Glossary · Speech
What is voice activity detection (VAD)?
Voice activity detection (VAD)
Voice activity detection, or VAD, is the part of a voice system that tells speech apart from silence and background noise, many times a second. Everything else in the call waits on its answer.
On a call
A delivery customer takes the call at a busy signal, with a horn going off every few seconds. A weak VAD hears the horn as speech, keeps the turn open, and the agent never replies.
AI agents from ₹1 a minute.¹
Explained
Here is how it works, in plain words.
A voice agent can’t think about what you said until it knows you said something. VAD answers that question on short slices of audio, many times a second: speech, or not speech.
Phone networks have used it for decades. ITU-T G.729 Annex B, from the 1990s, pairs a voice detector with silence compression, so a line sends less data while nobody is talking and plays soft comfort noise instead. Voice agents use the same idea for a different job: to know when the caller started and stopped.
Three decisions hang on it. Barge-in needs VAD to notice the caller talking over the agent. Endpointing needs it to notice the caller going quiet. Speech recognition uses it to skip stretches with nothing to transcribe.
Indian calls are hard on it. People pick up on a scooter, or with a pressure cooker whistling in the kitchen. A detector tuned for a quiet office treats some of that as speech, and the agent waits for a sentence that never comes. Tune it the other way and soft-spoken callers get cut off.
Most modern VADs are small neural networks trained on noisy audio, and they cope far better than an old loudness threshold. They still can’t read intent. A cough and the first syllable of an answer can look alike for a fraction of a second.
That’s why good agents don’t let VAD decide alone. They wait for the recogniser to produce words before treating a sound as the caller’s turn, and they recover when a false alarm cancels something.
In Frequency
This is what Frequency does about it.
A cough or a hum on the line used to be able to cancel a reply our agent had already started. Now a sound with no words in it can't: the agent picks the reply back up and carries on. With adaptive interruptions on, the agent also learns the caller's own volume, so quieter voices in the room are ignored.