Glossary · Conversation
How does end-of-turn detection work?
End-of-turn detection
End-of-turn detection predicts whether the caller has finished what they wanted to say and is handing the conversation back. It listens to the words and the voice as well as the silence.
On a call
Caller: "मेरा order number है four five…" and a pause. The silence is long, but the words say more is coming. A good detector waits for the last two digits before the agent replies.
AI agents from ₹1 a minute.¹
Explained
Here is how it works, in plain words.
Silence is a weak clue. People stop to breathe, to look something up, or to let a child finish shouting in the background. If the agent treats every pause as the end of the turn, it keeps talking over half-finished sentences.
End-of-turn detection adds two more clues. The first is the words. “मेरा नाम है…” ends on an open slot, and so does “pincode है four zero zero”. A finished answer like “हाँ, कल ठीक है” closes cleanly. The second is the voice. Pitch tends to fall at the end of a statement, and people often stretch the last syllable when they’re about to hand over.
Modern detectors are small models that read the live transcript, and sometimes the audio, and give a probability that the turn is over. When the chance is high the agent can reply almost at once. When it’s low the agent waits longer, up to a ceiling, in case the caller is simply slow.
That’s the trade. Every millisecond you wait after a finished turn is dead air. Every early guess is a caller cut off mid-thought.
OpenAI’s Realtime API exposes the two approaches by name. Server VAD splits turns on silence. Semantic VAD uses a classifier over the user’s words to decide whether they’re done. Researchers who study turn-taking describe the same shift: from fixed silence rules toward models that predict where a turn will end.
Indian calls add a twist. Numbers, addresses and names come in chunks, often in a mix of Hindi and English, and callers pause between the chunks. A detector trained only on English audio can cut those calls short.
In Frequency
This is what Frequency does about it.
On the Realtime API, server_vad and semantic_vad both use our own turn detection, so a client built for OpenAI's protocol works without retuning. Our agents also wait longer when the words say a number or an address isn't finished yet, and reply quickly after a plain yes or no.
Questions
People often ask us these questions.
Is end-of-turn detection the same as endpointing?
They overlap. Endpointing asks whether the speech has stopped. End-of-turn detection asks whether the caller meant to hand the turn over, which a pause alone can't tell you.
What is semantic VAD?
It's OpenAI's name for a turn detector that reads the words. Its docs describe a classifier that estimates, from what the user has said, how likely they are to be done, and waits longer when that chance is low.
Still deciding? Ask us.
Fifteen minutes with the team that builds it. Bring your script, or your hardest call.
Book 15 minutes