Skip to content

Phoenix: a voice-based, non-embodied LLM conversational peer agent for group learning

Design · All designs
Evidence · 3 claims (3 for) · 1 study (1 causal), q2 · 0 of 1 report an effect size · 3 claims rest on one study

Description

Phoenix is a "voice-based, non-embodied conversational agent that joins real-time group discussions through spoken dialogue with multiple participants", positioned as an equal peer supporting knowledge co-construction rather than an authoritative tutor. Its architecture integrates Google speech-to-text, Microsoft Azure text-to-speech, and OpenAI's GPT-4.1-mini on a Django backend, with a GPT-4.1-mini-based gatekeeping classifier analyzing the eight most recent dialogue turns to decide when Phoenix should respond. Phoenix was implemented with Indian, British, and African American accents while keeping the LLM prompt and tasks constant.

Design Implications

Context

Requirements

  • Runs on a shared laptop at the group table with a web interface displaying spoken output alongside a live transcript
  • A gatekeeping classifier that analyzes the eight most recent dialogue turns to determine when the agent should respond
  • The LLM generating responses is not informed of the synthesized accent type, to prevent accent-related cultural biases in outputs

Constraints

  • Designed for one-off, lab-based workshop interactions in this study rather than longitudinal classroom use
  • Technical constraints, including latency and limited model transparency, affected interaction flow and trust

Target Learners

  • K-12 teachers in small-group problem-solving tasks

Learning Goals

  • Collaborative problem-solving and consensus building
  • Equitable, human-led participation in group discussion

Claims

Examples

-

Key Sources

  • Ravi, P., Stevens, C., Hurt, B., Hanks, B., Lin, G., and Anderson, E. (2026). Exploring How Agent Voice Accents Shape Human-AI Collaboration in K-12 Group Learning. https://arxiv.org/abs/2606.12805. doi:10.48550/arxiv.2606.12805
Page metadata
Field Value
type design
id phoenix-voice-llm-peer-agent-group-learning
title Phoenix: a voice-based, non-embodied LLM conversational peer agent for group learning
description Phoenix is a "voice-based, non-embodied conversational agent that joins real-time group discussions through spoken dialogue with multiple participants", positioned as an equal peer supporting knowledge co-construction...
status review
generated by process:settle-candidates on 2026-10-10
trust unverified — no one has confirmed this page's content
sources 1 — ravi-2026

Frontmatter is the page's machine-readable record — see the Schema & Guide.