Glossary · Conversation
What is endpointing in speech recognition?
Endpointing
Endpointing is how a speech system decides that the caller has stopped speaking. It is the moment the agent starts working on its reply.
On a call
A customer reads out a pincode: "four zero zero… एक second… zero zero one." If endpointing fires at that pause, the agent hears half a pincode and asks for it again.
AI agents from ₹1 a minute.¹
Explained
Here is how it works, in plain words.
Every voice agent has to guess when you’re done talking. People pause mid-sentence to think, to read a number off a slip, or to ask someone in the room. A pause doesn’t always mean the turn is over.
The simplest endpointer waits for a fixed stretch of silence after speech. Set it short and the agent jumps in while the caller is still reading out an address. Set it long and every reply starts with dead air, because nothing else can start until the endpointer says go.
That wait sits inside every turn of every call. It’s often a bigger share of the delay a caller feels than the language model or the voice.
Better endpointers listen for more than silence. Researchers at Google trained a detector to predict whether a pause was final, using cues like filler sounds and how fast the person had been speaking. On voice search it ended queries about 100 ms sooner at the same accuracy.
Words help as well. “मेरा नंबर है…” tells you more is coming, however long the pause. A recogniser that streams partial text lets the agent use that, which is where endpointing starts to blur into end-of-turn detection.
The two get mixed up. Endpointing is mostly the acoustic question: has the speech stopped? End-of-turn detection asks whether the speaker means to hand the turn over. A good agent uses both, and treats a pincode read in two halves differently from a quick “हाँ”.
In Frequency
This is what Frequency does about it.
Our agents wait longer while someone is reading out a number or an address, and answer sooner after a plain yes or no. On the Realtime API our turn detection makes this call for you, so the silence timers other providers expose are accepted and ignored.