio segmentation correction unit determines that the correction method for the fixed segmentation time length is positive feedback correction; wherein the positive feedback correction is to extend the fixed segmentation time length;

[0024] If the number of the execution verbs is less than or equal to the audio segmentation evaluation value, the audio segmentation correction unit determines that the correction method for the fixed segmentation time length is negative feedback correction; wherein the negative feedback correction is to shorten the fixed segmentation time length;

[0025] The audio segment evaluation value is set in the audio segment correction unit.

[0026] Furthermore, the audio segmentation correction unit classifies and integrates each initial audio segment according to the type of the correction method, including: a number of positive feedback points and a number of negative feedback points; and corrects the fixed segmentation duration according to the number of positively executed verbs and the number of negatively executed verbs in the initial audio segment corresponding to each positive feedback point and each negative feedback point to obtain the actual segmentation duration.

[0027] Further, for any standard key information, the matching judgment unit determines the initial matching sensitivity according to the final key information and the standard key information;

[0028] The matching judgment unit determines whether the standard key information is the actual key information and the sensitivity level of the final key information according to the absolute value of the first difference;

[0029] If the absolute value of the first difference is less than or equal to the sensitivity level evaluation value, the matching judgment unit determines that the standard key information is the actual key information and determines that the sensitivity level of the final key information is level one;

[0030] If the absolute value of the first difference is greater than the sensitivity level evaluation value and the initial matching sensitivity is greater than the standard matching sensitivity; the matching judgment unit determines that the standard key information is the actual key information and determines that the sensitivity level of the final key information is level 2;

[0031] The sensitivity level evaluation value is set in the matching judgment unit; the first difference absolute value is the absolute value of the difference between the initial matching sensitivity and the standard matching sensitivity.

[0032] Further, if the sensitivity level of the final key information is level one, the semantic calibration unit performs semantic calibration on the final key information to obtain the key information to be matched;

[0033] For any actual key information, the buffer determination module determines whether the actual key information is the target key information according to the key information to be matched and the actual key information;

[0034] If the key information to be matched is consistent with the actual key information, the buffer determination module determines that the actual key information is the target key information.

[0035] Compared with the prior art, the beneficial effects of the present invention are that, through the combination of a voice acquisition module, a preprocessing module, a key information extraction module, and a key information matching module, the present invention realizes the automatic acquisition, processing, extraction and matching of voice information, thereby improving efficiency and reducing labor costs; the various modules cooperate with each other, and can quickly and accurately process, extract and match voice, thereby improving processing speed and accuracy; by providing an audio quality assessment unit and related processing modules, the quality of the collected voice can be assessed and calibrated, thereby ensuring the accuracy of subsequent information extraction and matching; through the information integration unit, various key information fragments can be integrated to obtain the final key information, making the information more complete and comprehensive; through the matching judgment unit and the semantic calibration unit, the key information can be matched and adjusted according to actual conditions, thereby improving the accuracy and flexibility of matching; through the buffer determination module, the target key information can be determined according to the key information to be matched, thereby facilitating the accurate positioning of the key information and improving the accuracy of matching.

[0036] In particular, several sound sources in the external environment are identified and determined by the first audio filtering unit, and the sound sources determined to be human voices are retained to obtain the first audio composed of multiple characters; by analyzing and judging the waveform characteristics of the audio frequency, different sound sources, such as human voices and non-human voices, can be effectively distinguished, thereby achieving accurate filtering and retention; and by using the waveform characteristics for judgment, the sound sources can be intelligently classified, which not only improves the accuracy of screening but also simplifies the processing flow; by judging the waveform characteristics presented by the audio frequency, the properties of the sound source can be more accurately determined, and then the corresponding retention or deletion operations can be performed, effectively avoiding misjudgment; through automated audio frequency recognition and processing, the sound source can be efficiently filtered, saving the time and labor cost of manual processing; judging the waveform characteristics of different sound sources can be applicable to various environments and voice samples, and has certain versatility and flexibility; it achieves high efficiency and accuracy in processing audio frequencies, identifying different sound sources and filtering processing, providing effective data preparation and screening guarantees for subsequent processing.

