News live voice anomaly real-time monitoring correction method and system thereof

Abstract

The application relates to the technical field of speech signal processing and live broadcast quality control, and discloses a news live broadcast speech abnormality real-time monitoring and correction method and system, which comprises the following steps: extracting a multi-dimensional acoustic feature vector from a real-time frame of a broadcaster's speech signal and storing the feature vector in a ring feature buffer; synchronously performing parallel detection of abnormal pause detection, speech error deviation detection and volume mutation detection based on the feature vector; pushing a prompt sound through an ear return, highlighting a deviation position on a teleprompter or triggering a dynamic compressor to smooth the volume according to the type of an abnormal event; recording the abnormal event to generate a broadcast quality analysis report and correcting the detection threshold according to the report; and the application realizes millisecond-level real-time perception and instant correction assistance for broadcasting abnormalities, and significantly shortens the response delay of traditional manual monitoring.

Description

Technical Field

[0001] This invention relates to the field of speech signal processing and live broadcast quality control technology, and in particular to a method and system for real-time monitoring and correction of abnormal speech in news live broadcasts. Background Technology

[0002] In live television and online news broadcasts, the quality of the broadcast directly impacts the accuracy of information delivery and the credibility of the media. During live broadcasts, announcers may experience speech errors such as repeated slips of the tongue, abnormal pauses exceeding normal intervals, and sudden changes in volume due to sudden tension, equipment malfunctions, or last-minute changes to the script. Once these anomalies are transmitted to viewers, they severely affect the professionalism of the news program and the audience's listening experience. This is especially true during live coverage of major breaking news events, where announcers need to process constantly updated script information under high pressure, significantly increasing the probability of speech errors and posing a more severe challenge to broadcast quality.

[0003] In existing technologies, audio anomaly handling for live streaming scenarios mainly focuses on content compliance monitoring. For example, Chinese invention application CN110085213A discloses a method for monitoring audio anomalies. This method acquires segmented audio corresponding to a target audio program being played in real time, identifies these segmented audio segments as segmented text, performs anomaly identification on the segmented text to discover illegal content, and generates a pending review record based on the identified abnormal text for manual review. This solution has certain practical value in monitoring audio content compliance. It can automatically screen illegal audio content in live streams and short videos through speech-to-text technology and sensitive word database matching, and pushes the screening results to the review platform for reviewers to process, thereby reducing the workload of manual review and improving the efficiency of discovering illegal content to a certain extent.

[0004] However, the aforementioned existing technical solutions have the following shortcomings: First, the detection target of this solution is limited to the identification of illegal content at the text semantic level, relying on keyword comparison using a sensitive word database. Its technical approach essentially falls under the category of content review and cannot perceive or handle speech anomalies at the broadcast quality level, such as non-content-related broadcast accidents like repeated slips of the tongue, abnormal pauses, and sudden volume changes. These broadcast accidents are precisely the key factors affecting the professional quality of news programs. Second, the processing result of this solution is only to generate a record to be reviewed for subsequent manual review, which is a post-processing mechanism. It does not have the ability to intervene in real time during broadcast and assist the announcer in making corrections. This means that even if the system detects an anomaly, the announcer cannot receive any feedback during the broadcast. Third, this solution does not involve interactive linkage with broadcast auxiliary equipment. It cannot provide immediate feedback to the announcer through broadcast equipment such as in-ear monitors or teleprompters at the first moment an anomaly occurs, nor does it have the ability to perform real-time audio processing of the voice signal to correct problems such as sudden volume changes.

[0005] In traditional news live broadcasts, the detection and handling of broadcast anomalies primarily rely on manual monitoring by the director. The director judges the announcer's broadcast status by listening to the audio signal and, upon detecting an anomaly, issues a voice prompt to the announcer via in-ear monitors. However, this manual monitoring method has significant limitations. Firstly, during long, continuous live broadcasts, the director's attention gradually diminishes over time, especially in news broadcasts exceeding 60 minutes or special programs featuring major events. The rate of missed detections during manual monitoring increases significantly, and related research indicates that sustained attention declines noticeably after 45 minutes. Secondly, the entire process from the director detecting an anomaly to transmitting the prompt to the announcer via in-ear monitor typically involves a 3-5 second delay. This delay often causes the announcer to miss the optimal opportunity for remediation, resulting in irreversible broadcast damage as the abnormal audio has already been transmitted to the audience. Furthermore, with multiple tasks such as simultaneously monitoring screen switching, subtitle overlay, and signal transmission, it is difficult for the director to maintain continuous focused monitoring of audio quality, further exacerbating the unreliability of manual monitoring.

[0006] In summary, existing technologies lack a complete technical solution that can detect various types of speech anomalies in news live broadcasts in real time, provide immediate assistance and correction to the broadcaster through broadcasting equipment such as in-ear monitors and teleprompters as soon as an anomaly is detected, and adaptively optimize the detection parameters based on historical live broadcast data. Summary of the Invention

[0007] To address the aforementioned problems in existing technologies, this invention provides a method and system for real-time monitoring and correction of audio anomalies in live news broadcasts. This method extracts features from the announcer's speech signal in real-time, and executes three anomaly detection modes in parallel on a unified frame processing pipeline: abnormal pause detection, speech error / deviation detection, and sudden volume change detection. Upon detecting an anomaly, it immediately provides differentiated, real-time correction assistance to the announcer through broadcast equipment such as in-ear monitors and teleprompters. Simultaneously, it records all anomaly events to form a live broadcast quality log and generates a broadcast quality analysis report after the broadcast ends. Furthermore, it optimizes the detection threshold based on the anomaly distribution characteristics, thus forming a closed-loop guarantee mechanism that links detection and correction, and adaptively adjusts the threshold.

[0008] The first aspect of the present invention provides a method for real-time monitoring and correction of audio anomalies in live news broadcasts, comprising the following steps: a real-time audio stream framing and feature extraction step, wherein the raw audio signal collected by the announcer's wireless microphone is segmented into frames with a preset frame length, and a multi-dimensional acoustic feature vector, including short-time energy, short-time zero-crossing rate, and Mel-frequency cepstral coefficients, is extracted from each frame of audio signal and stored in a circular feature buffer in chronological order; and a multi-modal anomaly mode parallel detection step, wherein abnormal pause detection, speech error deviation detection, and volume change detection are performed synchronously based on the multi-dimensional acoustic feature vector in the circular feature buffer, wherein abnormal pause detection is determined by monitoring whether the duration of silent segments exceeds a genre-adaptive threshold. The event detection mechanism includes: Speech deviation detection, which identifies speech deviation events by comparing the edit distance between the real-time transcribed text and the teleprompter script; and volume change detection, which identifies volume change events by monitoring the variance of a sliding window. The intelligent teleprompter assistance and audio correction linkage step performs differentiated correction responses based on the type of abnormal event, including pushing ear-received prompts for abnormal pauses, highlighting the deviation position on the teleprompter for speech deviation events, and triggering a dynamic compressor to smooth the volume for volume change events. The live broadcast quality log generation and feedback optimization step records all abnormal events and generates a broadcast quality analysis report, dynamically correcting genre-adaptive thresholds and energy jump thresholds based on the distribution characteristics of the abnormalities.

