Automatic real-time video editing of individual clips from text using artificial intelligence features

Abstract

A system and method for real-time processing and segmenting of a video stream of a procession of multiple people participating in a filmed event being filmed. The method includes segmenting the video stream into packets as the video is being captured; analysing each packet with at least two artificial intelligence classifiers as the video stream is being captured, one of which is used to classify a human movement action, another of which is used to align any spoken audio content with transcribed text to identify a person in the video stream. The method also includes creating a video clip of the identified person, the video clip being segmented based on a combination of the behaviour movement and identification of the person through the AI classifiers.

Description

[0001]AUTOMATIC REAL-TIME VIDEO EDITING OF INDIVIDUAL CLIPS FROM TEXT USING ARTIFICIAL INTELLIGENCE FEATURES Field of the Invention The present disclosure relates to improvements in systems, platforms, and methods for facilitating video editing while an event is being recorded. Background of the Invention Video editing systems exist which are labour-intensive, and require a high degree of subjective judgement, often leading to inefficient processing with inaccuracies. It is important in environments involving video capture, such as graduations, award presentations, etc., that video editing and processing is delivered in a manner that is expedient, and reasonably error-free. Therefore what is needed is a system, platform and method that permits real-time video editing as an event is being filmed. Summary In a preferred aspect, the present disclosure sets forth a method for real-time processing and segmenting of a video stream of a procession of multiple people participating in a filmed event being recorded. The method includes segmenting the video stream into packets as the video is being captured; analysing each packet with at least two artificial intelligence classifiers as the video stream is being captured, a first of the artificial intelligence (AI) classifiers being used to classify a human movement action based on a feature set of digital behaviour movements, a second of the AI classifiers being used to align any spoken audio content with transcribed text to identify a person in the video stream. The method further includes creating a video clip of the identified person, the video clip being segmented based on a combination of the behaviour movement of the person as classified by the first artificial intelligence classifier, and identification of the person though alignment of spoken audio content with the transcribed text performed with the second artificial intelligence classifier. In a further preferred aspect, the present disclosure sets forth a system for creating a video clip of a person from a video stream of an event. The system includes a database configured to store a list of people scheduled to participate in the event. The system also includes a processor configured to: digitally recognise a human behaviour movement action using a first artificial intelligence classifier; align any spoken audio content with transcribed text to identify the person in the video stream using a second artificial intelligence classifier; and create a video clip of the person from the video stream by segmenting the video clip based on a combination of the behaviour movement of the person as classified by the first artificial intelligence classifier, and identification of the person though alignment of spoken audio content with the transcribed text performed with the second artificial intelligence classifier. In another additional preferred aspect, the present disclosure sets forth a method for capturing live video audio / video signal and process the content to repurpose into new video assets. The method includes connecting the client's live video feed via a wide area network (WAN) over the internet; connecting the client live video feed through a video transport protocol such as RTMP, SRT or WebRTC; capturing the client's live video feed through a Content Delivery Network CDN such as AWS IVS or a self-hosted NGINX server; segmenting the live stream video feed in packets (HLS .ts files); analysing packets by artificial intelligence processing within a virtual container; providing timecode and instructions from a database to create segments and clips; assembling the segments and clips by transcoding the video live stream packets with a video processing engine such as FFMPEG; and delivering repurposed branded and multi-format video content through a web graphic user interface GUI for previewing, downloading and sharing. In yet an additional preferred aspect, the present disclosure sets forth a method to automate live video editing with artificial intelligence and machine learning algorithm. The method includes analysing the live video feed audio and video signals via a Python script container instance on a cloud-based or local server; processing of an audio signal to determine significant sound elements of the video feed such as background noise, crowd cheering, music and speech; and processing the audio signal through speech recognition for speech-to-text fuzzy logic analysis and captioning with multi-language capabilities (foreign language speech recognition). It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed, unless otherwise stated. In the present specification and claims, the word “comprising” and its derivatives including “comprises” and “comprise” include each of the stated integers, but does not exclude the inclusion of one or more further integers. The claims as filed with this application are hereby incorporated by reference in the description. The disclosures of U.S. Provisional Patent Application No.63 / 682,777, and Australian Provisional Application No.2024902535 are each incorporated by reference herein in their entirety. The disclosure of U.S. Patent No.10,943,127 describes background media processing, the details of which are incorporated by reference herein to the extent they are needed for context, etc. