Compare neural and non-neural scores. Inspect the spectrogram and the CNN's Grad-CAM for the same audio context.
Loading models…
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.
Playback keeps the recording's original level. Model preprocessing differs: the spectrogram CNNs/SVM use a per-context peak reference; the YAMNet control has no shared waveform RMS target. A level change inside a context can remain visible after peak normalization.
Results remain loaded when switching recordings. Each cell shows whether an alarm occurred, the highest score and the score timeline; dashed lines show that model's threshold. Scores are not calibrated probabilities.
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) |
|---|
Choose any trained spectrogram CNN here, independently of the score comparison above. Grad-CAM is available for these eight CNNs.
Drag a rectangle in either plot to zoom both. Mouse wheel zooms at the cursor. Shift-drag pans. Double-click resets.
Hover over either plot to inspect time, frequency and values.
The horizontal axis is time within the selected context; low frequencies are at the bottom. Both plots use a shared linear Hz display; values from the model's original linear or mel bins are projected onto that axis without adding spectral detail. 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.
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.
Thresholds were selected on this development set under a 5% negative-file budget: nine alarmed files out of 188. The matching false-positive counts largely reflect calibration, not independent validation or identical mistakes. Recall is compared at this budget; false-alarm rates on new recordings can differ.
| Model | Method | Input tensor / descriptor | Development positives detected | Development negatives 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.
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.