Voice activity detection (VAD)
Voice activity detection is the technique that determines when human speech starts and stops in an audio stream by separating it from silence and noise.
Voice activity detection (VAD) splits incoming audio into very short chunks and decides, for each one, whether it contains speech. Phone systems have long used it to avoid sending data during silence; in voice AI it is one of the core pieces that manage the flow of a conversation. A voice agent notices that the caller has started or stopped talking thanks to VAD. Background noise, music, breathing and line interference are the main things it has to tell apart from speech.
How it works
Early methods were threshold rules based on the energy and frequency of the signal. Today, small neural network models calculate a speech probability for each chunk of audio. The system combines that probability with a threshold and minimum-duration rules to produce "speech started" and "speech ended" events. These events decide which audio is sent to speech recognition, when the agent stops talking because the caller has cut in (barge-in) and when endpointing kicks in.
Why it matters
If VAD is too sensitive, a cough or a door slamming silences the agent; if it is not sensitive enough, the agent keeps talking after the caller has started to speak. Deciding too late that the caller has gone quiet adds response latency, while deciding too early cuts the caller off. Businesses whose customers call from cars, building sites or busy shops should test this behavior in real conditions before going live.
In Speakline
In Speakline, interruption and end-of-turn settings live in the agent settings: you choose whether callers can interrupt the agent, the minimum number of words needed to make it stop, and how long it waits for the end of a turn. You can try the settings with test calls before publishing. See Voice agents for details.