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate several embodiments of the invention and together with the description, serve to explain the principles of one or more forms of the invention. Brief Description of the Figures Fig.1 is flow diagram of a method to capture live video and process the content to repurpose into new video assets in accordance with a preferred embodiment of the present disclosure. Fig.2 is a flow diagram between a recorder node and a renderer node in accordance with a preferred embodiment of the present disclosure. Detailed Description of the Drawings Reference will now be made in detail to the present preferred embodiments of the disclosure, examples of which are illustrated in the accompanying drawings. Fig.1 shows a preferred embodiment of a platform or system having a database 1.5, a backend processor 1.4, and a router 1.6. In use, an operator accesses database 1.5 and processor 1.4 through a workbench App 1.2. During of an event, such as a graduation ceremony or awards presentation, processor 1.4 segments the video stream into packets, and analyses each packet as part of a method of organizing packets into a video clip of a particular person in the video stream. One or more audio transducers record audio content to accompany the video footage. Multiple aspects of an initially unidentified person walking into view of the camera are analysed to identify the person, and the context of the event the person is participating in. The video stream may then be segmented into video clips involving primarily the identified person, and the video clips tagged / coded within the video stream for easy video clip identification and access. Referring now to Fig.1, the system includes a postgresql database 1.5 storing a list of people scheduled to participate in the event along with metadata for editing and delivering videos clips top customers. Backend system processor 1.4 is configured to access data stored in database 1.5 through via SQL and through to Router 1.6 via OMQ Socket. In a preferred form, processor 1.4 uses pattern recognition and machine learning to identify a person in a video, and the context of the event the person is participating in. It will be appreciated that pattern recognition and machine learning may be implemented through appropriate classifiers, such as an artificial neural network. Other exemplary classifiers include Long Term and Short-Term Temporal Difference Vision Transformer (LS-VIT), which may be used to recognise and classify human movement action; and a Generative Adversarial Transformer, which may be used to process an audio signal of the video stream to match items enunciated in a preconfigured list with a transcription of the video stream produced with speech to text technology. Other classifiers may be used as appropriate such as fuzzy logic models. Backend is interfaced with OMQ / SocketIO router which analyses the HLS .TS files from the Worker, in this case NGINX RTMP stream recorder, along with the Worbench UI via WebSocket which displays the automatic cuts and allows for Quality Control QC and monitoring in real-time. The suggested editing data is stored in the Postgres Database via SQL. REST APIs allow for the flow of metadata and clips delivery between UIs: Workbench, Customer App, Client App. The basic details of machine learning and pattern recognition methods would be understood by those of ordinary skill in the art and are therefore not repeated here for simplicity. Examples of features include: graduation ceremony, awards presentation, sport competition or any event where participants cross a stage or platform individually as part of a procession of people. Many of these features may involve numerical assignment for a set of characteristics within each feature type. Computer-extracted features typically include the image processing features calculated based on the pixels and grey levels related to the extracted areas. Once a set of selection features has been generated, the set is preferably classified using a classifier. There are a variety of techniques suitable for use as a classifier. Suitable classifiers include, but at not limited to statistical applications (e.g., K- nearest neighbor, Bayesian classifiers, rank nearest neighbor, fuzzy pyramid linking, discriminant analysis, logistic regression, multivariant adaptive regression splines, support vector machine, and Hidden Markov Model), neural networks, decision trees, associated rule mining, LS-VIT, Generative Adversarial Transformer, and case- based reasoning, or a combination of any of the foregoing. As shown in Fig.2, recorder node 2.1 interacts with Renderer node 2.2 by processing and analysing the video live stream footage to store editing reference points timecode to / var / record-data 2.6. Renderer node 2.2 consists of a load balancing Network Service via NFS Mount 2.7 for Instant Video Delivery from Real- Time Editing. On Recorder Node 2.1, NGINX RTMP HLS .TS Stream