Voice activity detection Deciding, moment by moment, whether anyone is speaking.
Short answer. Voice activity detection (VAD) decides, moment by moment, whether an audio signal contains human speech or only silence and noise. Speech systems use it to cut recordings into speech segments, to skip silence and to decide when a speaker has finished. Silero VAD, an open-source detector under the MIT licence, reports that a chunk of 30 milliseconds or more takes under 1 millisecond on one CPU thread and that its model is about two megabytes. VAD matters because recognisers can invent words in silence: OpenAI’s Whisper paper says its own no-speech signal alone was not enough to detect empty segments.
01
What VAD does
VAD answers one question: is someone speaking right now? It does not say which words, which language or which person. It is a different job from wake-word detection, which listens for one particular word to wake a device, as Jurafsky and Martin describe it.
Speech tools use it in three ways. Before transcription it removes empty audio. Before speaker diarization it is the first step: pyannote’s documentation lists speech activity detection as the stage that gets rid of all non-speech regions. In a live voice agent it helps decide that the caller has finished talking.
02
How it is measured
pyannote’s metrics library defines the detection error rate as false alarms plus missed detections, divided by the total speech in the reference. Its detection cost function weights misses at 0.75 and false alarms at 0.25 by default, so missed speech counts three times as much.
Error
What it means
What you notice in an office
False alarm
Non-speech is treated as speech
A fan, typing or a door is transcribed as words
Missed detection
Speech is treated as non-speech
Clipped first words, or a quiet speaker lost
03
Where VAD sits in a pipeline
Stage
What VAD does
Example
Before transcription
Removes silence so the recogniser is not fed empty audio
faster-whisper’s vad_filter uses Silero VAD; by default it removes only silence longer than 2 seconds
Inside the recogniser
Whisper was trained on segments with no speech and predicts a no-speech token
The paper combines a no-speech threshold of 0.6 with an average log-probability threshold of −1
Before diarization
Finds the stretches of speech to be split by speaker
The first stage in pyannote’s pipeline
In a live voice agent
Helps decide when the caller has stopped
Part of the delay before a reply starts
04
What a small open VAD looks like
Silero VAD’s README says its model is around two megabytes, that a chunk of 30 ms or more takes under 1 ms to process on a single CPU thread, and that it supports 8,000 Hz and 16,000 Hz audio, so telephone recordings work. It says it was trained on corpora covering over 6,000 languages and is published under the MIT licence with no telemetry, keys or registration. The stated requirements are Python 3.8 or later, 1 GB of RAM and a modern CPU.
05
What to check
Tune the threshold on your own room: too strict clips short words such as “haan” and “nahin”, too loose passes fan noise.
Look for hallucinated text in silent stretches of a transcript. Whisper’s model card warns that predictions may include text that was not spoken.
Check the sampling rate your recorder and phone line produce against what the detector supports.
Test with Hindi and with overlapping speech, not only quiet English.
Do not treat VAD as language or speaker identification; it only separates speech from non-speech.
06
Where ANVE stands
ANVE comes from the makers of AnveVoice, a voice AI for websites whose replies start a median 487 milliseconds after a caller stops speaking, measured end to end. Deciding that a caller has stopped is the VAD part of that problem. A voice front desk that answers callers in Hindi or English is one of the agents ANVE builds with a paid pilot, on AnveVoice; it is not a shipped product. The sealed unit, which is in design, adds a physical switch with a light that cuts the microphones.
07
Questions people ask
Is VAD the same as speech recognition?
No. VAD only says whether speech is present. Speech recognition works out which words were said.
Why does my transcript contain words in silent parts?
Recognisers can hallucinate when fed silence. Whisper’s model card warns that predictions may include text not actually spoken. Filtering silence with a VAD before transcription helps.
Does VAD work for Hindi?
VAD detects speech, not the language. Silero says its model was trained on corpora covering over 6,000 languages, but test it on your own recordings and room noise.
Can VAD run on a CPU?
Yes. Silero says a chunk of 30 ms or more takes under 1 ms on a single CPU thread.
08
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