[0009] A second aspect of this invention provides a real-time monitoring and correction system for abnormal audio during live news broadcasts, comprising a real-time audio stream framing and feature extraction module, a multimodal abnormal mode parallel detection module, an intelligent prompting and audio correction linkage module, and a live broadcast quality log generation and feedback optimization module. The real-time audio stream framing and feature extraction module is used to complete the framing of the audio signal and the extraction and caching of multi-dimensional acoustic feature vectors; the multimodal abnormal mode parallel detection module is used to simultaneously perform the detection of three abnormal modes and output abnormal events; the intelligent prompting and audio correction linkage module is used to drive the earphone channel, prompter display, and dynamic compressor to perform differentiated corrections according to the type of abnormal event; and the live broadcast quality log generation and feedback optimization module is used to complete the recording and archiving of abnormal events, the generation of quality reports, and the feedback correction of detection thresholds. Each module corresponds one-to-one with each step in the above method.

[0010] The beneficial effects of this invention are as follows: By deploying three heterogeneous detection methods—voice activity detection, real-time speech recognition and editing distance comparison, and short-time energy variance monitoring—in parallel within the same speech frame processing pipeline, millisecond-level real-time perception of three types of broadcasting anomalies—abnormal pauses, slips of the tongue, and sudden volume changes—is achieved. The detection delay for abnormal pauses can be controlled within a single frame processing cycle, while the detection delay for slips of the tongue depends on the response time of the speech recognition engine and is typically between 200 and 500 ms. By linking the detection results with the earphone channel and teleprompter display, immediate auxiliary feedback is provided to the announcer within hundreds of milliseconds after an anomaly occurs, greatly shortening the 3-5 second response delay in traditional manual monitoring methods. Through the feedback loop of live broadcast quality logs and broadcast quality analysis reports, adaptive correction of the detection threshold is achieved, enabling the system to be personalized and optimized according to the language habits of different announcers and the broadcast characteristics of different programs, continuously improving detection accuracy and correction effects during continuous use. Attached Figure Description

[0011] Figure 1 This is a flowchart of a method for real-time monitoring and correction of abnormal audio in live news broadcasts provided in an embodiment of the present invention.

[0012] Figure 2 This is an architecture diagram of the real-time monitoring and correction system for abnormal audio during live news broadcasts provided in this embodiment of the invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the present invention and not for limiting the present invention. It should be noted that when an element is considered to be connected to another element, it can be directly connected to the other element or there may be an intervening element present.

[0014] See Figure 1 The real-time monitoring and correction method for abnormal audio in live news broadcasts provided in this embodiment of the invention includes the following steps S1 to S4. In this embodiment, the method runs on an audio processing server in a live news broadcast studio. The server establishes data connections with the wireless microphone system, in-ear monitor transmission system, and teleprompter control host worn by the announcer through an audio interface card.

[0015] Step S1: Real-time framing and feature extraction of the audio stream. In this embodiment, the wireless microphone worn by the announcer transmits the collected raw audio signal to the audio receiver via an RF link. After analog-to-digital conversion, it enters the audio processing server as a digital audio stream. In one embodiment of the present invention, the sampling rate of the raw audio signal is set to 16kHz, and the quantization precision is 16bit. This sampling configuration can fully cover the fundamental frequency range of human voice (typically 80~300Hz) and the main harmonic components, while maintaining a low data transmission bandwidth to meet the latency requirements of real-time processing.

[0016] Preferably, when performing framing processing on the digital audio stream, the frame length is set to 20ms and the frame shift is set to 10ms, meaning there is a 50% overlap between adjacent frames. At a 16kHz sampling rate, each frame contains 320 sampling points, and the frame shift corresponds to 160 sampling points. The technical basis for choosing a 20ms frame length is that this duration is within the effective range of the short-time stationarity assumption of the speech signal (typically 10~30ms), providing sufficient temporal resolution to capture rapidly changing speech events while ensuring the accuracy of feature extraction. Setting the frame shift to 10ms ensures that, in the worst case, the detection delay of abnormal events does not exceed one frame processing cycle.

[0017] For each frame of speech signal, the system extracts a multi-dimensional acoustic feature vector in three categories. The first category of features is short-time energy. The calculation formula is as follows:

[0018] ,

[0019] in, For the first The short-time energy of a frame is expressed as the square of the digital amplitude (i.e., the dimensionless normalized energy value). In this embodiment, the number of sampling points corresponds to the frame length. ; For the first The first frame The time-domain amplitude value of each sampling point; In this embodiment, the number of sampling points corresponding to the frame shift is... ; The Hamming window function is expressed as follows: The purpose of using a Hamming window is to reduce spectral leakage at frame boundaries. Short-time energy features primarily serve the subsequent abnormal pause detection and volume change detection sub-modules.

[0020] The second type of feature is the short-time zero-crossing rate. , defined as the number of times the signal crosses zero level in each frame, is calculated using the following formula:

[0021] ,

[0022] in, For the first The short-time zero-crossing rate of the frame, with a value range of 100%. Dimensionless; This is a sign function that returns 1 when the input is greater than 0, 0 when it is equal to 0, and 0 when it is less than 0. Short-time zero-crossing rate can help distinguish between audible and silent segments under low signal-to-noise ratio conditions, forming a complementary criterion with short-time energy.

[0023] The third type of feature is the 13-dimensional Mel-frequency cepstral coefficient (MFCC) vector. Its extraction process is as follows: a 512-point Fast Fourier Transform is performed on the windowed speech frame to obtain the spectrum; the spectrum is then weighted and summed using a Mel filter bank consisting of 26 triangular filters to obtain the Mel spectrum; the logarithm of the Mel spectrum is taken, and a Discrete Cosine Transform is performed, retaining the first 13 coefficients as the MFCC feature vector. The MFCC feature vector primarily serves the real-time speech recognition engine in the slip-of-speech detection submodule, providing input features for the acoustic model.