Recorder 2.3 provides footage for Media Worker / Recorder. See, for example, Fig.2. Output metadata are saved in / var / record-data 2.6 and shared to the Renderer Node 2.2 via / var / record- data 2.6 path to a load balancing pool of Media Workers / Renderers. Renderers to concatenate and export the video output according to specs from a FFMPEG script. A preferred use for the invention is to automate the editing process of video for events content dealing with large amounts of content and long-form videos. It is preferably configured to repurpose the content in various formats (length, aspect ratio, selection of content, titling and other effects). An advantage over existing systems is its efficiency at processing many hours of content or tasks requiring a multitude of clipping in minutes that would require several hours of editing with a manual or semi-automated workflow. Examples of an operation of the system are set forth below. In one commercial application example, the system delivers multiple assets of repurposed video content to its client as the live feed is still ongoing. It is designed to defer the complexity and time-consuming nature of video editing by improving the delivery time and cost- effectiveness of the editing process. The system preferably functions by remotely acquiring the video source via a wide area network (WAN) to a video processing instance, assisted by a machine learning algorithm in video signal and natural language processing. The results are then represented visually via a graphic interface that composes video assets based on the processed instances. Processor 104 locates editing points when audio is matched to a preloaded list of individuals: speakers, awards, graduations and creates individual clips in various aspect ratios and with preloaded branding assets. In a preferred form, system 100 permits individual clips of a live-streamed event available in real-time via a GUI customer portal where it is possible to preview, download and share the clips. Example 1 Example 1 relates to a method to capture live video audio / video signal and process the content to repurpose into new video assets. The exemplary method includes: 1. Connecting the client's live video feed via a Wide Area Network (WAN) over the Internet; 2. Connecting the client live video feed through a video transport protocol such as Remote Transmission Messaging Protocol (RTMP), Secure Reliable Transport (SRT) or Web Real-Time Communication (WebRTC); 3. Capturing the client's live video feed through a Content Delivery Network (CDN) such as Nginx-RTMP streaming server; 4. Segmenting the live stream video feed in packets with HTTP Live Stream (HLS), Transport Stream (TS) Video files; 5. Analysing packets by artificial intelligence (AI) processing within a virtual container; 6. Multimodal AI analysis processes including Long and Short-term Temporal Difference Vision Transformer (LS-VIT) and Audio Generative Adversary Networks (GAN) to determine in the footage when the list is being used and to recognise the patterns of the enumeration of the list items; 7. Match the list of items, i.e., name, surname, titles, etc with the footage combining steps of speech-to-text and text-to-speech analysis to determine a high degree of matching confidence; 8. Providing timecode and instructions from database to create segments and clips; 9. Assembling segments and clips by transcoding the video live stream packets with a video processing engine such as FFMPEG; and 10. Delivering repurposed branded and multi-format video content through web graphic user interface GUI for previewing, downloading and sharing. Example 2 Example 2 relates to a method to automate live video editing with artificial intelligence and machine learning algorithm. The method includes: 1. Analysing the live video feed audio and video signals via a Python script container instance on cloud-based or local servers; 2. Processing of audio signal to determine significant sound elements of the video feed such as background noise, crowd cheering, music and speech; and 3. Processing of audio signal through speech recognition for speech-to-text analysis and captioning with multi-language capabilities (foreign language speech recognition). Example 3 Example 3 relates to a method to automate real-time video editing with multimodal AI, comprehensive data integration processing multiple data types simultaneously— visual frames, audio, text, and sensor data—to create unified understanding that surpasses single-modality approaches. The system in this example preferably uses hybrid methods of fusion strategies, including early fusion (combining raw data) and late fusion (merging processed outputs). For live events, multimodal AI enables real-time analysis that combines participant movements with audio patterns to automatically generate clips. This approach leverages techniques like contrastive learning to align embeddings from multiple modalities in shared spaces, enabling complex tasks like searching videos using text queries or generating contextual captions. 1. Analysing the live video feed audio and video signals via a container instance on cloud-based servers using modern frameworks supporting real- time multimodal processing through tools like PyTorch or TensorFlow, with specialized libraries for audio processing (Librosa, Whisper) and video analysis (OpenCV, MediaPipe). 2. Using Long and Short-term Temporal Difference Vision Transformer (LS- VIT) specifically incorporating temporal modelling for action recognition, achieving high accuracy across multiple benchmarks while maintaining real- time performance capabilities. 