Engine Sound Classifier

A fine-tuned PANNs CNN14 (pretrained on AudioSet) that classifies short engine-audio clips into one of 34 configurations — cylinder count, layout, and stroke type (e.g. i4, v8_cross, single_two_stroke, 2_rotor).

Early checkpoint · ~0.33 balanced accuracy · 34 classes

This is a static, informational Space — it doesn't run the model. It exists so you can hear a few labeled example clips and see how the model is doing without needing to spin up compute. To actually run inference, download the checkpoint from the model repo linked below.

Example clips

i2_360
v2_45
single_two_stroke

Performance by run

Run A
balanced accuracy not tracked this run (micro 0.380, overfit)
Run B
0.297 balanced · 0.311 micro
0.297
Run C
0.323 balanced · 0.343 micro
0.323
Run D
0.330 balanced · 0.376 micro (best)
0.330
0.0 0.2 0.4
Recording-level balanced accuracy is the headline metric; random guessing over 34 classes is ~0.03.
RunChangeBalanced accMicro acc
Abaseline (no regularization)0.380 (overfit)
B+ SpecAugment / noise / mixup0.2970.311 (underfit)
Cfixed SpecAugment time-mask scale0.3230.343
D+ softened class weights (counts⁻⁰·⁵)0.3300.376

Known weak spots

Links