Audio model lab

Runs on your device. Audio stays in your browser. Choose PCM or float WAV, up to 2 minutes / 64 MB. The first comparison downloads model weights. Per-window clipping and low-frequency flags are available in the Python app; this browser comparison does not measure them. These are experimental detectors; development results do not establish field reliability. Grad-CAM shows sensitivity of a selected CNN logit, not proof of a useful acoustic cue. Read the formulation and training diagnosis.

Audio model lab

Compare neural and non-neural scores. Inspect the spectrogram and the CNN's Grad-CAM for the same audio context.

Work through real examples

Publisher classes describe recordings; audible event boundaries have not been independently verified.

PCM or float WAV · up to 64 MB and 2 minutes. Inference runs on your device; your audio is never uploaded. Processing speed depends on your device.

Solid lines: scores. Dashed lines: each model’s alarm-on threshold. Scores are uncalibrated and cannot be compared as probabilities. Each alarm uses its own development-selected threshold and confirmation rule.

Model	Score at playback time	Alarm / input	Quality flags (Python app only)

Inspect a CNN decision

Model Context Target Layer

Frequency from Hzto Hz

Model input spectrogram

Grad-CAM overlay

The horizontal axis is time within the selected context; low frequencies are at the bottom. The vertical axis follows the model's linear or mel bin locations. Warm overlay colours mark positive attribution to the selected logit. Maps are normalized independently; brightness is relative within a map. A zero map is reported rather than filled with invented detail. Grad-CAM is computed from gradients of the trained CNN, at the selected layer's native resolution; interpolation adds no explanatory detail. Attribution can reveal shortcuts but does not prove a useful or causal acoustic cue.

Training history for selected CNN

Green: balanced training objective including regularization. Orange: balanced development bag loss, checked every two epochs. White dashed line: selected checkpoint. The two losses have different sampling/regularization terms; values are not accuracies.

Training and development evidence

Model	Method	Input tensor / descriptor	Development positives detected	Negative files alarmed	Selected epoch	Contexts seen at checkpoint

Frequency-axis warping, log intensity and tensor shape are separate choices. These models use the same real training/development protocol. Counts describe correlated publisher-labeled files. The previously assessed default is preserved; this lab supplies development evidence.

Listening notes

Listen and inspect a specific interval. Save an audible-drone, background, or uncertain judgment with your reasons. These draft notes stay on your device and are not used for training.

Start (s) End (s)