3. Using Generative Adversarial Transformers for processing audio signal to match items enunciated in a list with the transcription of event footage using speech recognition (speech-to-text) algorithms. Below, the Long and Short-term Temporal Difference Vision Transformer (LS-VIT) is described more particularly in a video context. LS-VIT Algorithm Equations for Recognizing List Enunciation in Events The Long and Short-term Temporal Difference Vision Transformer (LS-VIT) is designed to capture both short-term and long-term motion information in video streams, making it suitable for recognizing complex temporal patterns such as when a list is being enunciated during an event. Set forth below are core algorithmic equations and modules that underpin LS-VIT, adapted for the task of detecting list enunciation. 1. Short-term Motion Information Frame (SMIF) SMIF enriches each frame with local motion details by integrating temporal differences from adjacent frames. This is crucial for capturing the audio patterns and gesture movements typical when a speaker enumerates items in a list. Equation 1: SMIF Construction ^ˆ^ = ^^ + ^(^^)● ^ˆ^: Image enhanced with details. ● ^^: Original frame at time ^. ●^(^^): Differential image highlighting changes between adjacent frames[1][2]. Equation 2: Short-term Temporal Differences ^^ = [^^^^ − ^^^^,● Captures bidirectional motion information for each frame. Equation 3: Forward and Backward Differences ^^^ = ^^ − ^^^^ ^^ = ^^^ + ^^^Equation 4: Averaged Temporal Difference ^(^^) = ^^^(^^^^, ^^^^, ^^ , ^^^^, ^^^^)● Averages over a local window to robustly capture motion trends. Equation 5: Temporal Difference Inhibition (TDI) ^(^ − 1, ^, ℎ, ^) = {0, ^^^ ≤ ^ × !^ + " × !, ^^^ > ^ × ! ● ^: Threshold for noise suppression. ● ": Enhancement multiplier. ● Scale factor for pixel intensity range. 2. Long-term Motion Information Module (LMIM) LMIM captures extended temporal dependencies, essential for recognizing the sequential structure of list enunciation (e.g., repeated pauses, camera angle switching, gestures, or vocal inflections). Equation 6: Channel Reduction for Long-term Features $% = &ℎ'(()*+),-^)($, .)● $: Feature tensor. ●.: Reduction ratio (e.g., . = 6 for optimal performance).Equation 7: Long-term Temporal Differences &= % %^ $^ − $^^^Equation 8: Temporal ^0 = 1(&^ + &^) − 2● 1: Sigmoid function. ● 2: Hyperparameter (commonly 0.5). Equation 9: Feature Enhancement with Attention $ˆ^ = $^ + $^ ⊙ ^0● ⊙: Element-wise multiplication. 3. Integration with Vision Transformer (ViT) The enhanced frames (^ˆ^) and features ($ˆ^) are fed into a standard Vision Transformer backbone, which applies self-attention and feed-forward layers to model complex spatio-temporal dependencies. 4. Application to List Enunciation Recognition To adapt LS-VIT for recognizing when a list is being enunciated: ● Input: Video frames and synchronized audio features. ● SMIF: Captures rapid camera changes and movements, as well as visual cues between list items. ● LMIM: Detects the recurring temporal structure (e.g., pauses, gestures, or vocal inflections) typical of list enumeration. ● Classifier Head: A final classification layer is trained to output a binary or multi- class label indicating whether a list is being enunciated, based on the spatio- temporal features extracted by LS-VIT. Summary Table: Key Equations Module Equation Purpose $ˆ^ = $^ + $^ ⊙ ^0 Feature enhancement Generative Adversarial Network (GAN) Algorithm for Audio-Visual List Matching To match items enunciated in a list with the transcription of event footage—achieving a high matching confidence (i.e., over 95%), a specialised GAN-based architecture is used. This approach leverages both audio and visual modalities, aligning spoken content with transcribed text using adversarial learning and modality-invariant representations. 1. Algorithm Overview ● Input: Video footage with audio (containing list enunciation) and its transcription. ● Goal: Accurately align and match each spoken list item to its corresponding text segment, with high confidence. ● Core Components: o Audio and Visual Encoders: Extract modality-specific features from the footage. o Modality-Invariant Representation Generator: Produces shared representations capturing the commonality between audio and visual streams. o Modality Discriminator: Ensures the shared representation is agnostic to modality via adversarial training. o Mutual Information Maximization: Aligns the shared representation with both modalities, preserving semantic content. o Matching Head: Computes similarity between generated representations and transcription segments, outputting a confidence score. 2. Key Equations A. Feature Extraction Let $ f_v $ and $ f_a $ be the frame-level features from the visual and audio encoders, respectively, for $ T $ frames and embedding size $ D $: ^:×;7, ^8 ∈ +B. Modality-Invariant Representation (Generator) The generator $ G $ produces a modality-invariant representation $ f^{inv}_{va} $: ^^^<7 = ^ GWhere: ● $ f_{va} = [f_v; f_a] $ (concatenation) ● $ s_m = f^{share}_m \otimes \sigma(Conv([f^{spe}m; f{va}])) $ ● $ f^{share}m = CrossAttention(f{va}, f^{spe}_m, f^{spe}_m) $ ● $ \otimes $ is element-wise multiplication, $ \sigma $ is the sigmoid function C. Adversarial Loss (Discriminator) The modality discriminator $ D $ is trained to distinguish between modality-specific and modality-invariant representations: H= L [*>M ^(^OPQ) + *>M (1 − ^(^OPQ))] + L [−*>M ^ ^<7 ^<7IJK ^ 8 7 ^ (^78 ) − *>M (1 − ^(^78 ))] , output 1 and 0, respectively. ● For $ f^{inv}_{va} $, the discriminator should output 0.5 (modality-agnostic). D. Mutual Information Maximization To ensure semantic alignment between the shared and specific representations: :)UV (^^^<7, ^OPQ^ / Z): )UV (^^^<7, ^OPQ^ / Z)H = − A ^ *>M 78,^ 7,^ − A ^ 78,^ 8,^RSR ^*>M ^^OPQ parameter. E. Matching and Confidence Calculation For each list item $ l_k $ in the transcription, compute the similarity with the corresponding segment in $ f^{inv}_{va} $: &>(^^,)(^)(* ) = ?'U F^ ^<7] ^∈_ ?