What Is Voice Activity Detection (VAD)?
Voice activity detection tells speech apart from silence and noise. Learn how VAD works, where codecs, AMD and AI voice agents use it, and why it can misfire.
Risposta rapida
Voice activity detection (VAD) decides whether a slice of audio is speech or just silence and noise. Codecs use it to stop sending during silence and add comfort noise. Voice agents use similar detection to know when a caller has finished or interrupted. RFC 3389 leaves the algorithms implementation-specific.
Voice activity detection, or VAD, answers a simple question many times a second: is somebody speaking right now, or is this silence and background noise? Many voice technologies depend on the answer, and when VAD gets it wrong the symptoms are easy to hear.
What Voice Activity Detection Is
VAD is a decision step that classifies a slice of audio as speech or not speech. It does not understand words. It only marks where speech seems to start and stop. The method varies between implementations, from simple energy thresholds to trained models. RFC 3389 treats the VAD algorithm itself as unspecified and left implementation-specific, so two products that both say they have VAD may behave quite differently.
VAD in Codecs: Silence Suppression and Comfort Noise
In a phone call, each person is silent for much of the time. VAD lets a sender stop transmitting during silence to save bandwidth, a technique known as discontinuous transmission, or DTX. The problem is that complete silence sounds like a dead line, so the receiver plays a faint, artificial background sound instead: comfort noise.
RFC 3389 defines an RTP payload format for transporting comfort noise, mainly for audio codecs without built-in comfort noise, such as G.711 and G.722. It states that comfort noise analysis and synthesis, along with the VAD and DTX algorithms, are unspecified and left implementation-specific, and gives an example for G.711 that uses the VAD and DTX of G.729 Annex B. The ITU lists that annex as a silence compression scheme for G.729, dated 10/96, and shows it as superseded. For the codecs themselves, see the voice codec page.
VAD in Answering-Machine Detection
Detecting an answering machine involves listening to how the answered call begins. A person says a short greeting and then waits. A recorded greeting runs on for longer. Telling these apart means knowing when speech starts and stops, which is VAD-style information, along with other signals. The rilevamento della segreteria telefonica page covers the methods in detail, and the AI AMD page describes how DialerBee approaches it. This page does not claim that any specific DialerBee component uses VAD, and no accuracy figure is implied.
VAD in AI Voice Agents: Knowing When the Caller Has Stopped
A voice agent has to decide when the caller has finished their sentence before it answers. Answer too soon and it cuts the caller off. Wait too long and the pause feels unnatural, which ties directly to latency. The same detection also lets an agent notice when the caller starts speaking over it, which is the basis of irruzione. The agente vocale basato sull'intelligenza artificiale page describes the wider picture; the product page states that the agent handles barge-in. The site does not describe which VAD method it uses, so none is claimed here.
Why VAD Misfires
- Noise. Traffic, a TV or a crowded office can be mistaken for speech, or can mask quiet speech.
- Short replies. A one-word "yes" or "no" is easy to miss or to clip, especially at the start.
- Soft starts and trailing words. The first and last syllables are the quietest, so detection can clip them.
- Pauses mid-sentence. A caller who stops to think can be judged to have finished.
How to Use It
- Test with real conditions. Noisy lines and mobile calls expose VAD weaknesses that quiet lab calls do not.
- Listen for clipped starts and early cut-ins. Those are the two classic symptoms.
- Treat vendor claims carefully. Since the algorithm is implementation-specific, ask how it performs on your calls.
Domande frequenti
What is voice activity detection?
A step that decides whether an audio segment contains speech or only silence and noise.
What is the difference between VAD and DTX?
VAD detects speech. DTX, discontinuous transmission, uses that decision to stop sending packets during silence.
What is comfort noise?
A faint artificial background sound played during silence so the line does not sound dead. RFC 3389 defines how to carry it over RTP.
Is the VAD algorithm standardised?
RFC 3389 says the VAD and DTX algorithms are unspecified and left implementation-specific, so implementations differ.
Why does a voice agent sometimes cut me off?
It may have judged a pause to be the end of your turn. Detecting turn ends is a core part of how voice agents manage conversation.
Related terms and guides: rilevamento della segreteria telefonica, voice codec, agente vocale basato sull'intelligenza artificiale, irruzione, VoIP latency, AI AMD.
Fonti
- RFC 3389, RTP Payload for Comfort Noise (introduction: VAD and DTX left implementation-specific; G.729 Annex B example) (source, checked November 7, 2026)
- ITU-T G.729 page (Annex B, 10/96, silence compression scheme; listed as superseded) (source, checked November 7, 2026)
Termini correlati
What Is VoIP QoS? Quality of Service
VoIP QoS (quality of service) gives voice packets priority on a network so calls avoid delay, jitter and loss. Learn DiffServ, DSCP EF 46 and where to use it.
What Is VoIP? Voice over IP Explained
VoIP (Voice over IP) carries phone calls as data packets over IP networks instead of copper circuits. How it works, SIP and RTP, and what shapes call quality.
What Is Word Error Rate (WER)?
Word error rate (WER) scores speech recognition by comparing its output to a correct transcript and counting substitutions, deletions and insertions.
Che cos'è un contact center BPO?
Un contact center BPO gestisce le chiamate per conto di aziende clienti. Il modello richiede isolamento tenant, regole per cliente, branding e reporting.
What Is a Call Detail Record (CDR)?
A call detail record logs who called whom, when, how long and how the call ended. See what a CDR holds, how teams use it, and how it differs from a recording.
Che cos'è una classificazione di chiamata?
L'esito di chiamata è il risultato codificato registrato a fine contatto. Scopri come si progettano i codici, perché variano e cosa dipende da essi.
Pronti a vedere DialerBee in azione?
Richiedi una demo dal vivo di 15 minuti oppure avvia una prova gratuita di 14 giorni: niente slide, nessun impegno. Le chiamate partono quando il tuo trunk SIP o PBX è collegato e approvato.
Prova gratuita di 14 giorni · credito IA gratuito · senza carta · 11 lingue · BYOC · controlli a supporto della conformità