A Fast Solver for Interpolating Stochastic Differential Equation Diffusion Models for Speech Restoration

Researchers Bunlong Lay and Timo Gerkmann have published a paper introducing a novel fast solver specifically designed for interpolating Stochastic Differential Equation (iSDE) diffusion models, a category that includes the established speech enhancement model SGMSE+. The core innovation addresses a major bottleneck: traditional diffusion models for tasks like speech restoration require solving a complex reverse process, which can demand hundreds of evaluations of a large neural network, making them slow and computationally expensive. The team's new solver framework is tailored to the unique mathematics of iSDEs, which interpolate between a target clean signal and a noisy observation, unlike standard image diffusion models that move between data and pure noise.

This technical breakthrough enables remarkably efficient sampling. The proposed solver can generate high-quality restored speech with as few as 10 neural network evaluations across multiple tasks like denoising and enhancement. This represents a potential order-of-magnitude speedup, transforming these powerful but previously sluggish models from research curiosities into practical tools. By drastically reducing the computational cost, the work paves the way for real-time or near-real-time application of state-of-the-art diffusion models in audio processing, from cleaning up podcast recordings to restoring historical audio archives.