Meow Decoder classifier heads
Small logistic-regression heads that turn a meow's audio embedding into the likeliest situation: waiting for food, enjoying attention, or stressed somewhere new. Used by meow-decoder, which downloads them automatically.
| File | Encoder | First guess right (unseen cats) | Top two | Macro recall | Calibration error | License |
|---|---|---|---|---|---|---|
beats.npz |
Earth Species Project esp_aves2_sl_beats_all |
65.9% (57β74%) | 89.1% | 58.1% | 2.2% | CC BY-NC-SA 4.0 (follows the encoder) |
ast.npz |
MIT AST fine-tuned on AudioSet | 61.8% (52β72%) | 84.1% | 53.2% | 4.7% | CC BY 4.0 (follows the CatMeows data) |
Always guessing the most common situation scores 50.2% (macro recall 33%). Scores are 5-fold cross-validation over
whole cats (splits.json in the code repo), so every test meow comes from a cat the head never saw. The range is a
95% interval from resampling cats: with 21 cats, the benchmark is small.
These scores don't carry over to other recordings. On 58 meows from Wikimedia Commons and Freesound whose
uploaders said what was happening (45 uploaders), beats.npz got the first guess right 14% of the time, and called
2 of 48 food-begging meows "food". See the code repo's README, "Tested on real recordings from the internet".
Since v0.5.0 the app doesn't use these heads for cats nobody has taught: it guesses from the meow's pitch contour
(pitch.npz, shipped in the package), which got 45% on 55 real meows. These heads are used once a cat is taught,
refitted with that cat's own meows.
Re-run the benchmark from these files alone, in seconds, with no audio or audio models:
meow benchmark --from-weights
Format
Plain NumPy arrays, no pickle: classes, mean, scale, coef, intercept (standardize, then multinomial
logistic regression), plus X, y (the 440 training embeddings, kept so the app can refit with your own cat's meows),
cats (which cat each embedding came from) and train_only (marks augmented copies; none in this release).
from meow_decoder import Decoder
Decoder("beats").predict("meow.wav")["results"]
Training data
CatMeows (Ludovico et al., 2020): 440 meows from 21 Maine Coon and European Shorthair cats, CC BY 4.0. Audio is 8 kHz, so nothing above 4 kHz was heard.
Limits
Three situations only. It cannot detect pain or illness. Pet cats from one dataset; other cats may sound different.