Audio Mamba: Bidirectional State Space Model for Audio Representation Learning
Mehmet Hamza Erol, Arda Senocak, Jiu Feng, Joon Son Chung
TL;DR
This work introduces Audio Mamba (AuM), the first self-attention-free, purely SSM-based model for audio classification, addressing the high $O(n^2)$ cost of self-attention by employing bidirectional state-space models to capture global context. AuM processes spectrogram patches with a mid-sequence classification token in a patch-embedded sequence and uses a bidirectional Mamba encoder to achieve competitive accuracy against AST while exhibiting linear-time complexity $O(n)$ with respect to sequence length. Across six diverse datasets, AuM demonstrates comparable or superior performance to AST, along with memory and speed advantages, and the authors support their claims with extensive ablations and pre-training analyses that highlight AuM's potential for self-supervised and multimodal learning. The findings suggest AuM as a strong, efficient audio backbone suitable for long sequences and future self-supervised or multimodal integration.
Abstract
Transformers have rapidly become the preferred choice for audio classification, surpassing methods based on CNNs. However, Audio Spectrogram Transformers (ASTs) exhibit quadratic scaling due to self-attention. The removal of this quadratic self-attention cost presents an appealing direction. Recently, state space models (SSMs), such as Mamba, have demonstrated potential in language and vision tasks in this regard. In this study, we explore whether reliance on self-attention is necessary for audio classification tasks. By introducing Audio Mamba (AuM), the first self-attention-free, purely SSM-based model for audio classification, we aim to address this question. We evaluate AuM on various audio datasets - comprising six different benchmarks - where it achieves comparable or better performance compared to well-established AST model.