A close photograph of a microphone grille in soft light

Technology & privacy

Voice activity detection, and why your first word goes missing

A small model runs before the speech model and decides which parts of the audio are speech. It saves time and battery, and it is responsible for the most common dictation complaint.

Last updated

What it does, and why it exists

When you hold a dictation shortcut, the app captures everything the microphone hears — including the half-second before you start, the pauses while you think, and the moment after you finish but before you release the key.

Handing all of that to a speech model is wasteful in three ways. It costs time you notice as latency. It costs battery. And it costs accuracy, because a model asked to transcribe a fan and a distant conversation will sometimes produce words from them.

VAD solves this by being small and cheap. Silero VAD, which several dictation apps use including ours, is a few megabytes and runs in a fraction of the time the speech model takes. It marks the speech regions; the recogniser only sees those.

Voice activity detectionSpeech recognition
Question it answersIs this speech?What words are these?
Model sizeA few megabytesHundreds of megabytes to gigabytes
SpeedEffectively instantThe slow part
OutputTimestampsText
RunsFirst, on everythingSecond, on the trimmed audio only

Why the first word goes missing

This is the most useful practical consequence and it explains a complaint that otherwise looks like a bug.

The detector needs a moment of audio to decide that speech has begun. If you press the shortcut and start talking in the same instant, the beginning of your first word arrives while it is still deciding, and part of it can be trimmed with the preceding silence.

The same happens at the end, in a slightly different way: people drop their volume on the last few words of a sentence, the detector reads the quiet tail as silence, and "...before the meeting" becomes "...before the".

The trade-off VAD is tuned against

Every detector sits on a dial between two failure modes, and moving away from one moves towards the other.

SettingFailure mode
Too eager (marks more as speech)Room noise, keyboard clicks and a colleague's conversation get transcribed. Stray words appear in your text
Too strict (marks less as speech)Quiet starts, soft speakers and the ends of sentences get trimmed away

There is no setting that is right for every room and every voice, which is why an app's default is a judgement about its typical user rather than a correct answer. Softly spoken people in quiet rooms and loud people in busy offices genuinely want different tuning.

Where else VAD shows up

  • Latched dictation. When the microphone is held open rather than push-to-talk, VAD decides where one utterance ends and the next begins.
  • Video calls. The same class of model drives the speaking indicator and noise suppression.
  • Voice assistants. Deciding when you have finished talking is a VAD problem, and it is why an assistant sometimes cuts you off mid-sentence.
  • Transcription tools. Chunking a long recording into segments before transcribing it.

Troubleshooting VAD-shaped problems

SymptomLikely causeFix
First word missing or clippedSpeaking as you pressHalf a second of pause first
Last few words missingTrailing off in volumeKeep the level up to the end; release a beat after you finish
Words appear that you did not sayBackground speech being marked as yoursCloser microphone, or a quieter room
Long pauses split a sentence in twoThe detector read the thinking pause as an endingSpeak in complete phrases; think before you press
Nothing is transcribed at allWrong input device, so the detector hears nothingSystem Settings → Sound → Input

Questions

What is VAD short for?

Voice activity detection. It is a small model that identifies which parts of an audio stream contain speech, running before the speech recogniser so the expensive model only sees the parts that matter.

Why does my dictation cut off the first word?

Almost always because you started speaking as you pressed the shortcut. The detector needs a moment of audio to decide speech has begun, and part of the first word can be trimmed with the leading silence. Pause half a second before speaking.

Does VAD mean my microphone is always listening?

Not in a push-to-talk app. The microphone is live only while you hold the shortcut, and VAD runs on that captured audio. That is a different design from an always-listening assistant waiting for a wake word.

Can I turn VAD off?

Most dictation apps do not expose it, and turning it off would make things worse — the speech model would be asked to transcribe silence and room noise, which is slower and produces stray words.