[0037] In particular, by marking and calculating the weights of different types of audio clips (repeated, blank, blurred), the quality of the audio can be comprehensively evaluated and various problems can be quantitatively evaluated; by introducing weight coefficients of different types of markings, the weights of various types of problems in the final evaluation value can be adjusted according to actual conditions, making the evaluation closer to actual conditions; the actual audio quality evaluation value obtained by calculation can quantify the degree of audio quality, which is convenient for further processing and decision-making; and according to the number of different types of markings, the actual audio quality evaluation value is automatically calculated without human intervention, thereby improving the evaluation efficiency and accuracy; it achieves effective evaluation of audio quality, provides important reference information for subsequent processing, and helps improve audio processing efficiency and quality.

[0038] In particular, the second audio filtering unit determines whether audio re-collection is required based on the actual audio quality evaluation value, which helps to automate the process and ensure that the sound quality meets certain standards. When it is determined that the audio does not need to be re-collected, the second audio filtering unit processes the audio segment according to the marking result of the first audio, and achieves efficient processing of audio data by deleting, skipping or repairing the corresponding type of audio segment; the second audio filtering unit can determine the human voice of a single role, that is, by comparing the sound spectrum with the standard human voice spectrum, the human voice part in the audio is determined, which helps to identify and separate effective human voice audio; and the first audio segmentation processing unit sets a fixed segmentation time and corrects it according to the number of execution verbs to optimize the audio segmentation method so that it can better adapt to the processing requirements of different audio content. For different correction methods (positive feedback and negative feedback), the second audio filtering unit integrates and renumbers all initial audio segments to uniformly manage and process various types of audio segments, thereby improving processing efficiency and accuracy; the screening, correction and optimization of audio data are achieved, providing key support for the final acquisition of high-quality audio data.

[0039] In particular, by calculating the initial matching sensitivity and the absolute value of the first difference, matching judgment and sensitivity level evaluation of key information can be implemented according to actual conditions; this makes it possible to adapt to different matching requirements and sensitivity levels; and helps to improve the accuracy and reliability of matching; based on the size of the absolute value of the first difference and the relative size of the initial matching sensitivity, the sensitivity level of the final key information can be determined to improve the matching degree and credibility, and when the sensitivity level of the final key information is level one, semantic calibration is performed on it to further improve the accuracy and semantic consistency of the key information, which helps to ensure the quality and availability of the final key information. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic diagram of the structure of a system for extracting key speech information according to an embodiment of the present invention;

[0041] Figure 2 This is a structural diagram of a voice acquisition module in a voice key information extraction system according to an embodiment of the present invention;

[0042] Figure 3 A logic diagram for obtaining a first audio in a system for extracting key voice information according to an embodiment of the present invention;

[0043] Figure 4 A logic diagram for evaluating the sensitivity level of final key information in the voice key information extraction system according to an embodiment of the present invention;

[0044] The figure includes: a voice acquisition module 1, a preprocessing module 2, a key information extraction module 3, a key information matching module 4, a buffer determination module 5, a microphone array acquisition unit 11, a first audio filtering unit 12 and an audio quality evaluation unit 13. DETAILED DESCRIPTION

[0045] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0046] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0047] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is merely for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0048] See also Figure 1-Figure 4 As shown, Figure 1 A schematic diagram of the structure of a system for extracting key speech information according to an embodiment of the present invention; Figure 2 This is a structural diagram of a voice acquisition module in a voice key information extraction system according to an embodiment of the present invention; Figure 3 A logic diagram for obtaining a first audio in a system for extracting key voice information according to an embodiment of the present invention; Figure 4 This is a logic diagram for evaluating the final key information sensitivity level in the speech key information extraction system according to an embodiment of the present invention.

