Software & Data Downloads — SIPS

Stochastic Interpolants for Speech Enhancement and Separation for refining processed or cleaning speech signals to improve perceptual quality.

This repository contains the PyTorch training and inference code for the paper “Stochastic Interpolants for Speech Enhancement and Separation.” It is implemented as a Python package and includes the model checkpoint for SIPS, a generative speech prior designed to refine processed or clean speech signals to improve perceptual quality. The package is designed to operate on speech files, allowing users to provide either model-enhanced outputs or corrupted inputs. When given corrupted inputs, it can automatically run supported predictors such as SEMamba, Conv-TasNet, and NCSN++ for speech enhancement, and and FlexIO for speech separation.