Model-assisted event review

Find more candidate sounds while keeping the analyst in control.

Search recordings for time-frequency events, classify candidates with a project-relevant model and review every proposed annotation against the spectrogram and audio.

Analyze · Reviewed auto-annotation
Analyze acoustic detection model trainer and reviewed classification workflow

Detection with evidence

Move from candidate event to reviewed record.

Automated detection is most useful when its result can be checked quickly. Analyze keeps the proposed event, model output, source recording and analyst decision connected.

Propose candidate regions

Identify connected time-frequency features against the local spectrogram background and place candidate boxes around events for review.

Apply the active classifier

Display the predicted class, confidence, window agreement and strongest alternatives from the selected acoustic model.

Make the authoritative decision

Listen, inspect and accept, relabel or reject each candidate. The reviewed result—not the initial prediction—becomes the project record.

Detection applications

Adapt the workflow to the sounds that matter.

Marine mammal calls

Find candidate calls, whistles or clicks using models trained for the project and recording conditions.

Vessel and machinery sound

Review recurring tonal, broadband or transient events across deployment recordings.

Impulsive events

Locate short-duration acoustic events and retain their time-frequency evidence.

Batch verification

Test independent labelled files and surface mismatches or ambiguous predictions before operational use.

Reviewed datasets

Return accepted events and useful background examples to organised sound-class folders.

Traceable decisions

Retain model output, confidence, source times and the analyst's final classification.

Responsible use

Automation supports review; it does not replace it.

Detection and classification quality depends on representative labelled material, recorder response, background conditions and the target sounds. Models should be verified on independent recordings from the intended operating environment, and important results should remain subject to analyst review.

Bring representative target and background sounds.

We can demonstrate the detection and review loop using classes and recording conditions relevant to your project.

Plan a detection workflow