SSMphony: Linear-Complexity Audio Generation with Mamba & State Space Models

Explored modern State Space Models (SSMs) and selective structured state space architectures (Mamba) for continuous long-context audio and waveform synthesis.

Technical Focus & Architectures

• Overcoming Quadratic Bottlenecks: Standard transformer-based autoregressive audio synthesis suffers from (O(L^2)) computational complexity, making high-sample-rate audio generation memory-prohibitive.
• Continuous-Time State Space Discretization: Formulated continuous state space dynamics: [ h’(t) = \mathbf{A} h(t) + \mathbf{B} x(t), \quad y(t) = \mathbf{C} h(t) + \mathbf{D} x(t) ] discretized via Zero-Order Hold (ZOH) to yield parallelizable convolutional training and (O(1)) recurrent inference per step.
• Selective State Representation: Leveraged input-dependent selection mechanisms to selectively propagate or filter out audio frequency bands, achieving high acoustic fidelity with sub-quadratic memory scaling.