[0049] The present invention provides a speech key information extraction system, comprising:

[0050] A speech acquisition module 1 is used to acquire a first audio and determine an actual audio quality evaluation value of the first audio; it comprises: a microphone array acquisition unit 11, a first audio filtering unit 12 and an audio quality evaluation unit 13; the microphone array acquisition unit 11 is used to acquire an initial audio in an external environment; the first audio filtering unit 12 is used to identify a speech activity part in the initial audio and filter out non-human voice audio to obtain a first audio; the audio quality evaluation unit 13 is used to mark the first audio to obtain different types of marked audio and determine an actual audio quality evaluation value according to the marking result;

[0051] A preprocessing module 2, which is used to preprocess the first audio to obtain a second audio; and to segment the second audio to obtain a number of initial audio segments; it includes: a second audio filtering unit, a first audio segmentation processing unit and an audio segmentation correction unit; the second audio filtering unit is used to determine whether audio re-collection is required according to the actual audio quality evaluation value, and to obtain the second audio according to the standard human voice spectrum diagram and each marked audio in the first audio; the first audio segmentation processing unit is used to segment the second audio according to a preset fixed segmentation time length to obtain a number of initial audio segments; the audio segmentation correction unit is used to correct the fixed segmentation time length according to the number of executive verbs in each initial audio segment to obtain the actual segmentation time length;

[0052] The key information extraction module 3 is used to identify the second audio and extract key information according to the identification result; it includes: a second audio segmentation processing unit, a feature extraction unit and an information integration unit; the second audio segmentation processing unit is used to segment the second audio according to the actual segmentation duration to obtain a plurality of actual audio segments; the feature extraction unit is used to extract feature information in each actual audio segment to obtain a plurality of key information segments; the information integration unit is used to integrate each key information segment to obtain the final key information;

[0053] The key information matching module 4 is used to match and correct the final key information to obtain a number of actual key information and key information to be matched; it includes: a matching judgment unit and a semantic calibration unit; the matching judgment unit is used to determine a number of actual key information that meets the standard matching sensitivity according to the matching results of the final key information and a number of standard key information in the standard information library, and determine the sensitivity level of the final key information; the semantic calibration unit is used to perform semantic calibration on the final key information according to the difference of the sensitivity level to obtain the key information to be matched;

[0054] The buffer determination module 5 is used to determine the target key information in each actual key information according to the key information to be matched.

[0055] The embodiment of the present invention realizes the automatic collection, processing, extraction and matching of voice information through the combination of a voice collection module, a preprocessing module, a key information extraction module and a key information matching module, thereby improving efficiency and reducing labor costs; the modules cooperate with each other, and can quickly and accurately process, extract and match voice, thereby improving processing speed and accuracy; by providing an audio quality evaluation unit and related processing modules, the quality of the collected voice can be evaluated and calibrated, thereby ensuring the accuracy of subsequent information extraction and matching; by means of an information integration unit, various key information fragments can be integrated to obtain the final key information, thereby making the information more complete and comprehensive; by means of a matching judgment unit and a semantic calibration unit, key information can be matched and adjusted according to actual conditions, thereby improving the accuracy and flexibility of matching; by means of a buffer determination module, target key information can be determined according to the key information to be matched, thereby facilitating accurate positioning of key information and improving the accuracy of matching.

[0056] Specifically, in this embodiment, the first audio filtering unit 12 recognizes the voice activity part in the initial audio to obtain the audio frequencies of different sound sources in the external environment;

[0057] For any sound source, if the audio frequency corresponding to the sound source presents a complex waveform feature, the first audio filter unit 12 determines that the sound source is a human voice and retains the sound source;

[0058] If the audio frequency corresponding to the sound source presents a single waveform feature, the first audio filtering unit 12 determines that the sound source is non-human voice and deletes the sound source.

[0059] In this embodiment, a plurality of sound sources in the external environment are identified and determined by the first audio filtering unit, and the sound sources determined to be human voices are retained to obtain a first audio composed of multiple roles; by analyzing and judging the waveform characteristics of the audio frequency, different sound sources, such as human voices and non-human voices, can be effectively distinguished, thereby achieving accurate filtering and retention; and by using the waveform characteristics for judgment, the sound sources can be intelligently classified, which not only improves the accuracy of screening but also simplifies the processing flow; by judging the waveform characteristics presented by the audio frequency, the properties of the sound source can be more accurately determined, and then the corresponding retention or deletion operations can be performed, effectively avoiding misjudgment; through automated audio frequency recognition and processing, the sound source can be efficiently filtered, saving the time and labor cost of manual processing; judging the waveform characteristics of different sound sources can be applicable to various environments and voice samples, and has certain versatility and flexibility; it achieves high efficiency and accuracy in processing audio frequencies, identifying different sound sources and filtering processing, and provides effective data preparation and screening guarantees for subsequent processing.

