From Masking to Merging: Rethinking SpecAugment for Efficient Audio Spectrogram Transformer

Increasing depth of transformer models improves recognition, but it comes at a substantial cost. Each additional layer requires more parameters, which makes the process computationally inefficient. We ask whether additional processing can focus on integrating features already computed. Looped Audio Spectrogram Transformer (LAST) first processes all tokens, then reuses the same blocks to refine only the class token over fixed audio features, thereby making later passes inexpensive. On AudioSet, ten-pass LAST achieves 0.345 mean average precision, exceeding a twelve-layer sequential transformer by 2.1% relative with 49.4% fewer parameters, 42% fewer multiply-accumulate operations, and 9.8% higher measured throughput. Across separately trained models, increasing the pass count from two to ten improves accuracy while adding only 1.2% computation. Further evaluations show improved robustness to temporal masking and various other auditory augmentations, with better generalization on classification tasks with music, environmental, and event sounds.

Pretrained audio spectrogram transformers are widely used as backbones of machine listening systems for sound event classification. However, they are typically trained using spectrograms with a fixed Short-Time Fourier Transform (STFT) configuration, implicitly binding the model to the time-frequency resolution seen during pretraining. In practice, adapting these systems to new domains or computational constraints often requires different STFT frontend parameters. In this work, we conduct a controlled study of the sensitivity of the foundational Audio Spectrogram Transformer to variations in STFT window and hop size, showing that audio classification performance degrades substantially as frontend parameters deviate from pretraining settings. As a mitigation strategy, we introduce a lightweight conditioning mechanism using FiLM that enables a pretrained model to adapt to variable input configurations using limited additional training data. Our method generalizes well to unseen STFT configurations, improving model robustness without sacrificing performance in the original setting.

Convolutional neural networks (CNNs), used widely and successfully in audio enhancement, capture long-range time-frequency dependencies only indirectly, through successive convolution and pooling. Here, we present U-PAST, a hybrid transformer-U-Net architecture that addresses this limitation through self-attention dependency-modeling in the complex spectrogram domain. U-PAST tokenizes a complex STFT representation, similarly to the magnitude spectrogram tokenization of the Audio Spectrogram Transformer (AST), applies a multi-layer transformer encoder, and reconstructs the enhanced complex spectrogram with a U-Net-style decoder. We evaluate four architectural variants with between 1.17M and 2.40M parameters on the DNS Challenge, VoiceBank-DEMAND, and LibriMix corpora under matched, acoustic mismatch, and two-dataset mismatch conditions. U-PAST attains the best SI-SDR of any evaluated model under acoustic mismatch and closely trails substantially larger convolutional and time-domain baselines by 0.26 dB to 0.63 dB SI-SDR under the remaining three conditions while achieving the strongest perceptual (DNSMOS) quality under dataset mismatch. The largest evaluated configuration, U-PAST-H (2.40M parameters), is consistently the strongest variant of the family, offering an attractive performance-to-cost trade-off at a small parameter footprint.