Conference Paper
Open Access
Voice-Controlled Audio Routing in Professional Dante Networks Using Large Language Models and the Model Context Protocol
AES Show 2026 Nashville · Paper 57 · October 2026
Abstract
Professional Audio-over-IP (AoIP) networks based on Dante are widely deployed in broadcast, live production,
and educational facilities, yet their management interfaces demand expert knowledge of network topology and
device-naming conventions. This paper presents a peer-reviewed implementation of a voice-controlled Dante
routing system built on Large Language Models (LLMs) and the Model Context Protocol (MCP). We contribute a
purpose-built Dante MCP Server exposing six structured tools for querying and modifying audio routes through the
Dante Domain Manager (DDM) GraphQL API, integrated with local automatic speech recognition and commercial
LLM tool use, and protected by a domain-level safety layer that prevents changes to production infrastructure. The
pipeline was validated on a real, multi-domain Dante infrastructure of approximately 39 devices; end-to-end latency
is approximately 1770 ms, dominated by LLM inference and acceptable for non-real-time supervisory routing.
A controlled within-subject study (N=10; five experts, five non-experts) compared voice against conventional
graphical user interface (GUI)-based routing using the NASA Task Load Index (NASA-TLX) and the System
Usability Scale (SUS). Contrary to our hypothesis, voice imposed significantly higher cognitive load (+41%;
Wilcoxon p=.002, r=.89) and lower usability (SUS 66 vs. 80; p=.008); non-experts nonetheless completed 85% of
voice tasks, and a modality-specific failure signature emerged in real-time correction (four voice failures vs. zero
GUI). We trace these failures to delete-command disambiguation, a concrete and addressable design target rather
than a limit of voice control. The Dante MCP Server is LLM-agnostic and will be released as open source.
