TBLT-AgentSLA × NLP Research Project

TBLT-AgentVoice AI Task Interlocutor

TBLT-Agent is a collaborative effort of second language acquisition (SLA) and natural language processing (NLP) researchers to build a real-time voice AI system. Grounded in task-based language teaching (TBLT) and computer-supported collaborative learning (CSCL), the AI carries out communicative tasks with elementary EFL learners in spoken English.

We design the AI not as a response tool but as a Task Interlocutor, and study in real classrooms how its interactional role shapes learners’ task performance and L2 learning.

Try the System
  • Georgia State University
  • Seoul National UniversitySeoul National University
  • KAIST
  • Cheongju National University of EducationCheongju National University of Education

System

Starting from one-on-one learner–AI conversation, TBLT-Agent is expanding into three-party dialogue in which the AI supports collaboration between two learners.

  1. System 01 · Realtime

    Realtime Task Interlocutor

    • In Classroom

    An AI character carries out a spoken English task with one learner in real time. Its interactional role is set to either dominant or collaborative.

    Participants
    1 learner · 1 AI
    Voice
    Real-time conversation on the OpenAI Realtime model
    AI role
    Dominant · Collaborative — after Storch’s pair dynamics model

  2. System 02 · Pipeline

    Multiparty Tutor

    • In Development

    Two learners develop one task together over three steps. The AI waits while they talk; once both finish a step, it speaks as an English tutor, noting what they did well together and then coaching each learner on one learning point.

    Participants
    2 learners · 1 AI
    Voice
    Per-learner speech recognition → LLM → speech synthesis pipeline
    Design
    A state machine controls steps and practice turns; the LLM generates grounded feedback (review · encourage · repair · examples). Up to two practice tries per learner

  3. System 03 · Pipeline

    Multiparty Facilitator

    • Planned

    The AI joins the peer conversation as a participant with its own name and voice — a more proficient English-speaking friend who supports the two learners’ collaboration. At each moment it decides when to speak, to whom, and with which dialogue act, and waits when the learners need the floor.

    Participants
    2 learners · 1 AI
    Design
    Utterance understanding (NLU) → dialogue state tracking (DST) → dialogue act selection (Policy) → response generation (NLG)
    Research
    A dialogue act schema for the LLM, LLM evaluators validated against teacher annotations, and a benchmark of off-the-shelf LLMs’ intervention decisions

Get Started

How to access

Use the test account below to experience the system as a learner.

  1. STEP 01

    Open the learner system

    Follow the link below to open the TBLT-Agent login screen.

    tblt-agent.net
  2. STEP 02

    Enter the test account

    Enter the student test account details in the corresponding fields.

    Student ID
    99999
    Name
    test
    Access code
    aa00
  3. STEP 03

    Start the English task

    Allow microphone access when prompted, then begin the real-time voice task.

    A Chromium-based browser with a connected microphone is recommended.
Open the system

Studies

Studies conducted by deploying TBLT-Agent in elementary EFL classrooms.

Study 02 · 2026 Fall

When, to Whom, and How LLM Participates?

We establish pedagogical criteria for when an LLM should speak, to whom, and with which dialogue act as it joins two learners’ collaborative English dialogue. Collecting learner–learner–AI conversations in real elementary EFL classrooms, we build a dataset and benchmark.

  • In Progress
  • Dataset & Benchmark
RQ1
How can a dialogue act schema be defined for an LLM that facilitates collaborative language learning among young EFL learners?
RQ2
How closely do LLM evaluators match teachers when judging the appropriateness of LLM dialogue acts under this schema?
RQ3
How appropriately do off-the-shelf LLMs choose when to intervene and which dialogue act to use?
Methods
Learner–learner–AI dialogues from ~100 students in four elementary EFL classes → initial schema → finalized with teacher feedback and annotation → comparison of GPT, Claude, Gemini, Llama, Qwen, and others

Team

The SLA group grounds tasks and interactional roles in theory; the NLP group turns them into a real-time voice system.

Faculty

  • YouJin KimProfessor
    Applied Linguistics and ESL
    Georgia State University
  • Yohan JoAssistant Professor
    Graduate School of Data Science
    Seoul National University

SLA Group

  • Sak LeePh.D. Student
    Georgia State University
  • Junbo KohResearcher
    Learning Sciences Research Institute, SNU

NLP Group

  • Junbo KohResearcher
    Learning Sciences Research Institute, SNU
  • Seongyeub ChuPh.D. in Data Science
    KAIST
  • Minje KimPh.D. Student in Elementary English Education
    CNUE

Publications

  • Generative AI as a Task Interlocutor in Task-Based Language Learning: Interactional Roles, Corrective Feedback, and L2 Development
    YouJin Kim, Sak Lee, and Junbo Koh
    Modern Language Journal, (In-Progress) [SSCI, Q1, IF 4.5]
  • Generative AI as a Task Interlocutor: Effects of AI Interactional Roles on Collaborative Task Performance and L2 Learning
    YouJin Kim, Sak Lee, and Junbo Koh
    System, (In-Progress) [SSCI, Q1, IF 8.0]
  • Effects of AI Interactional Roles on Collaborative Task Performance and L2 Learning among Young Learners
    YouJin Kim, Sak Lee, and Junbo Koh
    AAAL 2027: American Association for Applied Linguistics, (Accepted) (Link)