Skip to content

Feature · Voice activity detection

Voice activity detection for AI voice agents.

PyAI hears when a user is talking and when they have finished, in an app, on a website or on a phone line, so the agent never cuts in.

Get a sandbox key$50 credit. No credit card required.
Knows the user is done in ~0.2 secondsWaits while numbers are read outStops talking when interruptedLive transcription at $0.001/min, silence not billed

What PyAI gives you for voice activity detection

  • PyAI voice activity detection spots the end of a user's turn about 196 ms after they stop, checking the words as well as the silence so a pause to think never cuts the speaker off.
  • Card numbers read aloud in groups of four stay inside one turn, so the agent reads all sixteen digits at once instead of asking the user for them twice.
  • When a user talks over the agent, PyAI sends a flush signal and your app stops the agent's voice, and developers can raise the end-of-turn wait mid-session for slow speakers without reconnecting the stream.
  • Voice activity detection on live transcription bills $0.001 per minute of active speech, so silent stretches on a connected socket cost nothing.

What happens when a user pauses, reads digits or interrupts?

Pick a moment that confuses a simple silence timer, then play it. The lines are a sample script, not a real recording.

Pick a moment, then press play

UserAI voice agentTurn detector
Ready00:00

How does PyAI voice activity detection work in a live conversation?

Everything runs in one engine, in one place. Hearing the user, deciding they are done and replying happen together, so nothing waits on another company's service.

The user speaks

The user's voice reaches PyAI as PCM16 audio in 20 ms frames, from an app, a browser mic or a device. Phone line audio stays at 8 kHz end to end.

PyAI hears it

PyAI speech recognition starts transcribing at once. The first partial transcript arrives in about 200 ms.

PyAI decides: pause or done?

PyAI turn detection weighs the length of the silence and the words so far. If the user stopped on “to…” or halfway through a card number, the turn stays open.

The agent replies

Once the user is done, PyAI marks the turn as finished and the agent starts answering.

The user has finished

PyAI knows about 0.2 seconds after they stop. The agent's first audio comes about 0.4 seconds after that pause began.

The user cuts in

PyAI notices the user talking over the agent and tells your app to stop the agent's voice.

Which voice agents need turn detection most?

Good turn-taking is the difference between an agent that sounds natural and one that keeps talking over people. Apps, websites, kiosks, cars, smart devices and support lines all show the difference.

In-app assistant

Users who think out loud

A user explaining a billing problem inside a mobile app often stops and restarts mid-sentence. The AI voice agent waits for the full explanation, so it answers the real question the first time.

Website voice bot

Card and account numbers

A shopper reads a 16-digit card number in groups of four with gaps through the browser mic. PyAI keeps the whole number in one turn, so the agent does not ask for it twice.

Kiosk ordering

Orders with long pauses

A customer at a self-service kiosk pauses to read the menu between items. PyAI checks the words, so the agent waits for the full order instead of replying after the first item.

In-car assistant

Drivers who speak slowly

A driver splits attention between the road and the request and speaks with long gaps. Raising the end-of-turn wait gives the driver more time before the agent replies.

Tutoring app

Learners who interrupt

A student cuts in with “wait, why?” mid-explanation. PyAI stops the agent's voice at once, and the agent answers the new question.

Phone support line

Callers on 8 kHz audio

A caller describing an order issue over a phone line gets the same turn detection as app users, with audio kept at 8 kHz end to end and PyAI telephony at $0.01 a minute.

Voice activity detection questions people ask

How fast does PyAI detect the end of a turn?

PyAI knows a user has finished about 196 ms, or 0.2 seconds, after they stop speaking. PyAI measured this in its US region on 24 June 2026. The agent's first audio follows about 390 ms after the user stops, measured on 8 kHz phone audio.

Can I adjust how long PyAI waits before ending a turn?

Yes. If you use the API, you can change how long PyAI waits before ending a turn, even in the middle of a live session. A longer wait gives slower or older speakers more time to finish. A shorter wait makes quick back-and-forth conversations feel faster.

What happens when a user interrupts the agent?

PyAI notices when the user starts talking while the agent is speaking. PyAI then sends your app a flush signal to stop the agent's voice straight away, whether the agent runs in an app, a browser or on a phone line. The agent then answers what the user just said instead of finishing its old sentence.

Does PyAI offer a standalone VAD API?

No. There is no standalone PyAI VAD product or API. Speech detection comes built into PyAI live transcription and Omni voice agents, where it works together with speech-to-text, turn detection and interruption handling, so you do not have to connect those parts yourself.

PyAI is voice infrastructure. It hands you the audio and transcript, and does not replace your own models or review tools.

Pause, read a number, interrupt. Test it free in your browser.

Start building with $50 creditSandbox keys never bill, and you can start without a credit card.

One stack: Omni, Hear, Speak, Recap and Trace.