A Guth Labs publication ↗
Back to News

RealtimeVAD’s faster voice-agent signal comes with a cancellation bill

A voice assistant can waste time waiting for certainty that a person has stopped speaking. RealtimeVAD tries an earlier guess, but the useful story is the tradeoff rather than a universal speed claim.

The project’s current README reports a median 92-millisecond lead over Silero’s speech-off signal on 3,153 close-microphone turns from one speaker. It reports no median lead on a distant table microphone. The early pause hint intentionally fires often, so a pipeline must be able to cancel speculative work when speech resumes.

That distinction matters: an earlier hint is not the same as an equally large reduction in user-perceived response time. The developer reports workload-dependent results and warns that speculative requests can increase.

The changelog says version 0.1.0 made the normal final endpoint match Silero while improving the separate pause signal. Version 0.1.1 adds configurable quiet streams without changing existing events when those streams are unused. PyPI dates those releases November 13 and 14 respectively and classifies the package as alpha.

Guth analysis

This is an interesting design for local voice stacks because it moves the optimization question from “How early can we stop listening?” to “How cheaply can we start and undo work?” The latter is safer engineering territory: keep a reliable final endpoint, then make early processing disposable.

For a pipeline with expensive requests or poor cancellation, speculation could simply turn latency into a larger bill. For fast streaming components, it may add little. The right evaluation therefore measures first audible response, discarded computation and interruption behavior together.

Evidence boundary

The measurements are the developer’s, not independently reproduced here. Older launch summaries quoting roughly 110 milliseconds should not replace the current, explicitly qualified documentation. The source pages are mutable; their exact latest edit times were not established.

Sources and citations

Revision history

Published .