StoRM speech enhancement's performance 'kink' traced to predictor stage via new flow analysis
Researchers localize the sudden SI-SDR drop at training noise amplitude to a single model component.
Diffusion-based speech enhancement models like StoRM pair a deterministic predictor with a learned score network. These systems exhibit a sharp, non-smooth transition (a "kink") in their SI-SDR degradation curve exactly at the noise amplitude used during training. Until now, engineers could only work around this anomaly without understanding its root cause.
Shubham Ojha's new theoretical work provides a pathwise variational-flow analysis that isolates the kink to the predictor stage. The key insight is an exact factorization: ∂σ^(M)/∂M = K(M) · ∂C_M/∂M, where K(M) depends on the score Jacobian along the reverse trajectory, and C_M is the predictor output. Under three reasonable hypotheses, the kink exists if and only if the predictor output C_M is non-smooth at the same point. The analysis also covers the finite-step Euler–Maruyama sampler used in practice, making it directly relevant to deployed systems. Empirical results are promised in a separate experimental report.
- StoRM's SI-SDR degradation curve has a sharp kink exactly at the training noise amplitude, previously unexplained.
- New variational-flow analysis shows the non-smoothness originates in the predictor stage, not the score network.
- A matrix-valued functional K(M) links predictor output C_M to overall performance degradation, enabling targeted fixes.
Why It Matters
Pinpointing the cause lets researchers redesign predictor modules to eliminate the kink, improving speech enhancement robustness.