[0060] Specifically, the audio quality assessment unit 13 in this embodiment marks the first audio;

[0061] When it is determined that a repeated audio segment exists in the first audio, the repeated audio segment is marked and recorded as a Class I mark;

[0062] When it is determined that a blank audio segment appears in the first audio, the blank audio segment is marked and recorded as a Class II mark;

[0063] When it is determined that a blank audio segment appears in the first audio, the ambiguous audio segment is marked and recorded as a Class III mark;

[0064] The audio quality evaluation unit 13 calculates the actual audio quality evaluation value K according to the number A1 of class I marks, the number A2 of class II marks and the number A3 of class III marks;

[0065] K=(A1×a1+A2×a2+A3×a3) / 100;

[0066] Among them, a1 is the first weight coefficient of the number of Class I marks A1 to the actual audio quality assessment value K; a2 is the second weight coefficient of the number of Class II marks A2 to the actual audio quality assessment value K; a3 is the third weight coefficient of the number of Class III marks A3 to the actual audio quality assessment value K.

[0067] In a specific implementation process, for example, if the first weight coefficient corresponding to the number of repeated segments is set to 0.4, the second weight coefficient corresponding to the number of blank segments is set to 0.3, and the third weight coefficient corresponding to the number of blurred segments is set to 0.3, and there are a total of 100 segments in the audio, of which there are 10 repeated segments, 5 blank segments and 3 blurred segments, then the actual audio quality evaluation value K is calculated to be (0.4×10+0.3×5+0.3×3) / 100=0.16; the actual audio quality evaluation value obtained is 0.16, indicating that the quality of the first audio is relatively low; of course, in this embodiment, the first weight coefficient is set to The actual weights and calculation methods may vary depending on specific requirements. For example, the value of the weight coefficient is also affected by the real-time sampling rate, where the real-time sampling rate is the number of times the microphone array acquisition unit samples per second. In this embodiment, the real-time sampling rate is 80%. The real-time sampling rate is related to the sampling quantization accuracy index of the application scenario. For example, when the application scenario is real-time diplomatic translation, the real-time sampling rate at this time reaches more than 98%. The first weight coefficient should be set to 0.6, the second weight coefficient should be set to 0.8, and the third weight coefficient should be set to 0.9.

[0068] The embodiment of the present invention can comprehensively evaluate the quality of audio and perform quantitative evaluation on various problems by marking and weighting different types of audio clips (repeated, blank, blurred); by introducing weight coefficients of different types of markings, the weights of various types of problems in the final evaluation value can be adjusted according to actual conditions, so that the evaluation is closer to the actual conditions; the actual audio quality evaluation value obtained by calculation can quantitatively represent the quality of the audio, which is convenient for further processing and decision-making; and according to the number of different types of markings, the actual audio quality evaluation value is automatically calculated without manual intervention, thereby improving the evaluation efficiency and accuracy; it realizes the effective evaluation of audio quality, provides important reference information for subsequent processing, and helps improve the efficiency and quality of audio processing.

[0069] Specifically, in this embodiment, the second audio filtering unit determines whether to perform audio re-collection according to the actual audio quality evaluation value;

[0070] If the actual audio quality evaluation value is less than or equal to 0.5, the second audio filtering unit determines that audio re-capture is required;

[0071] If the actual audio quality evaluation value is greater than 0.5, the second audio filtering unit determines that audio re-capture is not required;

[0072] In this embodiment, the standard audio quality evaluation value is set to 0.5. The specific setting is also affected by the sampling quantization accuracy index of the application scenario. In this embodiment, the sampling quantization accuracy index of the application scenario is that the lossless fidelity of the collected audio within one audio collection cycle must reach more than 93% of the original audio. The sampling quantization accuracy index of the application scenario is set according to the noisy environment and the type of application scenario. For example, when the application environment is ordinary real-time conversation translation and the surrounding environment is quiet and noise-free, the sampling quantization accuracy index of the application scenario at this time is 85%.

