Voice activity detection (VAD) is a signal-processing technique that detects when human speech is present in an audio stream and when it is not. It works by continuously analyzing incoming audio and classifying each short slice — typically 10 to 30 milliseconds — as either "speech" or "non-speech" (silence, noise, or background sound). VAD is the invisible gatekeeper inside every modern voice application: it decides when to start listening, when a speaker has finished, and when to stop wasting compute on empty air. For revenue teams running AI phone agents, VAD quality is the difference between a call that feels human and one that talks over your lead.
How voice activity detection works
VAD works by extracting acoustic features from an audio signal and scoring them against a model that distinguishes speech from everything else.
Older VAD systems relied on simple thresholds: energy level, zero-crossing rate, and spectral characteristics. If the audio crossed a loudness threshold, it was flagged as speech. These are fast and cheap but fail badly in noisy environments — a slamming door or highway hum gets misread as talking.
Modern VAD uses machine-learning models (often lightweight neural networks) trained on thousands of hours of labeled speech. They tolerate background noise, accents, and cross-talk far better.
A typical VAD pipeline:
- Framing — split audio into 10–30 ms windows.
- Feature extraction — compute energy, frequency bands, or learned embeddings.
- Classification — score each frame as speech or non-speech.
- Smoothing — apply hangover logic so a brief pause mid-sentence isn't misread as "done."
Where VAD is used
VAD is used anywhere a system needs to react to human speech in real time.
Common applications:
- AI voice agents and IVR — knowing when a caller has finished speaking so the agent can respond.
- Speech-to-text (ASR) — trimming silence so only speech is transcribed, cutting cost and latency.
- Video conferencing — muting inactive mics and prioritizing active speakers.
- Voice assistants — detecting the start of a command after a wake word.
- Telecom — silence suppression to save bandwidth on VoIP calls.
In each case, VAD reduces the volume of audio the downstream system has to process, which lowers cost and speeds up responses.
Why VAD matters for AI calling and speed to lead
VAD matters because it determines whether an AI phone call feels natural or robotic — and natural conversations convert.
When an AI agent calls an inbound lead, it must detect the exact moment the person stops talking to reply without an awkward pause or, worse, interrupting them. Poor VAD creates the "talking over you" effect that makes callers hang up. Precise VAD lets the agent handle back-and-forth, pauses, and turn-taking like a human rep.
This is critical for speed to lead. Contacting a lead within 5 minutes makes them far more likely to qualify (MIT/Oldroyd Lead Response Management study), and contact within 1 minute drives dramatically higher conversion (Velocify). Since 30–40% of inbound leads commonly arrive after hours, an AI agent that calls instantly and converses smoothly captures demand human teams miss. Tools like Lead to Speed rely on tight VAD to call inbound leads in seconds and hold a real conversation — see the complete guide to speed to lead for how response time drives revenue.
VAD vs. related speech technologies
VAD is one layer in a stack of speech technologies, and it's often confused with the others.
| Technology | What it does | Best for | Limitation |
|---|---|---|---|
| Voice Activity Detection (VAD) | Detects presence/absence of speech | Turn-taking, silence removal | Doesn't identify words or speakers |
| Automatic Speech Recognition (ASR) | Converts speech to text | Transcription, commands | Needs clean speech input to be accurate |
| Speaker Diarization | Identifies who spoke when | Multi-speaker transcripts | Doesn't understand content |
| Wake-Word Detection | Spots a specific trigger phrase | "Hey" assistant activation | Narrow; only one phrase |
| Endpointing | Decides when a turn has ended | Conversational AI timing | Builds on VAD; can misfire on pauses |
Note: capabilities evolve quickly across vendors — verify current specs before choosing a platform.
VAD usually runs first, feeding cleaner, speech-only audio to ASR and endpointing so the rest of the pipeline is faster and cheaper.