(^78,^, L?a),(*]` ))● $ S_k $ is the $ l_k $. ● $ Embed(l_k) $ is the embedding of the list item text. ● $ sim $ is a similarity function (e.g., cosine similarity). A match is accepted if $ Confidence(l_k) > 0.95 $. 3. Training Objective The total loss combines recognition, adversarial, and mutual information terms: H= HbQ0 + cIJKHIJK + cRSRHRSR● $ L_{rec} $: Cross- ground-truth transcription. ● $ \lambda_{GAN}, \lambda_{MIM} $: Weighting parameters. 4. Summary Table: Core Equations Step Equation Purpose ng Example 4 Example 4 relates to an AI-Powered Algorithm for Text Based Video Clipping System that automatically extracts individual clips from text based list by: 1. Transcribing the video footage audio to detect text items in list such as name and surname or other individual text-based identifiers; 2. Matching text-based identifiers against a pre-provided list with high confidence; and 3. Using computer vision to determine the timecode in / out points based on movement analysis in the footage. This algorithm has processing steps using the original footage and text based identifiers lists to generate timecode and process video into clips with branding overlays and other audio / video assets. Example 5 Example 5 relates to real-time performance optimization strategies using parallel processing implementation and memory / bandwidth optimization. 1. For the real-time performance, the system in this aspect preferably uses effective parallel processing at the frame level. Different frames can be processed simultaneously across multiple CPU cores and / or GPU streams. For a 30 FPS video stream, this means up to 30 processing operations can happen in parallel. 2. It is also using pipeline parallelization: different stages of video processing (decode, enhance, encode) can operate on different frames simultaneously, creating an efficient processing pipeline. 3. GPU Memory Management: keeping video data in GPU memory throughout the processing pipeline avoids costly CPU-GPU memory transfers. 4. Compression Optimization: using advanced codecs like H.265 / HEVC or AV1 can reduce bandwidth requirements by 50% or more compared to older codecs, enabling higher quality streams within the same bandwidth constraints. 5. Smart Caching: intelligent caching strategies at multiple levels (edge, CDN, client) can significantly reduce latency and improve user experience. The present disclosure in a preferred form provides at least one or more of the following advantages. 1. Speed of delivery to the client which is critical in events where there are awards or graduations and individual clips are delivered instantly when the event is still ongoing. 2. Recognising multi-language transcriptions with foreign language speech recognition. 3. The automated nature of the new system, driven by AI, allows for real-time processing and segmentation of video sequences. This scalability is particularly beneficial for large-scale events with numerous participants, where quick turnaround times are crucial. 4. The AI-driven approach reduces the need for human intervention by automating the entire process of video segmentation and participant identification. This not only speeds up the process, but also reduces the potential for human error. The system increases significantly the engagement of individual clip recipient and marketing opportunities for the client. In a preferred form, the system has technical advantages and commercial applications that innovate how live video content is repurposed for social media, broadcast and archival use cases. The system permits real-time processing of live events such as awards, prize-giving ceremonies, convocation and graduation ceremonies where participants pass the stage individually to produce their personalised clip in real-time. One exemplary main commercial application of the system is to deliver multiple assets of repurposed video content to its client as the live feed is still ongoing. It is preferably designed to defer the complexity and time-consuming nature of video editing by improving the delivery time and cost-effectiveness of the editing process. In a preferred embodiment, the system works by remotely acquiring the video source via a wide area network (WAN) to a video processing instance assisted by a machine learning algorithm in video signal and natural language processing. The results are then represented visually via a graphic interface that composes video assets based on the processed instances. The system locates editing points when audio is matched to a preloaded list of individuals: speakers, awards, graduations and creates individual clips in various aspect ratios and with preloaded branding. An advantageous outcome of the system is permitting individual clips of a live- streamed event to be made available in real-time via a GUI customer portal where it is possible to preview, download and share the clips. Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of forms of the embodiments disclosed herein. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

