Regularized Schrödinger Bridge (RSB) is a generative speech enhancement approach that reconciles fidelity and realism while mitigating exposure bias. RSB regularizes training with a Distortion-Perception Perturbation that constructs time-varying targets by interpolating between clean speech and posterior-mean estimates, and trains the network on perturbed intermediate states to correct toward the ground truth progressively. Consequently, such perturbation simulates inference-time prediction errors, mitigating the training–inference mismatch and thereby reducing exposure bias. Furthermore, it also injects posterior-mean estimates as fidelity-preserving guidance, facilitating reconstruction fidelity.
We have publicly released a checkpoint of MISB's generative model, which is based the ncsnpp_base architecture and was trained on the Voicebank+Demand dataset.
This project is licensed under the
Apache License 2.0.
