STELLA: An Open-Source Platform for Auditable Voice-First Conversational Health Coaching Agents
Journal
Proceedings of the 8th ACM Conference on Conversational User Interfaces
Series
ACM International Conference Proceedings Series
Type
conference paper
Date Issued
2026-07-21
Author(s)
Abstract
Deploying voice agents powered by LLMs creates tension between system auditability and the low latency required for natural conversation, while imposing heavy infrastructure burdens. We present STELLA, an open-source platform enabling the efficient deployment of voice-first agents. As a platform, STELLA provides ready-to-use infrastructure, including plan-based session management, modular expert pipelines, and a developer SDK. This eliminates the need for researchers to build complex audio and safety backends from scratch. To address the latency and auditability tension, STELLA’s default agent structurally decouples conversational presence from backend processing. Upon user input, the agent immediately delivers backchannel acknowledgments to maintain natural conversational flow, while safety arbitration and expert routing execute concurrently. A 32-participant field evaluation showed STELLA achieves 1.5-second median latency on commodity scientific hardware and above-median conversational quality in the 50th to 75th percentile ASAQ representative set, demonstrating feasibility for health coaching deployment and motivating future work on clinical applications such as cognitive stimulation therapy.
HSG Classification
contribution to scientific community
Refereed
Yes
Publisher
ACM
Pages
7
Event Title
ACM Conversational User Interfaces 2026 (CUI '26)
Event Location
Bremen, Date
Event Date
21 - 24 July 2026
Subject(s)
Division(s)