three decades of audio classifiers, same log-mel input
input: 10-sec log-mel spectrogram — shape [T=998, n_mels=80]
same frontend, three generations of backends
1990s — k-NN on MFCCs
MFCC-13 + mean/var pool
cosine → top-K vote
zero training, no GPU
strong on ESC-50
fails on AudioSet
ESC-50 acc: ~65%
your first baseline
2015–2021 — 2D CNN
log-mel as image
Conv + Pool × 3, softmax
3M params, 10-min train
SpecAugment + mixup
still the baseline in KWS
ESC-50 acc: ~82%
edge + speech commands
2024–2026 — BEATs / AST
patchify log-mel (16×16)
SSL-pretrained transformer
fine-tune on 1-10% labels
BEATs-iter3 leads AudioSet
default for non-speech 2026
ESC-50 acc: 97.0%
AudioSet mAP: 0.548
the features stayed the same for 30 years; the backend is the lever