[0024] The three types of features mentioned above are merged to form a multi-dimensional acoustic feature vector for each frame, which is then written into a circular feature buffer in chronological order. In one embodiment of the present invention, the capacity of the circular feature buffer is set to the feature vectors of the most recent 300 consecutive frames, with a 10ms frame shift corresponding to a 3s speech duration. When the buffer is full, the latest written feature vector will overwrite the oldest feature vector, thus ensuring that the buffer always retains the acoustic feature data of the most recent 3s. The buffer capacity is set based on the fact that a 3s time window can cover the maximum genre adaptive threshold required for abnormal pause detection (3.0s for talk shows), while providing sufficient historical data support for sliding window variance monitoring.

[0025] Step S2: Parallel Detection of Multimodal Anomalies. During the continuous updating of the circular feature buffer, the system simultaneously launches three independent detection sub-modules to perform parallel detection of three anomaly modes: abnormal pauses, verbal slips, and sudden volume changes. The three detection sub-modules share the data in the circular feature buffer but maintain independent detection state variables, thus achieving simultaneous monitoring of multiple anomaly modes without interference.

[0026] Step S2a: Abnormal pause detection. The abnormal pause detection submodule determines whether the announcer has made a speech interruption that exceeds the normal range by monitoring the time evolution pattern of the short-time energy sequence. In one embodiment of the present invention, the short-time energy of the current frame is first analyzed. Compared with the preset sound detection threshold When comparing, The current frame is marked as a frame with sound if it is active, otherwise it is marked as a frame without sound. Sound detection threshold. The value is typically set to 6 times the average environmental noise floor energy collected within 30 seconds before the start of the live broadcast. In one embodiment of the present invention, in a standard studio environment... The typical range of values ​​is to (Normalized energy value).

[0027] Based on frame-by-frame determination of the presence or absence of sound, the system maintains a silent segment duration counter. The counter increments by 1 when the current frame is marked as a silent frame and resets to 0 when the current frame is marked as a frame with sound. The value multiplied by the frame shift duration The duration of the silent segment obtained (in this embodiment, it is 10ms) Exceeding the genre adaptation threshold At that moment, the system determined that an abnormal pause event had occurred.

[0028] Genre-adaptive threshold The settings are associated with the genre of the current live broadcast program, with different genres corresponding to different allowable silence durations. In this embodiment, the silence duration for news broadcast programs is... The setting is 1.5 seconds because news broadcasts typically require a tight pace and continuous information output; silences exceeding 1.5 seconds are clearly abnormal in a news broadcast context. Talk shows, on the other hand, have a similar setting. The timeout is set to 3.0 seconds to account for the natural pauses and turn-taking between the host and guests in talk shows; for commentary programs... Set to 2.5 seconds, this falls between a news broadcast and a talk show. The program format is preset by the director through the system configuration interface before the live broadcast begins.

[0029] To avoid false alarms triggered by natural micro-pauses between words (typically lasting 50-200ms), the system introduces a de-jitter mechanism before the silence duration counter starts: only when the number of consecutive silent frames exceeds the minimum de-jitter frame count... (In this embodiment, it is set to 30 frames, corresponding to 300ms) After that, the silent segment duration counter begins to start timing. This de-jitter mechanism can effectively filter out normal breathing pauses and sentence breaks during the broadcast process, ensuring that the alarm is triggered only for truly abnormal long periods of silence.

[0030] Step S2b: Speech Deviation Detection. The core idea of ​​the speech deviation detection submodule is to continuously compare the announcer's real-time spoken content with the text displayed on the teleprompter, thereby quickly detecting instances where the announcer deviates from the intended text. This submodule consists of two interconnected stages: real-time transcription and text comparison.

[0031] During the real-time transcription stage, the system feeds the MFCC feature vectors from the circular feature buffer into the streaming speech recognition engine in real time. In one embodiment of the invention, the engine employs a streaming Transformer architecture based on an end-to-end attention mechanism, operating in a block processing manner. Each processing block covers approximately 640ms of speech data (corresponding to 64 frames), with processing latency controlled between 200 and 500ms. The transcribed text output by the engine is passed to the text comparison stage in a character-by-character incremental update manner; that is, each newly recognized character is appended to the transcribed text buffer. Preferably, the speech recognition engine is domain-adaptive fine-tuned for news broadcasting scenarios during deployment, and its character error rate on the standard news corpus test set is controlled within 5%. This level of accuracy ensures that the text comparison stage obtains sufficiently reliable transcribed input.

[0032] During the text comparison stage, the system needs to compare the real-time generated transcribed text with the manuscript text currently displayed by the teleprompter. In one embodiment of the present invention, the teleprompter control host synchronously transmits the manuscript text of the currently displayed paragraph in string form to the audio processing server via a network interface. The system uses a dynamic programming algorithm to calculate the transcribed text (as the source sequence). ) and the corresponding paragraphs of the manuscript text (as the target sequence) Edit distance matrix between ) Its state transition equation is:

[0033] ,

[0034] in, Indicates the source sequence before The characters and the target sequence before The minimum edit distance between characters, a dimensionless integer value; The first of the source sequence (transcribed text) One character; For the target sequence (manuscript text) of the first One character; To replace the cost function, when The value is 0 if the condition is met, and 1 otherwise. The three terms in the branch correspond to the three basic editing operations: character replacement, deletion, and insertion. Each operation has a unit cost of 1. The initial condition of the matrix is... , .

[0035] After obtaining the complete edit distance matrix, the system starts from... The optimal alignment path is traced back, and mismatched segments in the path where replacement or insertion operations occur consecutively are marked. In one embodiment of the present invention, the preset deviation tolerance number is set to 3, that is, when the number of consecutive mismatched characters in a certain mismatched segment reaches a certain threshold, the deviation tolerance is set to 3. When this occurs, the system determines that the segment is a speech error deviation event. The technical basis for choosing 3 characters as the deviation tolerance number is that: a single character mismatch is very likely to be a recognition error of the speech recognition engine rather than an actual slip of the tongue by the announcer; a 2-character mismatch also has a high probability of recognition error in fast speech flow; while a mismatch of 3 or more consecutive characters is highly likely to reflect the announcer's actual deviation behavior.

[0036] Preferably, considering that the real-time speech recognition engine may correct the output characters during the incremental output process (i.e., the so-called backtracking correction phenomenon), the system sets a delayed confirmation window for the two most recently output characters during comparison, and only performs edit distance comparison on the transcription results before the delayed confirmation window, so as to avoid false alarms caused by the intermediate output of the recognition engine.

[0037] Step S2c: Volume Sudden Change Detection. The volume sudden change detection submodule aims to capture discontinuous volume jumps during broadcasting caused by sudden changes in microphone distance, device gain jumps, or broadcaster emotional fluctuations. Unlike abnormal pause detection, which focuses on the absolute level of energy, volume sudden change detection focuses on the rapid changes in energy over time.

