Skip to main content
Scicom AI
Contact us

Semantic VAD cuts wrongly cut-off turns to 20.0%

False cutoff rate across ten end-of-turn systems at two latency budgets, measured over 300 real call turns on Malaysian telephony. The lowest of any system we tested, at both.

Getting cut off mid-sentence is the fastest way to make a voice agent feel broken. We measured how often it happens. 300 real call turns on Malaysian telephony.

The metric is false cutoff rate: how often a system cuts you off while you're still talking.

At 0.6s latency:

  • Scicom Semantic VAD: 20.0%
  • Next best: 27.8%
  • Plain VAD baseline: 32.1%

At the harder 0.3s budget the gap widens: 38.6% for us against 55.7%.

Two ranked bar charts of false-cutoff rate, lower is better, over a 300-turn private telephony test. At a 0.3 second latency budget: Scicom Semantic VAD 38.6, Semantic-VAD whisper-base v6 45.0, Semantic-VAD whisper-small v6 48.6, LiveKit v1-mini audio local 55.7, Semantic-VAD whisper-tiny v6 57.9, ultraVAD no context 65.7, LiveKit v1 cloud audio 66.7, smart-turn v3.2 73.6, smart-turn v2 77.1, VAD baseline 77.9. At a 0.6 second budget: Scicom Semantic VAD 20.0, Semantic-VAD whisper-small v6 22.1, Semantic-VAD whisper-base v6 23.6, LiveKit v1 cloud audio 27.8, LiveKit v1-mini audio local 28.6, Semantic-VAD whisper-tiny v6 28.6, smart-turn v3.2 31.4, VAD baseline 32.1, smart-turn v2 32.1, ultraVAD no context 32.1.
Best false-cutoff rate at each latency budget. Private telephony test, 300 turns, false-cutoff rate in per cent, lower is better. Bold blue labels are Scicom models.

We built this on LiveKit's open work on end of turn detection, and we have open sourced three of the models so you can run the comparison yourself.

Talk to us if you want to explore our frontier model.

First published on LinkedIn, 26 August 2026. The figures are reproduced as published; the test set is a private telephony corpus and is not redistributable.

More from blog

← Back to the Newsroom