HELIX: Scaling Raw Audio Understanding with Hybrid Mamba-Attention Beyond the Quadratic Limit
This work addresses scaling challenges in audio processing for researchers and practitioners, but it is incremental as it builds on existing Mamba and attention methods.
The paper tackled the problem of scaling raw audio understanding by showing that design choices like input frontend and sequence backbone are coupled, and introduced HELIX, a hybrid Mamba-attention framework that outperforms pure Mamba by closing an 11.5-point gap on a 5-minute speaker identification task with 30,000 tokens.
Audio representation learning typically evaluates design choices such as input frontend, sequence backbone, and sequence length in isolation. We show that these axes are coupled, and conclusions from one setting often do not transfer to others. We introduce HELIX, a controlled framework comparing pure Mamba, pure attention, and a minimal hybrid with a single attention bottleneck. All models are parameter-matched at about 8.3M parameters to isolate architectural effects. Across six datasets, we find that the preferred input representation depends on the backbone, and that attention hurts performance on short, stationary audio but becomes important at longer sequence lengths. On a 5-minute speaker identification task with 30,000 tokens, pure attention fails with out-of-memory errors, while HELIX closes an 11.5-point gap over pure Mamba.