[0038] In one embodiment of the present invention, the system uses a preset sliding window length (Set to 50 frames, corresponding to a time span of 500ms) A sliding analysis window is constructed for the short-time energy sequence in the circular feature buffer. At each analysis time, the system performs an arithmetic mean operation on the short-time energy of all frames within the sliding analysis window to obtain the short-time energy mean. The variance of the short-time energy is calculated by performing variance operation on the short-time energy of all frames within the sliding analysis window to obtain the short-time energy variance. The calculation formulas are as follows:

[0039] ,

[0040] ,

[0041] in, The average energy over a short period within the sliding window, its dimensions are... Consistent; Let Variance be the short-time energy variance within the sliding window, with dimensions of The square of; The length of the sliding window is specified in this embodiment. (frame); The current frame number; For the first The short-time energy of a frame.

[0042] Subsequently, the system calculates the inter-frame energy difference of the current frame. This value, expressed in decibels, represents the degree of deviation of the current frame's short-time energy from the window mean.

[0043] ,

[0044] in, This represents the inter-frame energy difference of the current frame, in dB. To prevent the extremely small constant from being divided by zero, in this embodiment... Its function is to prevent numerical anomalies where the denominator is zero under completely silent conditions.

[0045] when Exceeding the preset energy jump threshold (Set to 12dB in this embodiment) and short-time energy variance Exceeding the preset variance threshold At this time, the system determines that the current frame is the frame triggered by a sudden volume change event. A variance condition is introduced. The purpose is to eliminate the following false alarm scenarios: during the brief silence when the announcer takes a normal breath, the energy transitions from the audible segment to the silent segment. It may exceed 12 dB, but in this case, the energy variance within the window is small, indicating that the energy change is a smooth transition rather than an abrupt change. Preferably, The typical range of values ​​is to (Normalized energy variance), the specific value of which is determined during the system calibration phase based on the studio acoustic environment. The technical basis for choosing 12dB as the energy jump threshold is that in a professional studio environment, the inter-frame energy fluctuation caused by changes in the announcer's normal speaking rate is usually in the range of 3~6dB, while the energy jump caused by sudden changes in microphone distance or equipment gain is usually more than 10dB. The 12dB threshold provides sufficient decision margin between the two.

[0046] Step S3: Intelligent prompting assistance and audio correction linkage step. The abnormal events output by the three parallel detection submodules in Step S2 are transmitted to Step S3 through a unified event message queue. The intelligent prompting assistance and audio correction linkage step then executes differentiated correction responses based on the type of abnormal event. A significant technical feature of this invention is that different correction methods and feedback channels are matched to different types of abnormal events, ensuring that each correction response accurately adapts to the characteristics of the corresponding abnormality, thereby achieving the highest correction efficiency with minimal interference during broadcasting.

[0047] In the event of an abnormal pause, the system pushes a preset prompt signal to the announcer via the in-ear monitoring channel. In one embodiment of the invention, the prompt signal is a 1kHz sinusoidal wave tone lasting 200ms. After fade-in and fade-out processing, the peak level is set 15dB below the announcer's in-ear monitoring level to ensure that the prompt does not interfere with the announcer's normal listening to their own voice, while still being sufficiently perceptible. The technical basis for choosing a frequency of 1kHz for the prompt is that this frequency is within the most sensitive frequency range for the human ear (usually 1~4kHz), and can be clearly perceived at a low level. Preferably, the system completes the generation and transmission of the prompt signal within 20ms after detecting an abnormal pause, thereby controlling the end-to-end delay of the in-ear monitoring prompt to within 30ms, far lower than the 3~5s response delay of traditional manual broadcasting methods. In addition, to avoid repeated triggering of the prompt during the same continuous silence, causing anxiety to the announcer, the system sets a prompt cooling-off time, and will not trigger repeated prompts for the same continuous silence segment within 5s after an abnormal pause triggers the prompt.

[0048] In the event of a slip of the tongue, the system sends a deviation marking instruction to the teleprompter control host, highlighting the deviation position on the teleprompter screen. In one embodiment of the invention, the highlighting is implemented by displaying the text at the deviation position with a red background, distinguishing it from the white background of the normal text. Simultaneously, a green triangle indicating a return to the original text is displayed at the beginning of the most recent complete sentence preceding the deviation position. The logic for selecting the position of the return to the original text is as follows: searching backward from the deviation position, the first character after the nearest period, question mark, or exclamation mark is used as the suggested return to the original text starting point. This design is based on the fact that after a slip of the tongue, the announcer needs to restart from the beginning of a semantically complete sentence to ensure the semantic continuity of the information heard by the audience. Preferably, the highlighting on the teleprompter screen is displayed continuously for 8 seconds, after which it automatically returns to normal display to avoid excessive accumulation of highlighting affecting the announcer's normal reading.

[0049] In the event of a sudden volume change, the system automatically triggers the dynamic compressor in the audio processing chain to smooth the volume back to a normal range. In one embodiment of the invention, the dynamic compressor employs a soft inflection point characteristic, with a compression ratio set to 4:1. This means that when the input signal exceeds the soft inflection point threshold, every 4dB increase in input signal only results in a 1dB increase in output signal. The soft inflection point threshold is set to the average energy of the current window. The corresponding decibel value is increased by 6dB, meaning compression begins 6dB above the normal broadcast level. The dynamic compressor's attack time is set to 5ms, and the release time to 50ms. The technical rationale for setting the attack time to 5ms is that it needs to be short enough to quickly intervene in compression after a sudden change in volume, while not too short to avoid unnecessary attenuation of the normal onset transient of the speech signal. The purpose of setting the release time to 50ms is to smoothly exit the compression state after the volume returns to normal, avoiding a breathing effect (i.e., background noise jerking caused by rapid compressor release). Through the processing of the dynamic compressor, the level fluctuation range of the audio signal received by the audience can be compressed from more than 15dB before processing to less than 6dB, significantly improving the listening experience.

[0050] In one embodiment of the present invention, the priority relationship of the above three correction responses is set as follows: the correction priority for volume change events is the highest, because volume change directly affects the auditory experience of the audience and needs to be processed in real time at the audio signal level; the correction priority for slips of the tongue events is the second highest, because slips of the tongue require the announcer to actively cooperate in rereading and correcting them; the correction priority for abnormal pause events is the lowest, because a moderate prompt can guide the announcer to resume broadcasting. When multiple abnormal events occur concurrently at the same time, the system processes them in order of priority, with higher-priority correction responses being executed before lower-priority ones.

