Skip to main content
Voice Activity Detection (VAD) examines your recording and removes the silence. HyperWhisper then sends the shorter audio to your transcription provider. Shorter audio costs less, returns results faster, and in many cases gives better accuracy.

​
Overview

When you stop a recording, HyperWhisper finds the parts of the audio that contain speech. HyperWhisper removes the silence and makes a trimmed copy of your audio. The transcription provider receives this trimmed copy, not the full original. VAD is enabled by default on macOS and on Linux. You can turn it off in Settings. The thresholds and the file behavior are different on each platform — see Platform Support.

​
How It Works

VAD uses the Silero VAD model. The macOS app includes this model (~864 KB). All processing occurs on your device. VAD runs when both of these conditions are true:
  • VAD is on in Settings.
  • The recording is 30 seconds long or longer.
HyperWhisper sends recordings shorter than 30 seconds to the provider without a change. These recordings are too short for silence removal to give a benefit. Processing flow:
1

Detect speech segments

The Silero VAD model examines your audio. It finds the parts that contain speech.
2

Remove silence

HyperWhisper joins all of the speech segments into one trimmed audio file. It removes each gap between segments that is longer than 200 ms. This includes the silence at the start, the silence at the end, and the pauses in the middle.
3

Validate the result

HyperWhisper makes sure that the trimmed file passes the minimum quality checks (see Validation & Quality Checks below). If the file fails a check, HyperWhisper uses the original audio.
4

Prepare for upload

If the trimmed file is 25 MB or larger, HyperWhisper converts it to M4A to make the upload smaller. HyperWhisper sends files smaller than 25 MB as WAV. Imported files that are already in a compressed format (M4A, MP3, and other compressed formats) are not encoded again, because a second encode lowers the quality.
HyperWhisper always keeps the original audio on disk. It stores the trimmed copy next to the original and uses the trimmed copy only for transcription and playback. HyperWhisper never writes over the original.

​
Benefits

  • Lower API costs — you pay for less audio when HyperWhisper removes the silence.
  • Faster transcription — smaller files upload and process faster.
  • Better accuracy in many cases — the model gets less silence and less background noise.
  • Good for any recording with pauses — interviews, dictation with thinking gaps, and recordings that you start early and stop late.

​
Enabling VAD

  1. Open Settings (click the menu bar icon → Settings).
  2. Go to the Sound section.
  3. Turn on Remove silence before transcription.
The toggle takes effect immediately. If your next recording is 30 seconds long or longer, HyperWhisper trims it with VAD.

​
Viewing Original vs. Trimmed Audio

If HyperWhisper trimmed a recording, the History detail view shows an Original / Trimmed toggle above the audio player. Select Trimmed to hear the audio without the silence. Select Original to hear the full recording.HyperWhisper keeps both files on disk. The toggle changes only which file plays. It deletes nothing.For more about the audio player, see Viewing Transcription History.

​
Validation & Quality Checks

After the trim, HyperWhisper examines the result before it uses the result. If the trimmed file fails one check or more, HyperWhisper uses the original audio for the transcription. You do not see an error.
HyperWhisper sends VAD failures to Sentry as breadcrumbs. These breadcrumbs help to diagnose rare faults. The original audio is always the fallback, so a VAD failure never stops the transcription.

​
Platform Support