[0073] Specifically, in this embodiment, when it is determined that the first audio does not need to be re-collected, the second audio filtering unit processes the first audio according to the marking result;

[0074] For the Class I mark, the second audio filtering unit deletes the marked audio segment;

[0075] For the type II mark, the second audio filtering unit skips the marked audio segment;

[0076] For the Class III mark, the second audio filtering unit repairs the marked audio segment;

[0077] The second audio filtering unit performs single-character human voice determination on the processed first processed audio;

[0078] Comparing the voiceprint spectrogram corresponding to each human voice audio in the first processed audio with the standard human voice spectrogram;

[0079] For any voiceprint spectrum, if the voiceprint spectrum is consistent with the standard human voice spectrum Figure 1 If the second audio filtering unit determines that the human voice audio corresponding to the voiceprint spectrum graph is the second audio;

[0080] For any voiceprint spectrum graph, if the voiceprint spectrum graph is inconsistent with the standard human voice spectrum graph, the second audio filtering unit determines that the human voice audio corresponding to the voiceprint spectrum graph is not the second audio;

[0081] Among them, the standard human voice spectrogram is arranged in the second audio filtering unit; the standard human voice spectrogram is a standardized spectrum representing the sound characteristics of a specific human role, which collects a large number of voice samples of the specific human role by pre-interacting with the voice acquisition module 1, and performs statistical analysis on the voice samples, so as to describe the characteristics of the standard human voice spectrogram corresponding to the specific human role according to the statistics. In this embodiment, the statistics include: the mean and standard deviation of each frequency band. Of course, other statistics may also be used, such as the maximum value and the minimum value, etc., which are not limited in this embodiment. According to the analysis results of the statistics, the parameters of the standard human voice spectrogram are determined, and the parameters include: spectrum range, spectrum resolution and spectrum smoothness.

[0082] Specifically, in this embodiment, the first audio segmentation processing unit is provided with the fixed segmentation time length t0, and the first audio segmentation processing unit segments the second audio according to the fixed segmentation time length t0 to obtain a plurality of initial audio segments, including: a first initial audio segment B1, a second initial audio segment B2, ..., an nth initial audio segment Bn;

[0083] Among them, in this embodiment, the fixed segmentation time t0 is set to 5s, and the specific setting is also affected by the audio segmentation standard. The audio segmentation standard is set according to the application scenario and the audio signal. For example, when the application scenario is music analysis, the fixed segmentation time should be set to 10s to ensure the integrity of the music, and the characteristics of the audio signal will also affect the setting of the fixed segmentation time. For example, when the speech paragraphs in the audio are short and appear frequently, a shorter fixed segmentation time is set to ensure that each segment contains complete speech information, and when the speech paragraphs in the audio are long and appear less frequently, a longer fixed segmentation time is set.

[0084] Specifically, in this embodiment, for any initial audio segment Bi, i=1, 2...n, the audio segment modification unit determines a modification method for the fixed segmentation time length t0 according to the number of executive verbs Xi in the initial audio segment Bi;

[0085] If Xi>X0, the audio segmentation correction unit determines that the correction method for the fixed segmentation time length t0 is positive feedback correction; the positive feedback correction is to extend the fixed segmentation time length t0;

[0086] If Xi≤X0, the audio segmentation correction unit determines that the correction method for the fixed segmentation time length t0 is negative feedback correction; the negative feedback correction is to shorten the fixed segmentation time length t0;

[0087] Wherein, X0 is the audio segment evaluation value set in the audio segment correction unit;

[0088] During the specific implementation process, the audio segmentation evaluation value is set to 3. The specific setting is also affected by the accuracy index required by the application scenario. In this embodiment, the accuracy index required by the application scenario is that the accuracy of identifying key information in the audio within a speech key information extraction cycle must reach more than 95%. The accuracy index required by the application scenario is set according to the audio task type. For example, when the audio task type is more difficult, the application scenario accuracy index required at this time is 96%, and the audio segmentation evaluation value should be set to 1.