[0051] Step S4: Live Stream Quality Log Generation and Feedback Optimization. Throughout the live stream, all abnormal events and their corrective responses generated in Steps S2 and S3 are written to the live stream quality log in the form of structured records. Each record contains the following fields: event timestamp (accurate to milliseconds), abnormal type identifier (values ​​are one of abnormal pause, slip of the tongue, or sudden volume change), abnormal severity score, type of triggering corrective response, and execution result of the corrective response. The abnormal severity score is automatically calculated based on the quantitative indicators of abnormal feature parameters. Specifically, the severity score for an abnormal pause event is proportional to the duration of the silent segment, the severity score for a slip of the tongue event is proportional to the number of consecutive mismatched words, and the severity score for a sudden volume change event is proportional to the absolute value of the inter-frame energy difference.

[0052] In one embodiment of the present invention, the system maintains an independent live broadcast quality log file for each live broadcast. The log file is stored in JSON format in the local storage of the audio processing server, and the filename includes the live broadcast date, program name, and announcer identification information for easy retrieval later. During the live broadcast, the log writing operation is performed asynchronously and non-blockingly to ensure that log recording does not affect the processing delay of the real-time detection and correction pipeline.

[0053] After the live broadcast ends, the system automatically reads the quality log of the live broadcast and generates a broadcast quality analysis report. In one embodiment of the present invention, the broadcast quality analysis report includes the following analysis dimensions: the total number of abnormal events categorized by anomaly type, a histogram of the time distribution of abnormal events divided into 5-minute time granularities, the average duration and maximum duration of each type of abnormal event categorized by anomaly type, and a comprehensive broadcast quality score. The formula for calculating the overall broadcast quality score is as follows:

[0054] ,

[0055] in, The overall broadcast quality score is calculated based on the following values: Dimensionless; This represents the total number of abnormal events during the entire live broadcast. For the first The weighting coefficients for the types of abnormal events are set as follows: in this embodiment, the weight for abnormal pauses is set to 1.0, the weight for verbal slips and deviations is set to 2.0, and the weight for sudden changes in volume is set to 1.5, reflecting the degree of impact of different abnormal types on broadcast quality. For the first The severity score for each abnormal event, with a value range of [value missing]. ; For the first The duration of each abnormal event, in seconds; The total duration of the live broadcast is expressed in seconds. Slips of the tongue have the highest weight because they directly lead to errors in information transmission, having the greatest impact on the accuracy of news reports.

[0056] A core function of the broadcast quality analysis report is to provide data support for the feedback optimization of the system's detection thresholds. In one embodiment of the present invention, the system adaptively adjusts the genre threshold based on the statistical data of abnormal events in the broadcast quality analysis report. and energy transition threshold Perform dynamic correction.

[0057] Genre-adaptive threshold The system corrects and counts the number of false alarms for abnormal pauses during the live broadcast. and number of missed reports The method for determining false alarms is as follows: After the live broadcast, the director or announcer manually reviews and annotates the abnormal pause events in the log, marking events that are actually normal intervals but were mistakenly identified as abnormal pauses by the system as false alarms. The method for determining missed alarms is as follows: During the live broadcast, the director records obvious abnormal pauses that they believe the system failed to detect through a separate manual annotation interface, and cross-compares them with the system log after the broadcast to determine missed events. When the false alarm rate is... (in The total number of abnormal pause events exceeds the preset false alarm ratio threshold. (In this embodiment, it is set to 0.2, i.e., 20%), when the system will Increase the threshold adjustment step by one preset step. (In this embodiment) s); when the false negative rate (in The expected number of detections (determined by the sum of the number of annotations by the director and the number detected by the system) exceeds a preset false negative threshold. (When set to 0.15, i.e., 15%, in this embodiment, the system will...) Lower This closed-loop correction mechanism allows the system to gradually approach the optimal detection threshold based on the actual performance of each live broadcast.

[0058] Regarding the energy jump threshold For correction, the system employs a similar false alarm-false alarm analysis mechanism. When the false alarm rate for a sudden volume change event exceeds 20%, the system will... Increase by 1dB; when the false negative rate exceeds 15%, the system will... Reduce by 1dB. Preferably, to prevent excessive oscillation of the threshold during multiple corrections, the system sets upper and lower limits for the cumulative correction amplitude of the threshold: The allowable correction range is the initial value. , The allowable correction range is the initial value. .

[0059] Through the coordinated operation of steps S1 to S4, this invention achieves a complete closed loop from speech signal acquisition, multimodal anomaly detection, real-time correction assistance to quality log feedback optimization. Step S1 provides a unified multidimensional acoustic feature input for step S2, the anomaly event output of step S2 drives step S3 to execute a differentiated correction response, the operation records of steps S3 and S2 are incorporated into step S4 to form a quality log and analysis report, and the threshold correction result of step S4 is fed back to step S2 to optimize the accuracy of subsequent detection, thus forming a deeply coupled closed-loop architecture of feedforward detection-correction and feedback optimization.

[0060] See Figure 2 The news live broadcast audio anomaly real-time monitoring and correction system provided in this embodiment of the invention includes a real-time audio stream framing and feature extraction module 1, a multimodal anomaly mode parallel detection module 2, an intelligent word prompting and audio correction linkage module 3, and a live broadcast quality log generation and feedback optimization module 4. The above four modules are deployed on the studio audio processing server and realize low-latency data interaction between modules through an internal data bus.

[0061] The real-time framing and feature extraction module 1 for the audio stream corresponds to step S1 in the above method embodiment, and is used to complete the entire processing flow from the acquisition of the original audio signal to the extraction and caching of multi-dimensional acoustic feature vectors. In one embodiment of the present invention, the hardware foundation of this module includes a professional audio interface card supporting ASIO driver and a real-time digital signal processing unit. The audio interface card receives the audio signal output from the announcer's wireless microphone receiver through the analog line input port, and performs analog-to-digital conversion at a sampling rate of 16kHz and a quantization precision of 16bit. After receiving the digital audio stream, the real-time digital signal processing unit performs framing processing with a frame length of 20ms and a frame shift of 10ms, extracting short-time energy, short-time zero-crossing rate, and 13-dimensional MFCC features frame by frame, and writes the extracted multi-dimensional acoustic feature vectors into a circular feature buffer with a capacity of 300 frames according to the first-in-first-out principle. Preferably, the frame processing delay of this module is controlled within 2ms. This performance indicator ensures that the subsequent detection module can acquire the latest acoustic feature data at a near real-time speed. This module also includes a signal preprocessing subunit for performing pre-emphasis filtering (pre-emphasis coefficient of 0.97) and DC offset cancellation on the raw audio signal to improve the accuracy of subsequent feature extraction. The transfer function of the pre-emphasis filter is... ,in The unit delay operator, with a coefficient of 0.97, results in an approximately 6 dB / oct boost to the high-frequency components, compensating for the inherent attenuation of high-frequency radiation during vocalization.

