Acoustic intelligence
No magic, no mind-reading — measurable acoustic signal, honest about what it can and can't tell you.
A CLAP audio encoder decides whether a clip is genuinely a dog bark, recalibrated on real field recordings so it works across breeds and noisy rooms.
Pitch, bark rate, harmonics-to-noise ratio, spectral shape and more — the acoustic dimensions that decades of bioacoustics research tie to emotional state.
A supervised arousal model — trained on labelled husky and shiba recordings — reads energy/arousal, combined with breed-relative acoustic priors for the final state.
Bark detection is measured — 99.7% recall / 99.3% specificity on real recordings. Reliably tells a bark from background noise.
Arousal is weak-but-real — R² ≈ 0.19 on held-out husky/shiba recordings (40% vs 34% chance). A genuine, validated energy signal.
Valence is the hard part — at chance from audio, so we read it from a photo/video instead (AUC 0.930). We show the estimate, don't oversell it, and your feedback improves it.
Every number, with its dataset and caveats, lives on the accuracy & methodology page.
Built on peer-reviewed research