[0089] Specifically, in this embodiment, the correction methods for all the initial audio clips are integrated, including:

[0090] 1) Renumbering the initial audio clip corrected by the positive feedback, recording as the first positive feedback point B11, the second positive feedback point B12, ... the mth positive feedback point B1m;

[0091] 2) renumbering the initial audio segment corrected by negative feedback, recording as the first negative feedback point B21, the second negative feedback point B22, ... the kth negative feedback point B2k;

[0092] The audio segmentation correction unit determines the actual segmentation duration t according to the number of executive verbs in each initial audio segment;

[0093]

[0094] Wherein, p=1, 2...m; q=1, 2...k; Xp is the number of positively executed verbs in the initial audio segment corresponding to the pth positive feedback point B1p; Xq is the number of negatively executed verbs in the initial audio segment corresponding to the qth negative feedback point B2q; a is the positive feedback correction coefficient; b is the negative feedback correction coefficient;

[0095] In the specific implementation process, the positive feedback correction coefficient a is set to 0.6, and the negative feedback correction coefficient is set to 0.4. The specific settings are also affected by the sampling quantization accuracy index of the application scenario. In this embodiment, the sampling quantization accuracy index of the application scenario is that the lossless fidelity of the collected audio within one audio collection cycle must reach more than 93% of the original audio. The sampling quantization accuracy index of the application scenario is set according to the noisy environment and the type of application scenario. For example, when the application environment is ordinary real-time conversation translation and the surrounding environment is quiet and noise-free, the sampling quantization accuracy index of the application scenario at this time is 85%. At this time, the positive feedback coefficient a should be set to 0.7, and the negative feedback coefficient b should be set to 0.2.

[0096] In the embodiment of the present invention, the second audio filtering unit determines whether audio re-collection is required based on the judgment of the actual audio quality evaluation value, which is helpful for automating the process and ensuring that the sound quality meets certain standards. When it is determined that the audio does not need to be re-collected, the second audio filtering unit processes the audio segment according to the marking result of the first audio, and achieves efficient processing of the audio data by deleting, skipping or repairing the corresponding type of audio segment; the second audio filtering unit can determine the human voice of a single role, that is, by comparing the sound spectrum with the standard human voice spectrum, the human voice part in the audio is determined, which is helpful for identifying and separating effective human voice audio; and the first audio segmentation processing unit sets a fixed segmentation time and corrects it according to the number of execution verbs to optimize the audio segmentation method, so that it can better adapt to the processing requirements of different audio contents. For different correction methods (positive feedback and negative feedback), the second audio filtering unit integrates and renumbers all initial audio segments to uniformly manage and process various types of audio segments, thereby improving processing efficiency and accuracy; the screening, correction and optimization of audio data are achieved, providing key support for finally obtaining high-quality audio data.

[0097] Specifically, in this embodiment, the second audio segmentation processing unit segments the second audio according to the actual segmentation duration t to obtain a plurality of actual audio segments;

[0098] For any actual audio segment, the feature extraction unit extracts the feature information in the actual audio segment; the feature information includes: an execution subject, an execution verb and an execution object;

[0099] The feature extraction unit extracts according to the feature information to obtain a plurality of key information fragments; the information integration unit is used to integrate the key information fragments to obtain the final key information;

[0100] For example, when any key information fragment is: [executing subject 1 and executing subject 2, executing verb 1, executing object 1 and executing object 2], and another key information fragment is: [executing subject 3, executing verb 2, executing object 3]; the information integration unit integrates the key information fragments to form the final key information:

[0101]

[0102] Specifically, in this embodiment, the standard information library stores a number of standard key information; for any standard key information, the matching judgment unit calculates the initial matching sensitivity D1 according to the final key information and the standard key information;

[0103] D1 = f (final key information, standard key information);

[0104] The matching judgment unit calculates a first difference absolute value S1=|D0-D1| according to the initial matching sensitivity D1 and the standard matching sensitivity D0;

[0105] If S1≤S10, the matching judgment unit determines that the standard key information is the actual key information and determines that the sensitivity level of the final key information is level one;

[0106] If S1>S10 and D1>D0; the matching judgment unit determines that the standard key information is the actual key information and determines that the sensitivity level of the final key information is level 2;

[0107] If S1>S10 and D1<D0; then the matching judgment unit determines that the standard key information is not the actual key information;

[0108] Wherein, the matching judgment unit is provided with the standard matching sensitivity D0 and the sensitivity level evaluation value S10;