[0062] The multimodal anomaly detection module 2 corresponds to step S2 in the above method embodiment. It is used to synchronously perform parallel detection of three anomaly modes—abnormal pause detection, speech deviation detection, and volume change detection—on a unified frame processing pipeline. In one embodiment of the present invention, this module is implemented using a multi-threaded parallel architecture, wherein abnormal pause detection, speech deviation detection, and volume change detection run in independent processing threads. Each thread shares access to the data in the circular feature buffer through a read-lock mechanism without blocking each other.

[0063] The abnormal pause detection submodule 2a reads the short-time energy value of the latest frame from the circular feature buffer, compares it with the sound determination threshold to determine whether the current frame is sound or silent, and maintains a silent segment duration counter to monitor the duration of the silent segment. When the duration of the silent segment exceeds a preset genre-adaptive threshold based on the current program genre, this submodule generates an abnormal pause event message and pushes it to the event message queue. As described in the method embodiment, the genre-adaptive threshold for news broadcast programs is 1.5s, for talk shows it is 3.0s, and for commentary programs it is 2.5s. The specific threshold is set by the director according to the program type through the system configuration interface before the live broadcast begins. This submodule also integrates a de-jitter mechanism, which only starts the formal timing after the number of consecutive silent frames exceeds 30 frames (corresponding to 300ms), thereby effectively filtering out brief silences caused by normal breathing and sentence breaks.

[0064] The slip-of-speech deviation detection submodule 2b functionally comprises two cascaded components: a real-time speech recognition unit and a text comparison unit. The real-time speech recognition unit reads MFCC feature vectors from a circular feature buffer and feeds them into a streaming Transformer speech recognition engine for incremental, character-by-character transcription. The transcription result is output to the text comparison unit as a character stream. Simultaneously, the text comparison unit obtains the manuscript text of the currently displayed paragraph from the teleprompter control host, calculates the edit distance matrix between the transcribed text and the manuscript text using a dynamic programming algorithm, and identifies consecutive mismatched segments by backtracking the optimal alignment path. When the number of consecutive mismatched characters in a mismatched segment reaches or exceeds the deviation tolerance limit (3 characters in this embodiment), this submodule generates a slip-of-speech deviation event message. Preferably, the text comparison unit employs an incremental update strategy to optimize computational efficiency: when the speech recognition engine outputs new transcribed characters, the text comparison unit only needs to update the latest row or column of the edit distance matrix, without recalculating the entire matrix, thereby reducing the computational complexity of a single comparison from... Reduce to ,in To transcribe the text length, This refers to the length of the manuscript text. This optimization is crucial for real-time processing performance in long-duration live streaming scenarios.

[0065] The volume change detection submodule 2c reads the short-time energy sequence of the most recent 50 frames from the circular feature buffer, calculates the energy mean and energy variance within the sliding window, and compares the inter-frame energy difference of the current frame with a preset energy jump threshold (12dB). When the inter-frame energy difference exceeds the threshold and the energy variance exceeds the variance threshold, this submodule generates a volume change event message. As detailed in the method embodiment, this submodule's design with dual decision conditions (energy difference exceeding the threshold and variance exceeding the threshold) effectively eliminates false alarms caused by energy fluctuations due to normal speech rhythm changes.

[0066] The intelligent prompting and audio correction linkage module 3 corresponds to step S3 in the above method embodiment. It is used to drive different broadcast devices to execute differentiated correction responses based on the abnormal event types output by the multimodal abnormality mode parallel detection module 2. In one embodiment of the present invention, this module internally maintains an event processing queue sorted by priority. Volume surge events have the highest priority (priority value 1), followed by slips of the tongue events (priority value 2), and abnormal pause events have the lowest priority (priority value 3). When multiple abnormal events occur concurrently within the same processing cycle, the module executes the corresponding correction responses sequentially according to their priority.

[0067] This module comprises three correction response subunits. Upon receiving an abnormal pause event, the in-ear monitoring prompt subunit 3a sends a 1kHz, 200ms prompt signal to the announcer's in-ear monitor receiver via a digital-to-analog converter and the in-ear monitoring transmission link. The prompt level is 15dB lower than the announcer's in-ear monitoring level, and a 5-second cooldown period is provided to prevent repeated triggering. Upon receiving a speech error deviation event, the teleprompter annotation subunit 3b sends an annotation command containing the deviation coordinates and the readback start point position to the teleprompter control host via a network interface. Upon receiving the command, the teleprompter screen highlights the deviation area with a red background and displays a green triangle marker at the readback start point. The highlighting lasts for 8 seconds and then automatically returns to normal. Upon receiving a volume surge event, the dynamic compression processing subunit 3c activates a dynamic compressor at the output stage of the audio processing link. It compresses signals exceeding the soft inflection threshold with a 4:1 compression ratio, a 5ms attack time, and a 50ms release time, ensuring that the audio signal level fluctuation received by the audience does not exceed 6dB.

[0068] The live broadcast quality log generation and feedback optimization module 4 corresponds to step S4 in the above method embodiment, and is used to complete the entire post-processing process from archiving abnormal event records to generating quality reports and then to feedback correction of detection thresholds. In one embodiment of the present invention, this module asynchronously writes each abnormal event record (including event timestamp, abnormality type, severity score, and correction response execution result) to a local live broadcast quality log file in JSON format during the live broadcast. After the live broadcast ends, the module automatically parses the quality log and generates statistics including abnormal event classification, time distribution histogram, duration statistics, and comprehensive broadcast quality score. The broadcast quality analysis report includes [the report].

[0069] After receiving manual review annotations from the director or announcer, the feedback optimization subunit of this module calculates the false alarm rate and false negative rate of abnormal pause events and volume change events, and applies an adaptive threshold to the genre according to the threshold correction rules described in the method embodiment. and energy transition threshold Dynamic correction is performed. The corrected threshold parameters are updated in real time to the corresponding detection submodule in the multimodal anomaly pattern parallel detection module 2 via the inter-module data bus, so that the optimized detection parameters can be used in subsequent live broadcasts. Preferably, the system also supports cross-session aggregation analysis of quality logs from multiple live broadcasts to identify systematic anomaly patterns under specific announcers or specific program types, providing data support for long-term broadcast quality improvement.

[0070] The data flow relationships among the above four functional modules form a deeply coupled closed-loop collaborative architecture: the feature output of the real-time audio stream framing and feature extraction module 1 is the sole data input source for the multimodal anomaly pattern parallel detection module 2, and the two are coupled through a circular feature buffer; the event output of the multimodal anomaly pattern parallel detection module 2 drives the intelligent prompting and audio correction linkage module 3 to perform correction actions, and the two are coupled through an event message queue; the operation records of the multimodal anomaly pattern parallel detection module 2 and the intelligent prompting and audio correction linkage module 3 converge to the live broadcast quality log generation and feedback optimization module 4 to form quality data accumulation, and the threshold correction results of the live broadcast quality log generation and feedback optimization module 4 in turn update the detection parameters of the multimodal anomaly pattern parallel detection module 2, forming a complete feedback closed loop. This closed-loop architecture enables the system's detection accuracy to continuously improve with the increase of usage, achieving the technical effect of adaptive evolution.

