New open-source RealtimeVAD detects end of speech faster than Silero VAD
A newly released tool called RealtimeVAD adds a fast loudness-comparison check on top of Silero VAD to flag when a speaker has stopped talking. Benchmarked on 3,279 real conversational turns, it signals speech-off a median 100ms before Silero's probability score drops and 206ms before Silero's VADIterator reports an end, beating Silero's end timing in 82% of turns by a median 70ms. The trade-off is that it also fires roughly ten times more often than Silero during brief pauses between words.
GoKawiil's interpretation of the reporting above, not reported fact.
Faster end-of-speech detection could shave meaningful latency off voice assistants, since speech recognition, LLM processing and voice output all wait for that single signal before starting. The higher false-positive rate suggests developers would need to tune or filter the extra triggers depending on how tolerant their application is to premature cutoffs. Running on a single CPU thread without requiring torch could also make it easier to deploy in lightweight or resource-constrained voice pipelines.
- RealtimeVAD layers a 10ms audio-level comparison on top of Silero VAD to detect speech pauses earlier.
- It beats Silero's end-of-speech timing by a median 70ms in 82% of tested turns, but triggers about ten times more false positives in short word gaps.
- It runs on a single CPU thread using only numpy and onnxruntime, with the Silero model bundled and no torch dependency.
Blue Yeti USB Microphone — If you're building low-latency voice pipelines like RealtimeVAD, clean input audio matters as much as fast endpointing logic. A quality USB mic like the Blue Yeti reduces background noise and room echo, giving your VAD and STT models a cleaner signal to work with for faster, more accurate speech detection.
See Blue Yeti USB Microphone on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: github.com, 2026-11-12
Published there as: “Show HN: RealtimeVAD – trading false positives for 110ms earlier end-of-speech”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.