Audio Analysis
VAD (Voice Activity Detection)
Finds where someone is speaking (or where there is loud sound) and drops the silence and background noise: one row and clip per segment, one clip per file with the gaps cut out, or just the list of segments.
When to use it
Before transcription or speaker analysis, to skip dead air, hold music and room noise — and to cut long recordings at natural pauses.
In a pipeline
Audio File InputVAD (Voice Activity Detection)ASR (Speech-to-Text)Metadata Joiner
Settings
| Setting | Type | Required | Default or hint |
|---|---|---|---|
| Detect | select | — | speech |
| Results | select | — | segments |
| Above background by (dB) | number | — | 12 |
| Shortest segment (ms) | number | — | 250 |
| Join gaps shorter than (ms) | number | — | 300 |
| Padding either side (ms) | number | — | 100 |
Example
A voicemail split into its spoken parts
- Detect=Human speech
- Results=One row (and clip) per segment
| File Name | Segment | Start (s) | End (s) |
|---|---|---|---|
| vm-5521.wav | 1 | 0.62 | 7.9 |
| vm-5521.wav | 2 | 9.1 | 18.35 |
Worth knowing
- Lower Above background by to catch quiet speech; raise it in noisy rooms. Human speech also checks that the energy is in the voice band, so a humming machine is not mistaken for talking.
- Start (s) and End (s) are times in the original file, even after a Slicer.
- Clips with no speech are dropped (and counted in the node's message), except with Keep the audio.