[0071] To verify the actual performance of the real-time monitoring and correction method for news live broadcast audio anomalies proposed in this invention, a 30-day test and evaluation was conducted in the standard high-definition studio of a provincial television station's news channel. The test covered various program types, including daily news broadcasts, special reports, and live broadcasts of breaking news, with a total test broadcast duration exceeding 180 hours. Eight announcers participated in the test.

[0072] The hardware configuration of the test environment is as follows: The audio processing server is a rack-mount server equipped with an Intel Xeon E5-2680v4 processor (14 cores 2.4GHz) and 64GB DDR4 memory, running Ubuntu 20.04 LTS; the audio interface card is an RME Fireface UCX II, supporting a maximum sampling rate of 192kHz and ASIO low-latency driver; the announcer's wireless microphone system uses a Sennheiser EW-DX series digital wireless system, with the receiver connected to the audio interface card via an XLR analog line output. The teleprompter system is an Autoscript intelligent teleprompter, synchronizing the teleprompter text with the audio processing server via a TCP / IP network. The in-ear monitoring system uses a Shure PSM300 wireless in-ear monitor, with the transmitter receiving prompt signals through the analog output port of the audio interface card.

[0073] In terms of abnormal pause detection performance, the system achieved an accuracy rate of 94.2% with a false positive rate of 4.8% and a false negative rate of 1.0% in news broadcast programs (genre-adaptive threshold 1.5s). In talk shows (genre-adaptive threshold 3.0s), the accuracy rate was 91.7% with a false positive rate of 6.3% and a false negative rate of 2.0%. In comparison, the traditional manual monitoring method under the same test conditions achieved an abnormal pause detection rate of only 72.5%, with an average response delay of 4.2s. The abnormal pause detection delay of the system in this invention is an average of 15ms (the end-to-end delay from the moment the abnormal pause duration exceeds the threshold to the arrival of the earpiece alert signal in the announcer's headset), which is 99.6% shorter than the traditional manual method. After feedback correction in 5 live broadcasts, the detection accuracy rate for news broadcast programs increased from the initial 94.2% to 96.8%, verifying the effectiveness of the threshold adaptive correction mechanism.

[0074] In terms of error detection performance, with a word error rate of 4.3% for the streaming speech recognition engine, the system achieved a recall rate of 88.6% and a precision rate of 82.4% for errors of 3 or more consecutive words. The average detection latency (end-to-end latency from the occurrence of an error to the appearance of a highlighted mark on the teleprompter screen) was 380ms, of which the speech recognition engine contributed approximately 320ms, and the edit distance calculation and annotation instruction transmission combined accounted for approximately 60ms. During the 30-day test, a survey of announcers' subjective feedback showed that 85% of announcers believed that the highlighted marks and rereading indicators for errors helped them find the rereading position more quickly after an error occurred, reducing the average rereading start time from 2.8s without system assistance to 0.9s.

[0075] In terms of volume change detection performance, the system achieved an accuracy rate of 96.1% for volume jumps exceeding 12dB, with a false alarm rate of 3.1% and a false negative rate of 0.8%. After the dynamic compressor was activated, the peak level fluctuation range of the audio signal received by the audience was compressed from an average of 18.5dB before processing to 5.2dB, an improvement of 71.9%. In subjective listening evaluation, a review panel of 15 professional audio engineers scored the compressed audio quality using the Mean Opinion Score (MOS), with an average score of 4.2 out of 5, significantly better than the 2.8 score of the unprocessed audio.

[0076] In terms of overall system performance, the comprehensive broadcast quality score After adopting the system of this invention, the score improved from a baseline of 78.3 to 92.6, an improvement of 18.3%. The system's CPU utilization during single-channel audio stream processing averaged 12.7%, with a peak of no more than 25%, and memory usage was approximately 1.2GB, indicating that the system has ample computational margin on mainstream server hardware to support the expansion requirements of multi-channel parallel processing. The end-to-end latency of the frame processing pipeline (from audio sampling to completion of abnormal event determination) averaged 8.3ms, with a maximum value not exceeding 15ms, fully meeting the requirements of real-time processing.

[0077] The test results above demonstrate that this invention effectively addresses the inherent shortcomings of traditional manual monitoring methods, such as attention decay, response delay, and multitasking, by integrating three heterogeneous detection methods—voice activity detection, real-time speech recognition and edit distance comparison, and short-time energy variance monitoring—into a unified frame processing pipeline and establishing real-time linkage with broadcast equipment like in-ear monitors and teleprompters. This provides an efficient, reliable, and continuously optimizable intelligent technical solution for ensuring the quality of live news broadcasts. Notably, the feedback optimization closed-loop mechanism employed in this invention enables the system to automatically adjust threshold parameters based on the actual detection results of each live broadcast. After accumulating learning from multiple live broadcasts, the system develops differentiated parameter configurations for different announcers' individual language habits and program characteristics. This adaptive characteristic is unattainable by traditional fixed-threshold detection schemes and is the core advantage that distinguishes this invention from existing live audio monitoring technologies.

[0078] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for real-time monitoring and correction of abnormal audio in live news broadcasts, characterized in that: Includes the following steps: Real-time framing and feature extraction steps for audio stream: The raw audio signal collected by the announcer's wireless microphone is segmented into frames with a preset frame length. Multidimensional acoustic feature vectors, including short-time energy, short-time zero-crossing rate and Mel frequency cepstral coefficients, are extracted from each frame of audio signal and stored in a circular feature buffer in chronological order. The multimodal abnormal pattern parallel detection steps are as follows: Based on the multidimensional acoustic feature vector in the circular feature buffer, abnormal pause detection, speech deviation detection, and volume change detection are performed simultaneously. Abnormal pause detection determines whether a segment is audible or silent based on short-time energy and monitors the duration of the silent segment. When the duration of the silent segment exceeds the genre-adaptive threshold, it is determined to be an abnormal pause event. Speech deviation detection uses a real-time speech recognition engine to transcribe the speech signal into transcribed text and compares the edit distance with the teleprompter's text. When the number of consecutive mismatched words exceeds the deviation tolerance limit, it is determined to be a speech deviation event. Volume change detection performs sliding window variance monitoring on the short-time energy sequence. When the energy difference between adjacent frames exceeds the energy jump threshold, it is determined to be a volume change event. The intelligent prompting assistance and audio correction linkage steps are as follows: Differentiated correction responses are executed according to the type of abnormal event. For abnormal pause events, a prompt signal is pushed to the announcer through the ear feedback channel. For mispronunciation and deviation events, the deviation position is highlighted on the prompter screen and the starting point of the rereading is indicated. For sudden volume events, the dynamic compressor in the audio processing link is triggered to smooth the volume to the target range with a soft inflection compression ratio. Live Streaming Quality Log Generation and Feedback Optimization Steps: Write the timestamp, anomaly type, and correction response record of each abnormal event during the entire live stream into the live streaming quality log. After the live stream ends, generate a broadcast quality analysis report containing the frequency and time distribution of various anomalies, and dynamically correct the genre adaptive threshold and energy jump threshold based on the anomaly distribution characteristics.

