What Is a VAD? Voice Activity Detection Explained

If you have ever wondered how a dictation app knows when you are actually talking versus just breathing, tapping keys, or sitting in a noisy cafe, the answer is a small piece of technology called a VAD. It is one of the quietest but most important parts of any modern voice to text mac setup.

What VAD actually means

VAD stands for Voice Activity Detection. It is a small algorithm, and these days usually a tiny neural network, that answers one question over and over: is someone speaking right now? It chops incoming audio into short chunks, often just twenty or thirty milliseconds each, and labels every chunk as either speech or not speech.

That sounds simple, but it is doing a surprising amount of work. A VAD has to tell the difference between your voice and a fan, a keyboard, street traffic, a passing siren, or the hum of an air conditioner. Older systems used raw signal energy, meaning anything loud enough counted as speech. Modern VAD models are trained on real recordings, so they recognize the texture of human speech rather than just loudness. That is why they still work when you talk softly in a loud room.

Why dictation software needs a VAD

Speech to text models like Whisper and Parakeet are powerful, but they are also relatively heavy. Running them on every second of audio, including long silences, wastes processing power and can produce strange output. If you feed pure silence or background hiss into a transcription model, it sometimes hallucinates words that were never said. A VAD prevents that by acting as a gate: only the chunks that actually contain speech get passed downstream.

The VAD also solves a subtler problem: knowing when you are done. When you pause at the end of a sentence, the VAD detects the trailing silence and signals that a phrase is complete. That is what lets a well built app finalize your text at natural moments instead of waiting awkwardly or cutting you off mid word. If you are curious how the wider process fits together, our guide to the voice first workflow for knowledge workers walks through where dictation slots into real work.

How a VAD fits into the pipeline

In a typical on-device dictation app, the audio flows through a short chain of steps. The microphone captures sound, the VAD filters it, the speech model transcribes what is left, and then optional AI cleanup polishes the result. Here is how those roles compare.

The reason this ordering matters is efficiency. The VAD is tiny and cheap to run, while the speech model is the expensive part. Putting the cheap gate first means the expensive stage does far less work. This is exactly how BlaBlaType keeps dictation feeling instant on a Mac, since every one of these stages runs locally on Apple Silicon. If you want to understand what makes that local hardware so quick, we cover it in how the Apple Neural Engine runs speech models.

The trade-offs of tuning a VAD

A VAD is not magic, and app makers have to tune how sensitive it is. Set the threshold too aggressive and it may clip the start of a quiet word. Set it too relaxed and it lets noise through. Here is an honest look at the upsides and the limits.

None of these are dealbreakers. They are simply the reason accuracy depends on more than just the VAD. The speech model, the microphone, and how the whole thing is measured all play a part. If you want to see how transcription accuracy is scored in the first place, read our explainer on word error rate and how accuracy is measured.

VAD, privacy, and running it locally

Because a modern VAD model is so small, there is no technical reason to send your audio to a server just to detect speech. The whole gate can run on your Mac in a fraction of a millisecond per chunk. BlaBlaType takes the same approach for the entire chain: the VAD, the Whisper and Parakeet speech models, and the AI cleanup powered by Apple Intelligence all run on-device, so your audio and transcripts never leave the Mac. If privacy is your reason for asking about VAD in the first place, our piece on whether Mac dictation is private goes deeper.

Not every voice tool works this way. Some accessibility and command systems, like the open source Talon Voice, take their own approach to local control, while cloud dictation services stream your audio elsewhere for processing. And the on-device AI that cleans up your text afterward, in BlaBlaType's case, is built on Apple Intelligence, which also runs locally on supported Macs. The pattern is the same across all of it: keep the small, fast steps like VAD on the device, and you get speed and privacy at once.