YAMNet

YAMNet is Google's lightweight answer to general audio classification: a MobileNetV1 backbone trained on AudioSet that outputs 521 per-class scores plus a 1024-dimensional embedding every ~0.48 seconds from 16 kHz mono audio. The ontology covers the instrument classes that matter for music work (electric/acoustic guitar, piano, violin, trumpet, saxophone, drums, synthesizer, and more), which is why it keeps showing up as a timbre classifier inside music tools.

It is distributed via TF Hub under Apache 2.0 and at 16 MB runs fine on modest hardware, including microcontrollers in quantized form. Google trimmed six classes from the original 527-class AudioSet ontology on fairness grounds (the gendered speech/singing labels, battle cry, and funny music). It is general audio rather than music-first, so dedicated music taggers beat it on music-specific benchmarks, but as a free, tiny, well-documented classifier it remains one of the most deployed audio models in existence.

No music-specific published benchmark; evaluated by Google on AudioSet event classification; widely re-used but provider numbers only.