Claims

What is claimed is:

1. A method for real-time processing and segmenting of a video stream of a procession of multiple people participating in a filmed event being filmed, comprising: segmenting the video stream into packets as the video is being captured; analysing each packet with at least two artificial intelligence classifiers as the video stream is being captured, a first of the classifiers being used to classify a human movement action based on a feature set of digital behaviour movements, a second of the artificial intelligence classifiers being used to align any spoken audio content with transcribed text to identify a person in the video stream; and creating a video clip of the identified person, the video clip being segmented based on a combination of the behaviour movement of the person as classified by the first artificial intelligence classifier, and identification of the person though alignment of spoken audio content with the transcribed text performed with the second artificial intelligence classifier.

2. The method of claim 1, wherein the first artificial intelligence classifier includes a Long Term and Short-Term Temporal Difference Vision Transformer to recognise and classify human movement action.

3. The method of either claim 1 or 2, wherein the second artificial intelligence classifier includes a Generative Adversarial Network for audio-visual list matching.

4. The method of any one of the above claims, further comprising using a Generative Adversarial Transformer to process an audio signal of the video stream to match items enunciated in a preconfigured list with a transcription of the video stream produced with speech to text technology.

5. The method of any one of the above claims, further comprising identifying a language being spoken in the audio content.

6. The method of any one of the above claims, wherein the analysis of each packet with the two artificial intelligence classifiers is conducted using parallel processing at a frame level.

7. The method of claim 6, wherein the parallel processing is conducted simultaneously with multiple CPU cores.

8. The method of any one of the above claims, wherein the person is identified in the video clip without using facial recognition technology.

9. A system for creating a video clip of a person from a video stream of an event, comprising: a database configured to store a list of people scheduled to participate in the event; and a processor configured to: digitally recognise a human behaviour movement action using a first artificial intelligence classifier; align any spoken audio content with transcribed text to identify the person in the video stream using a second artificial intelligence classifier; and create a video clip of the person from the video stream by segmenting the video clip based on a combination of the behaviour movement of the person as classified by the first artificial intelligence classifier, and identification of the person though alignment of spoken audio content with the transcribed text performed with the second artificial intelligence classifier.

10. The system of claim 9, wherein the first artificial intelligence classifier includes a Long Term and Short-Term Temporal Difference Vision Transformer to recognise and classify human movement action.

11. The system of either claim 9 or 10, wherein the second artificial intelligence classifier includes a Generative Adversarial Network for audio-visual list matching.

12. The system of any one of claims 9 to 11, further comprising using a Generative Adversarial Transformer to process an audio signal of the video stream to match items enunciated in a preconfigured list with a transcription of the video stream produced with speech to text technology.

13. The system of any one of claims 9 to 12, further comprising identifying a language being spoken in the audio content.

14. The system of any one of claims 9 to 13, wherein said processor is configured to digitally recognise the human behaviour movement and align any spoken audiocontent with transcribed text simultaneously using parallel processing of the video stream at a frame level.

15. The system of any one of claims 9 to 14, wherein the person is identified in the video clip without using facial recognition technology.

Images