TV noise crashes voice agents from 38.7% to 8.6%; full duplex is the fix
SIGKITTEN · x · 2026-10-14
- A benchmark of 14 voice agents found that simply playing a TV in the background crashed task completion from 38.7% to 8.6% — background noise is a fatal weakness.
- Commenter pbbakkum notes gpt-realtime is the only model tested with turn detection built into the audio model itself rather than a separate module, explaining its far better turn perception in noisy environments.
- This is enabled by its full duplex architecture: listening and speaking happen simultaneously, so the model reads turn boundaries straight from the audio stream.
- Takeaway for builders: end-to-end full duplex architectures may beat the standard ASR+VAD+LLM pipeline for noise robustness.
Related event: TV Noise Crashes Voice Agent Task Success From 38.7% to 8.6%(2 posts)→
More from coding & agent
- Beancat launches an AI agent for media production with transparent per-run dollar pricing — Scobleizer · 2026-10-14
- Two devs burn $2,025 in API costs building a game with Claude, screenshots show progress — zeeg · 2026-10-14
- Sketch first, code later: prototyping with tldraw before handing specs to Claude — max__drake · 2026-10-14
- Google ran Argon overnight on a 200K-line codebase: 40% faster sessions, 2 of 6 fixes shipped — rseroter · 2026-10-14
- Argon synthesizes its own voice from scratch: pure Python, no audio libraries — fofrAI · 2026-10-14
- Musk RTs: Grok Bot Multi-Agent Team Does Your Grocery Shopping in 3 Minutes — elonmusk · 2026-10-14