Merging masked patches speeds up AST training 13.9% with nearly no accuracy loss

kastnerkyle · x · 2026-09-16

Accepted to Interspeech 2026, this paper proposes SpecAugment-Patch Merging, a simple method to accelerate Audio Spectrogram Transformer (AST) training: SpecAugment is first applied to mask input spectrograms at the patch level; after positional embeddings are added, r pairs of masked patches are selected and merged, reducing the number of tokens the Transformer processes.

Experiments show that increasing merged pairs r from 0 to 100 keeps AudioSet mAP nearly unchanged (34.07 → 34.08) while throughput rises from 43.3 to 49.3 samples/sec, a 13.9% relative improvement. The same pattern holds on ESC-50 and Speech Commands V2, with steady throughput gains and only minor accuracy changes—demonstrating faster training at minimal performance cost.

Original post →

More from Research

Research channel →