AI Speak DogDecoding Dog Emotions
Get Started

Measured, not marketed

How accurate is it, really?

No mind-reading — measurable acoustic and visual signal, reported with its dataset, sample size, and honest caveats. Where a signal is weak or unproven, we label it plainly.

Is it a bark?

Measured

99.7% recall · 99.3% specificity

Held-out on 979 real breed barks + 300 ESC-50 non-bark negatives. Small field set (n=34): 100% / 100%. This is the part we trust most.

Arousal (energy / how worked-up)

Weak but real

R² ≈ 0.19 · 40% vs 34% chance

Supervised head, held-out Barkopedia n=192 (husky/shiba scope). Weak-but-real and validated — it leads the emotion read. Breed-stratified it's uneven (shiba R² 0.27, husky 0.12), so it's not yet proven across breeds — broadening is data-gated.

Valence from audio (positive ↔ negative)

Unvalidated

At chance

Every audio model we tried fails to beat chance on valence — a bark encodes energy, not positivity. We say so, and we don't oversell it.

Valence from a photo/video

Measured

AUC 0.930 · 84% accuracy

A CLIP vision head reads body-language valence (held-out on the Dewa dog-emotion set). Multi-frame video 0.907; robust to real webcam capture (−0.006). This is how we solve the axis audio can't.

“Is my dog off today?” (vs their own baseline)

Measured

≈1% false-alarm rate

Rather than an absolute emotion, we track how a dog's own voice moves against its own 3-week robust baseline (median/MAD). An off-day needs several acoustic features to shift together, tuned for a low false-positive rate — measured at ≈1% per day on stable synthetic data (target 5%). A prompt to look, never a diagnosis.

“Which breed does this bark sound like?”

For fun

33% top-1 (chance 20%)

Better than chance, but deliberately for entertainment — shown as a playful spread, never a confident claim.

How we keep ourselves honest

  • Frozen evaluation. Every number is on data the model never trained on, with the test set hash-excluded from training to prevent leakage.
  • Floor tests in CI. A shipped model that regresses below its measured floor fails the build — it can't quietly get worse.
  • We label the weak parts. Valence from audio is at chance; we show it as an estimate and solve it with vision instead of pretending.
  • Your corrections train it. Every thumbs-down + “what was it actually?” becomes a labelled example that improves the next model.

See the read on your own dog

Free, no card required.

Get Started
AI Speak Dog — Decoding Dog Emotions

AI acoustic analysis that decodes your dog's emotions in real time. Built on peer-reviewed canine bioacoustics.

Product

  • Features
  • Emotions
  • Science
  • How it works
  • FAQ

Company

  • Book a Demo
  • Privacy
  • Terms

© 2026 AI Speak Dog. All rights reserved.

Made for dogs and the people who love them.