Audience: a 33M-param classifier that tells multi-character agents who each line is for
alexcovo_eth · x · 2026-11-01
Developer Jon P. Taylor shared Audience, a tiny (33M-param, ONNX, built on Ettin) classifier model for use in Pipecat voice agent pipelines. It determines who each line in a scene is addressed to.
- It uses names, descriptions, nearby items and character positions to pick one character, several, the whole group, or return "unclear" when the speech is ambiguous.
- Runs in 6–7 ms on CPU, slotting into a Pipecat pipeline without hurting latency.
- Weights are publicly available; the author says multi-character agents are really fun to build.
More from coding & agent
- Cursor launches Annotate in design mode: give agents visual prompts by editing the page — soleio · 2026-11-01
- Open-source Garmin MCP turns training data into AI-built dashboards — indaco_dev · 2026-11-01
- GPT-5.6 Sol reportedly beats Factorio, launching a rocket autonomously — Realistic-Total-3553 · 2026-11-01
- Open-source CodeRail lets teams share one Claude Code session from any device — L_i_n_k0666 · 2026-11-01
- Dev observes AI has made systems-heavy games like RTS and strategy easier to build — IanArawjo · 2026-11-01
- Cursor ships Agent SDK, the framework behind Bugbot, with one-command deploy — dean_rie · 2026-11-01