English: Co-Listening AI Havruta Learning Support System and Method
The system addresses online lecture isolation and high conversion costs by integrating real-time Havruta-style questioning, enhancing social presence and motivation through synchronized AI-generated questions, thus improving educational outcomes.
Patent Information
- Application Number
- KR1020250057899
- Authority / Receiving Office
- KR · KR
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-05-01
- Publication Date
- 2026-11-10
AI Technical Summary
Online lectures suffer from social isolation, reduced sense of realism, decreased learning motivation, and high costs in converting existing video content to AI-friendly interactive formats, with existing systems lacking real-time question synchronization.
A system integrating a video player, speech-to-text conversion, large language model (LLM) question generation, and virtual human modules to enable real-time Havruta-style questioning during video lectures, using modules like STT, LLM, and virtual human to generate and deliver questions synchronized with the lecture timeline.
Enhances social presence, improves learning motivation and engagement, reduces content conversion costs, and increases completion rates by providing interactive learning experiences without reshooting videos.
Smart Images

Figure PAT00003_ABST
Abstract
Description
Technology Field
[0001] The present invention AI learning support technology for online education Regarding, more specifically Learners and AI agents listen to video lectures simultaneously do, Real-time Havruta-style questioning and discussion This relates to a system and method that alleviates learning isolation and enhances learning effectiveness through [this]. Background Technology
[0002] 1. Limitations of one-way video lectures
[0003] Online lectures lack interaction, causing learners' social isolation This is getting higher sense of realism This decreases. Accordingly Decline in motivation and engagement , Decline in completion rate Problems are frequent
[0004] 2. Content Recycling Issues in the AI Era
[0005] Generative AI is becoming widespread, but existing video lectures Recreated as AI-friendly interactive content It requires enormous costs and time to do so.
[0006] 3. Conventional technology
[0007] Chatbot-based Q&A systems exist, but Real-time questions synchronized with the lecture timeline It has no function.
[0008] · There is research on AI study partners, but Watching videos together while Trigger a Havruta conversation A structure that does so was not proposed. Prior art literature
[0009] KR 102124790 B1 - "Havruta Learning System and Havruta Learning Platform" (Registered 2020-06-30) Provides online Havruta partner matching and question generation functions. Differentiated from our invention as it is not directly synchronized with video. KR 2020-0045863 A - "Havruta Learning Support Device and Method" (Published 2020-04-24) Includes learning result analysis and vulnerability suggestion. No 'simultaneous listening' or [[PAUSE]] control. ETRI Technology Transfer Notice 1130-2022-02112 - "Real-time Streaming End-to-End Speech-to-Text Conversion Technology" (2022) Key feature is minimization of STT delay (≒1 s). Similar to the STT module of our system and can be cited as prior art.
[0010] Yu et al., "Context Selection and Rewriting for Video-based Educational Question Generation," arXiv 2504.19406. Proposal of a model that generates real-time questions by analyzing video subtitles. Similar to our LLM-Question module. Ryu & Lee, "Effects of the Havruta Learning Method Using ChatGPT on Pre-service Early-Childhood Teachers," Asia-Pacific J. of Research in Early Childhood Education, 2024. A study on the educational effects of simulating Havruta dialogue with ChatGPT. Suitable as a conceptual prior art study. ACM DL, "Examining the Feasibility of AI-Generated Questions in Educational Videos," Proc. CHI EA 2024. Experiment on LLM-based video question generation. Provides data on question-transmission latency and accuracy. The problem to be solved
[0011] · Occurring in online lectures Social isolation, decreased sense of reality, reduced learning motivation and completion rates to resolve and
[0012] · Without reshooting existing video content AI-based interactive learning The purpose is to provide technology that can be converted to. means of solving the problem
[0013] 1. System configuration (Fig. 2)
[0014] ·Video player module (100) Lecture Streaming
[0015] ·Speech-to-text conversion (STT) module (130) Real-time transcription using CLOVA Speech API
[0016] ·LLM API module (140) Transcriptional text analysis
[0017] ·Trigger inspection module (150) Detect question timing based on keywords and time
[0018] ·LLM question generation module (160) : Create Havruta template question, include [[PAUSE]] token
[0019] ·Virtual Human Module (120) : Question Voice / Avatar Animation Output
[0020] · Learner UI module (110) : Learner voice / text input, feedback display
[0021] · Video control : Pause / Play
[0022] 2. Operation procedure (Fig. 1, sequence)
[0023] · Lecture streaming → STT 0.8-1.5 s transcription
[0024] · Trigger detection ≤ 1 s → [[PAUSE]] → Question generation 0.6-1.2 s
[0025] · Learner Response → LLM Feedback → Resume Playback
[0026] · Query frequency: 1 time every 5 minutes (variable)
[0027] · Overall perceived latency ≤ 3 s
[0028] 3. Claim Summary
[0029] · Independent Port 1 (System) : Simultaneous listening · Havruta questioning · Pause / play control
[0030] · Dependent term STT delay, trigger algorithm, [[PAUSE]] token, virtual human, etc.
