Record your sandpapers
Record P60
ReadySlow, long strokes
Rub steadily across the sandpaper for 20 seconds.
20 seconds
Documents / Bela Recordings
Latest recording
All recordings →Recordings
Listen, refine labels, and choose the takes for training.
Copy original recordings and label history to Dev.
Train a surface model
Train now. Use fresh passes for validation and testing.
Dataset readiness
— eligible takesWhole sessions are held out for validation and testing.
Test each new finger or mounting on reference surfaces.
Fine-tune pretrained BAT
Adapt a pretrained audio model to your material and grit labels on a Modal GPU.
Modal · L40S · up to 2 h
Model details
BAT starts with pretrained audio weights. Fine-tuning updates the final two transformer blocks and the label classifier in PyTorch.
Analysis uses 16 kHz audio, covering frequencies below 8 kHz. Original WAV files retain the full recording.
Current output: your material and grit labels. Quantitative roughness needs measured reference values; spatial maps need contact position.
Test results
One-second windows from held-out recordings
Predictions by material
Window counts · rows: actual · columns: predictedSignal map
Each point is one second of recorded contact.
Arrow keys select a point. Enter plays its one-second audio segment.
Nearby points have similar features. This projection is a visual summary.
Live surface test
Rub a surface to see what the model recognizes.
Train the first model to test live.
Up to 3 minutes per test.
Training →