Chrome's new echo canceller halves voice agent word error rate, stops agents answering their own greeting
chadwallacehart · x · 2026-10-10
Philipp Hancke at webrtcHacks analyzes Chrome's new ML-based residual echo estimator: it removes more echo than AEC3 alone, halves a voice agent's word error rate when transcribing the near-end talker, and cuts how often an agent that greets you on connect answers its own greeting.
For regular calls leaked echo is an annoyance, but for a voice agent it's input: the VAD decides someone spoke, STT turns it into words, the LLM answers, and TTS speaks. The author shows a server log catching the agent replying to its own greeting echo as a phantom user turn ("USER: Hmm."), and explains how to diagnose this by comparing listen re-enable timestamps with speech-start events.
More from coding & agent
- mitsuhiko: opencode's isolation tradeoff is fine, but he wouldn't start there with Pi — mitsuhiko · 2026-10-10
- mitsuhiko: an AST interpreter alone is too fragile as an agent code-execution sandbox — mitsuhiko · 2026-10-10
- Agent tool ships global secrets/env/SSH key management, LAN egress next — lucasmeijer · 2026-10-10
- Sierra's Personal Agent Protocol draft out with Meta, Stripe, Walmart; Cloudflare joins — irvinebroque · 2026-10-10
- Claude Code Tip: Drop the 1M-Context Disable Flag, Keep Autocompact at 300-400K — dotey · 2026-10-10
- Humanlayer open-sources a serverless background agent that turns GitHub issues into PRs — samgoodwin89 · 2026-10-10