Diffusion-based Speech Enhancement: Demonstration of Performance and Generalization

Julius Richter, Timo Gerkmann

Keywords: diffusion models, speech enhancement, Schrödinger bridge

Abstract: This demo presents advanced techniques in speech enhancement using deep generative models. It highlights the generalization capabilities of score-based generative models for speech enhancement and compares directly with Schrödinger bridge approaches. The presented methods focus on generating high-quality super-wideband speech at a sampling rate of 48 kHz. Participants will record speech using a single microphone in a noisy environment, such as a conference venue. These recordings will then be enhanced and played back through headphones, demonstrating the model's effectiveness in improving speech quality and intelligibility.

Submission Number: 37