[0109] During the specific implementation process, the standard matching sensitivity D0 is set to 88%, and the sensitivity level evaluation value S10 is set to 5%; the specific setting is also affected by the accuracy index required by the application scenario. In this embodiment, the accuracy index required by the application scenario is that the accuracy of identifying key information in the audio within a speech key information extraction cycle must reach more than 95%. The accuracy index required by the application scenario is set according to the audio task type. For example, when the audio task type is more difficult and contains physical formulas, the application scenario accuracy index required at this time is 99%, then the standard matching sensitivity D0 should be set to 98%, and the sensitivity level evaluation value S10 is set to 2%.

[0110] When the sensitivity level of the final key information is level one, the semantic calibration unit performs semantic calibration on the final key information to obtain the key information to be matched.

[0111] The embodiment of the present invention calculates the initial matching sensitivity and the absolute value of the first difference to achieve matching judgment and sensitivity level evaluation of key information according to actual conditions; it can adapt to different matching requirements and sensitivity levels; and help to improve the accuracy and reliability of matching; according to the size of the absolute value of the first difference and the relative size of the initial matching sensitivity, the sensitivity level of the final key information can be determined to improve the matching degree and credibility, and when the sensitivity level of the final key information is level one, it is semantically calibrated to further improve the accuracy and semantic consistency of the key information, which helps to ensure the quality and availability of the final key information.

[0112] Specifically, in this embodiment, for any actual key information, the buffer determination module 5 determines whether the actual key information is the target key information according to the key information to be matched and the actual key information;

[0113] If the key information to be matched is consistent with the actual key information, the buffer determination module 5 determines that the actual key information is the target key information;

[0114] If the key information to be matched is inconsistent with the actual key information, the buffer determination module 5 determines that the actual key information is not the target key information.

[0115] The calculation compensation parameters and the calculation adjustment parameters described in the present invention have two functions: one is to balance the left and right dimensions of the formula, and the other is to adjust the numerical results. No specific assignment is performed in this embodiment. In addition, the calculation formulas in this embodiment are used to intuitively reflect the adjustment relationship between the numerical values, such as positive correlation and negative correlation. Unless otherwise specified, the parameter values ​​that are not specifically limited to numerical values ​​are all positive.

[0116] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A speech key information extraction system, characterized in that: include: A voice collection module, used to collect a first audio and determine an actual audio quality evaluation value of the first audio; A preprocessing module, configured to preprocess the first audio to obtain a second audio; and segment the second audio to obtain a plurality of initial audio segments; A key information extraction module, which is used to recognize the second audio and extract key information according to the recognition result; It includes: a second audio segmentation processing unit, a feature extraction unit and an information integration unit; the second audio segmentation processing unit is used to segment the second audio according to the actual segmentation duration to obtain a plurality of actual audio segments; the feature extraction unit is used to extract feature information in each actual audio segment to obtain a plurality of key information segments; the information integration unit is used to integrate the key information segments to obtain the final key information; The characteristic information includes: execution subject, execution verb and execution object; the key information segment is [execution subject, execution verb, execution object]; A key information matching module, which is used to match and correct the final key information to obtain a number of actual key information and key information to be matched; it includes: a matching judgment unit and a semantic calibration unit; the matching judgment unit is used to match the final key information with a number of standard key information in the standard information library to determine a number of actual key information that meets the standard matching sensitivity, and determine the sensitivity level of the final key information; the semantic calibration unit is used to perform semantic calibration on the final key information according to the difference in the sensitivity level to obtain the key information to be matched; A buffer determination module, which is used to determine target key information in each actual key information according to the key information to be matched; The voice acquisition module includes: a microphone array acquisition unit, a first audio filtering unit and an audio quality evaluation unit; the microphone array acquisition unit is used to collect initial audio in an external environment; the first audio filtering unit is used to identify the voice activity part in the initial audio, filter out non-human voice audio to obtain the first audio; the audio quality evaluation unit is used to mark the first audio to obtain different types of marked audio and determine the actual audio quality evaluation value according to the marking result; The preprocessing module includes: a second audio filtering unit, a first audio segmentation processing unit and an audio segmentation correction unit; the second audio filtering unit is used to determine whether audio re-collection is required according to the actual audio quality evaluation value, and obtain the second audio according to the standard human voice spectrum diagram and each marked audio in the first audio; the first audio segmentation processing unit is used to segment the second audio according to a preset fixed segmentation time length to obtain a plurality of initial audio segments; the audio segmentation correction unit is used to correct the fixed segmentation time length according to the number of executive verbs in each initial audio segment to obtain an actual segmentation time length; For any standard key information, the matching judgment unit determines the initial matching sensitivity according to the final key information and the standard key information; The matching judgment unit determines whether the standard key information is the actual key information and the sensitivity level of the final key information according to the absolute value of the first difference; If the absolute value of the first difference is less than or equal to the sensitivity level evaluation value, the matching judgment unit determines that the standard key information is the actual key information and determines that the sensitivity level of the final key information is level one; If the absolute value of the first difference is greater than the sensitivity level evaluation value and the initial matching sensitivity is greater than the standard matching sensitivity; the matching judgment unit determines that the standard key information is the actual key information and determines that the sensitivity level of the final key information is level 2; The sensitivity level evaluation value is set in the matching judgment unit; the first difference absolute value is the absolute value of the difference between the initial matching sensitivity and the standard matching sensitivity.

