Real-time voice agent finds Silero VAD too slow, seeks ultra-low-latency speech-offset detection

Lonligrin · reddit · 2026-09-09

A developer building a real-time voice agent with a speculative response pipeline found Silero VAD v6 too slow at detecting when the user stops speaking. Their architecture needs to trigger speculative ASR and LLM generation the moment possible silence appears, so they'd happily trade false positives for ultra-low latency — speculative work can simply be canceled.

They currently roll a lightweight RMS-energy detector that emits an early "possible silence" signal. Questions: what do others use for ultra-low-latency speech-offset detection; are there VADs that react to the first silent frame without large buffer windows; local, CPU-friendly solutions preferred.

Original post →

More from coding & agent

coding & agent channel →