[0031] · Independent Clause 9 (Method) , Independent Port 11 (Program Product) Effects of the invention
[0032] · Restoring social presence : Alleviation of isolation caused by AI co-students
[0033] · Improvement of learning motivation and engagement : Active participation → Increased attention duration
[0034] · Improvement in long-term memory transfer rate Question-and-Answer-Based Higher-Order Cognitive Activation
[0035] · Content Recycling Convert existing lectures to interactive format without re-recording
[0036] · Cost and time savings Easy to apply to both educational institutions and individual instructors Brief explanation of the drawing
[0037] ·Figure 1 : Sequence diagram representing the learning support procedure · Fig. 2 Block diagram of a learning support system Specific details for implementing the invention
[0038] 1. System prototype
[0039] · Client: Flutter-based mobile app, WebSocket control
[0040] · Server: Python FastAPI + 1 GPU environment (16 GB VRAM)
[0041] · STT: NAVER CLOVA Speech API (Korean model) - Estimated average latency 1.2 s
[0042] · LLM : GPT-4o API - Expected response 0.9 s (based on 256 tokens)
[0043] 2. Prophetic Example
[0044] When the system of the present invention is applied to 20 high school mathematics lectures (30 minutes each), based on prior research results and internal simulations, the following effects expectation It works.
[0045]
[0046] 3. Modified embodiment
[0047] · The STT module can be replaced with an English or Japanese model, and the latency is predicted to be within 1.5 s.
[0048] · The frequency of questions can be dynamically adjusted at intervals of 3 to 8 minutes depending on the difficulty of the lecture.
[0049] · The virtual human module can be changed into various visual formats, such as 2D avatars and 3D virtual humans.
[0050] The above figures are based on simulation and prior research values, not actual implementation results. Predicted value ...and may vary depending on the implementation environment.
[0051] [Example]
[0052] 1. System Configuration (See Fig. 2)
[0053] ① Video player module (100): HLS streaming, 1080p, 30 fps
[0054] ② Learner UI Module (110): Flutter 3.29-based mobile app
[0055] ③ Virtual Human Module (120): 2D Live Animation Avatar
[0056] ④ Speech-to-Text (STT) module (130): NAVER CLOVA Speech API (Korean-only model, mono / 16 kHz)
[0057] ⑤ LLM API module (140): GPT-4o, up to 256 token responses
[0058] ⑥ Trigger check module (150): Keyphrase occurrence frequency-based algorithm
[0059] ⑦ Question generation module (160): Havruta template + "[[PAUSE]]" token preceding output
[0060]
[0061] · The above value Prediction / Expectation This is finalized and subject to change during actual implementation.
[0062] 3. Expected learning effects
[0063] Based on a synthesis of internal simulations and prior research results, when applying this system
[0064] It is expected that the social isolation index will decrease by approximately 25%, and the learning motivation score will improve by more than 15%. expectation It works.
[0065] These figures may vary depending on the actual implementation environment. Industrial applicability
[0066] The present invention relates to an online education platform, a learning management system (LMS), a Massive Open Online Course (MOOC), an in-house education solution for companies and institutions, etc. Digital education service industry It can be applied across the board. It can be implemented and deployed software-wise in cloud server and mobile application environments, and takes the form of converting existing video lectures into interactive AI learning content. Mass production · Commercial service provision This is possible. Therefore, the present invention Industrial applicability It is fully equipped.
[0067] Therefore, the present invention It has industrial applicability.
Claims
Claim 1 A computer-implemented learning support system that enables an artificial intelligence agent to perform **real-time discussion** with a learner while a learner and an artificial intelligence agent simultaneously listen to a video lecture, comprising: (a) a video player module that plays lecture audio; (b) a speech-to-text conversion module that transcribes the audio in real time; (c) a trigger check module that analyzes the transcribed text to determine a question trigger point; (d) control logic that pauses the lecture at the trigger point and question generation logic that calls a large-scale language model to generate a Havruta question; and (e) a user interface that presents the question to the learner and receives the learner's response. Claim 2 A computer-implemented learning support system according to claim 1, wherein the pause command is transmitted to a video player module using a "[[PAUSE]]" token.
