EngineeringGlossary

Voice Activity Detection (VAD)

Voice Activity Detection (VAD) is software that detects when someone is speaking versus silent: the foundational layer beneath turn-taking, barge-in, and STT in every voice AI agent.

Humanlike Voice AI → Talk to an architect

Three things to remember

VAD answers one question: “is there speech in this audio chunk right now?”. The answer feeds turn-taking and barge-in.
Modern VAD is neural (Silero-class) and runs in 10-30ms windows. Old VAD was energy-based (just looking at audio volume) and failed on background noise.
EnableX uses a tuned neural VAD that's robust to call-centre background noise, music-on-hold leakage, and IVR tones.

The detailed explanation

VAD is the simplest layer in a voice AI stack, and the most important. It just answers “is the human speaking right now?” every 10-30ms. That binary signal is what powers turn-taking (the agent knows the customer started speaking → don't start responding yet) and barge-in (the agent knows the customer started speaking → stop the TTS playback).

Bad VAD breaks the agent in two ways: false positives (says “speech” when it's background noise → agent thinks customer is talking, gets confused) and false negatives (says “silence” when customer is whispering or speaking quietly → agent doesn't hear them). Old VAD models (energy-based, just checking audio volume) failed both ways constantly. Modern neural VAD models (Silero, MarbleNet) are dramatically better, especially in noisy environments.

The deployment tradeoff: aggressive VAD (low threshold) catches every utterance but false-positives on noise → bot interrupts itself. Conservative VAD (high threshold) is robust to noise but misses quiet speech → customer talks, bot doesn't hear. EnableX exposes per-deployment tuning, with sane defaults for cloud, contact centre, and noisy outbound environments.

How EnableX implements Voice Activity Detection (VAD)

EnableX Voice AI uses tuned neural VAD optimised for noisy contact-centre environments, music-on-hold leakage, and IVR tones common in Indian / SEA / ME telephony. Per-deployment tuning available for high-noise scenarios. Powers both turn-taking (sub-500ms response after customer stops) and barge-in (sub-200ms detection when customer starts).

Humanlike Voice AI → Build with EnableX Pro Services →

Related glossary terms

Turn-Taking

Turn-taking is how a voice AI agent decides when the customer is done speaking and when it should respond.

Barge-In

Barge-in is the ability for a customer to interrupt the voice AI mid-sentence.

Speech-to-Text (STT / ASR)

Speech-to-text (STT), also called Automatic Speech Recognition (ASR), is software that converts spoken audio into written text.

AI Voice Agent

An AI voice agent is software that holds a real-time spoken conversation with a customer over the phone or any voice channel.

← Back to glossary

Where to dive deeper

Telephony → On-Premise → vs Vapi →

FAQ — Voice Activity Detection (VAD)

VAD detects when someone is speaking. It's the foundational layer beneath turn-taking and barge-in: without good VAD, the bot either interrupts the customer (false positive on noise) or sits in silence (false negative on quiet speech). Modern neural VAD is dramatically better than the old energy-based approach.

No. STT transcribes speech into text. VAD just detects whether speech exists. STT is heavyweight (typically runs on a GPU); VAD is lightweight (runs on CPU in 10-30ms windows). VAD typically runs first; STT only processes audio when VAD says "yes, there's speech here".

Background noise (other voices, music, machinery) can fool a naive VAD into reporting speech where there is none. Modern neural VAD (Silero-class) handles this dramatically better than old energy-based VAD, but tuning thresholds per deployment is still important for high-noise scenarios.

Yes. EnableX exposes per-deployment VAD tuning: aggressive for clean cloud audio, conservative for noisy contact centres or street-noise outbound. The defaults work for most use cases out of the box.

See EnableX in action.

Talk to sales, or start a free trial. No credit card required.

Try for free Talk to sales

Free trial credits · No credit card · API keys in 2 minutes