2. The speech key information extraction system according to claim 1, characterized in that: The first audio filter unit is used to identify the voice activity part to obtain the audio frequencies of different sound sources in the external environment; For any sound source, if the audio frequency corresponding to the sound source presents a complex waveform feature, the first audio filtering unit determines that the sound source is a human voice and retains the sound source; If the audio frequency corresponding to the sound source presents a single waveform feature, the first audio filtering unit determines that the sound source is non-human voice and deletes the sound source.

3. The speech key information extraction system according to claim 2, characterized in that: The audio quality assessment unit marks the repeated audio segments, the blank audio segments and the blank audio segments in the first audio as: a Class I mark, a Class II mark and a Class III mark in sequence; The audio quality evaluation unit can determine the actual audio quality evaluation value according to the number of class I marks, the number of class II marks and the number of class III marks.

4. The speech key information extraction system according to claim 3, characterized in that: The second audio filtering unit deletes, skips or repairs the first audio according to each mark type to obtain a first processed audio; The second audio filtering unit determines the single-character human voice on the first processed audio; compares the voiceprint spectrogram corresponding to each human voice audio in the first processed audio with the standard human voice spectrogram; For any voiceprint spectrum graph, if the voiceprint spectrum graph is consistent with the standard human voice spectrum graph, the second audio filtering unit determines that the human voice audio corresponding to the voiceprint spectrum graph is the second audio.

5. The speech key information extraction system according to claim 4, characterized in that: For any initial audio segment, the audio segmentation correction unit determines a correction method for the fixed segmentation duration according to the number of the executive verbs in the initial audio segment; If the number of the execution verbs is greater than the audio segmentation evaluation value, the audio segmentation correction unit determines that the correction method for the fixed segmentation time length is positive feedback correction; wherein the positive feedback correction is to extend the fixed segmentation time length; If the number of the execution verbs is less than or equal to the audio segmentation evaluation value, the audio segmentation correction unit determines that the correction method for the fixed segmentation time length is negative feedback correction; wherein the negative feedback correction is to shorten the fixed segmentation time length; The audio segment evaluation value is set in the audio segment correction unit.

6. The speech key information extraction system according to claim 5, characterized in that: The audio segmentation correction unit classifies and integrates each initial audio segment according to the type of the correction method, including: a plurality of positive feedback points and a plurality of negative feedback points; and corrects the fixed segmentation duration according to the number of positively executed verbs and the number of negatively executed verbs in the initial audio segment corresponding to each positive feedback point and each negative feedback point to obtain the actual segmentation duration.

7. The speech key information extraction system according to claim 6, characterized in that: If the sensitivity level of the final key information is level one, the semantic calibration unit performs semantic calibration on the final key information to obtain the key information to be matched; For any actual key information, the buffer determination module determines whether the actual key information is the target key information according to the key information to be matched and the actual key information; If the key information to be matched is consistent with the actual key information, the buffer determination module determines that the actual key information is the target key information.