Voice activity detection, or VAD, is responsible for making a quick, low-cost determination of whether a snippet of audio contains human speech.
VADs range in complexity from simple frequency analyzers to heavier black-box neural models. The underlying implementation in Spokestack’s libraries varies based on tools available for the various platforms, but we try to strike a balance between speed and accuracy.
VAD is the first gatekeeper in a speech detection pipeline. Ignoring audio that’s not detected as speech saves processing power, and the savings grow with each downstream processor you have in your speech pipeline. The Spokestack speech pipeline is a soft real-time system, and its components must be as responsive as possible. It’s worse for a downstream component to miss user speech than to process too much, so our VADs tend to err on the side of producing false positives rather than rejecting actual speech.