External Model - Audio Understanding - Sound Events (Cochl)

What it does

Sound Events scans an audio file and lists every sound it recognizes, with a start time, an end time and a confidence score. It knows 100+ sound tags across 8 categories, such as Glass_break, Siren, Dog_bark and Applause.

The output is a segments list, the same shape Video Analysis produces. That means you can cut a video around the sounds it finds: every applause moment, every bark, every alarm.

It is about sounds, not words. For what people say, use Speech Analysis (Cochl).

Problem it solves

Input/Output

• Input: Sound — one audio file, up to 1 hour. MP3, WAV, FLAC and OGG are sent as-is; other formats are converted first. Starting from a video? Use Video Audio Extract first
• Output: Segments — a JSON list, one entry per event:
• start_sec / end_sec — when the event starts and ends
• span — the event's time range
• title — the Cochl tag, e.g. "Siren"
• score — the highest confidence seen during that event
• Requires: A Cochl project key, added under your avatar → Workspace Settings → Integrations

Settings

• cochl_min_confidence — default 0.5 (range 0.05–0.95). Lower it to catch faint or distant sounds, at the cost of more false hits. Raise it to keep only clear ones
• cochl_merge_gap_sec (advanced) — default 1.0 (range 0–5). Two runs of the same tag closer than this become one event. Raise it to treat an on-and-off sound, like barking or a stuttering alarm, as one event. 0 keeps every run separate

How to Use

1. Add your Cochl project key under Workspace Settings → Integrations.
2. Upload a recording with Sound Loader, or a video with Video Loader → Video Audio Extract.
3. Connect the audio to Sound Events (Cochl) and the output to Text Viewer to see the events.
4. To cut clips, connect the segments to Video Trim along with the original video.

Example Workflow

See what's in a recording: Sound Loader → Sound Events (Cochl) → Text Viewer

Cut highlight clips by sound: Video Loader → Video Audio Extract → Sound Events (Cochl) → Video Trim (with the original video) → Video Viewer

Choosing an audio understanding model

Performance

• Execution: Cochl cloud API; no local GPU
• Timing precision: about 1 second. Cochl listens in 2-second windows that move forward 1 second at a time
• Length: up to 1 hour of audio per run
• Billing: 0 CNAPS credits. Cochl bills your project per minute of audio
• Watch for: faint or distant sounds dropping out at the default confidence — lower cochl_min_confidence a little and check for false hits

Technical Details

Compliance & Provenance

For more on how we classify models and mark outputs, see our AI Policy.

Model Source

• Cochl.Sense

• Cloud API docs: Cochl.Sense Cloud API — Getting started