2. The method for real-time monitoring and correction of abnormal audio in live news broadcasts according to claim 1, characterized in that, In the real-time framing and feature extraction steps of the audio stream, the preset frame length is 20ms, the frame shift is 10ms, and the sampling rate is 16kHz. The capacity of the circular feature buffer is the multidimensional acoustic feature vector of the most recent 300 consecutive frames.

3. The method for real-time monitoring and correction of abnormal audio in live news broadcasts according to claim 1, characterized in that, In abnormal pause detection, the genre-adaptive threshold is determined based on the current program genre. The genre-adaptive threshold for news broadcasts is 1.5s, for talk shows it is 3.0s, and for commentary shows it is 2.5s.

4. The method for real-time monitoring and correction of abnormal audio in live news broadcasts according to claim 1, characterized in that, In volume change detection, the preset energy jump threshold is 12dB; the soft inflection compression ratio of the dynamic compressor is 4:1, the attack time is 5ms, and the release time is 50ms.

5. The method for real-time monitoring and correction of abnormal audio in live news broadcasts according to claim 1, characterized in that, The word-by-word edit distance comparison in slip-of-speech deviation detection includes the following sub-steps: Using the transcribed text output by the real-time speech recognition engine as the source sequence and the paragraph corresponding to the manuscript text currently displayed by the teleprompter as the target sequence, a dynamic programming algorithm is used to calculate the word-by-word edit distance matrix between the source sequence and the target sequence. Backtrack the optimal alignment path in the edit distance matrix and mark the mismatched segments where replacement or insertion operations occur consecutively; When the number of consecutive mismatched characters in a mismatched section exceeds the preset deviation tolerance number, the starting position of the mismatched section is marked as the deviation starting point and determined as a slip of the tongue deviation event.

6. The method for real-time monitoring and correction of abnormal audio in live news broadcasts according to claim 1, characterized in that, Sliding window variance monitoring in volume change detection includes the following sub-steps: A sliding analysis window is constructed for the short-time energy sequence in the circular feature buffer with a preset sliding window length. An arithmetic mean operation is performed on the short-time energy of all frames in the sliding analysis window to obtain the short-time energy mean. A variance operation is performed on the short-time energy of all frames in the sliding analysis window to obtain the short-time energy variance. Calculate the inter-frame energy difference between the short-time energy of the current frame and the mean short-time energy within the sliding analysis window; When the inter-frame energy difference exceeds the preset energy jump threshold and the short-term energy variance exceeds the preset variance threshold, the current frame is determined to be a volume change event triggered frame.

7. The method for real-time monitoring and correction of abnormal audio in live news broadcasts according to claim 1, characterized in that, In the live stream quality log generation and feedback optimization steps, the process of dynamically correcting the genre adaptive threshold includes: The broadcast quality analysis report includes the number of false alarms and missed alarms for abnormal pause events. When the proportion of false alarms to the total number of abnormal pause events exceeds the preset false alarm proportion threshold, the genre adaptive threshold will be increased by the preset threshold adjustment step. When the proportion of missed detections to the expected number of detections within the total live broadcast duration exceeds the preset missed detection ratio threshold, the genre adaptive threshold will be lowered by the preset threshold adjustment step.

8. The method for real-time monitoring and correction of abnormal audio in live news broadcasts according to claim 1, characterized in that, In the intelligent prompting and audio correction linkage process, the highlighting of speech error deviation events on the prompter screen includes: displaying the text of the deviation position in a preset highlight color that is different from the normal text, and displaying a rereading indicator mark at the beginning of the most recent complete sentence before the deviation position.

9. The method for real-time monitoring and correction of abnormal audio in live news broadcasts according to claim 1, characterized in that, The broadcast quality analysis report includes the following: The total number of abnormal events categorized by abnormality type, the histogram of the time distribution of abnormal events divided by preset time granularity, the average duration and maximum duration of each type of abnormal event categorized by abnormality type, and the comprehensive broadcast quality score calculated based on the total live broadcast duration and the total number of abnormal events.

10. A real-time monitoring and correction system for abnormal audio during live news broadcasts, used to implement the real-time monitoring and correction method for abnormal audio during live news broadcasts as described in any one of claims 1-9, characterized in that, include: The real-time framing and feature extraction module for audio stream is used to process the raw audio signal collected by the wireless microphone worn by the announcer into frames with a preset frame length. It extracts multi-dimensional acoustic feature vectors, including short-time energy, short-time zero-crossing rate and Mel frequency cepstral coefficients, from each frame of audio signal and stores the extracted multi-dimensional acoustic feature vectors into a circular feature buffer in chronological order. The multimodal anomaly parallel detection module is used to synchronously perform parallel detection of three anomaly modes based on the multidimensional acoustic feature vectors in the circular feature buffer: abnormal pause detection, speech deviation detection, and volume change detection. It determines the presence or absence of audio segments based on short-time energy and determines abnormal pause events when the duration of the silent segment exceeds the genre adaptive threshold. It transcribes the speech signal into transcribed text through a real-time speech recognition engine and compares the transcribed text with the manuscript text. It determines speech deviation events when the number of consecutive mismatched words exceeds the deviation tolerance number. It performs sliding window variance monitoring on the short-time energy sequence and determines volume change events when the energy difference between adjacent frames exceeds the energy jump threshold. The intelligent prompting and audio correction linkage module is used to perform differentiated correction responses based on the abnormal event types output by the multimodal abnormal mode parallel detection module. For abnormal pause events, a prompting sound signal is pushed through the earphone channel. For speech error deviation events, the deviation position is highlighted on the prompter screen and the reading start point is indicated. For sudden volume events, a dynamic compressor is triggered to smooth the volume. The live broadcast quality log generation and feedback optimization module records the timestamp, anomaly type, and correction response of each abnormal event during the entire live broadcast to the live broadcast quality log. After the live broadcast ends, it generates a broadcast quality analysis report and dynamically corrects the genre adaptive threshold and energy jump threshold based on the anomaly distribution characteristics.

Images