DeepMind's AudioFM hits best results on 10 of 12 audio tasks, beats prior SSL on heart murmur by +4.7 F1

iScienceLuvr · x · 2026-10-27

Google DeepMind researchers introduced AudioFM, a self-supervised framework for general-purpose audio representation learning (arXiv:2610.34183).

Method: combines recording-level contrastive learning with patch-wise masked latent discrimination — a student predicts representations of time/frequency-masked spectrogram patches and picks the matching full-view EMA-teacher target from candidates in other recordings, with weaker supervision from masked temporal neighbors for local consistency.
Results: best or joint-best linear-probing performance on 10 of 12 tasks, with SOTA event detection, scene understanding, and domain generalizability.
Highlight: outperforms prior audio SSL on medical audio — +4.7 F1 on heart murmur and +3.2 F1 on sleep apnea subtype classification, underscoring the promise of audio foundation models in medical AI.

More from Multimodal