Phase-Aware Spectral Speech Enhancement Using Deep Learning Techniques

Beschreibung

In everyday environments, speech is often degraded by background noise, reverberation, echo, or transmission losses. These distortions reduce quality and intelligibility, impairing communication. Speech enhancement techniques aim to overcome these challenges by improving the perceptual quality and clarity of speech under adverse conditions. This dissertation advances the emerging field of phase-aware speech enhancement, which extends conventional magnitude-based methods by also processing the often-overlooked phase spectrum. Novel concepts for deep learning-based approaches are proposed and evaluated, with a particular focus on phase estimation and its integration into speech enhancement. Beyond theoretical investigations that highlight the potential of phase processing, methods for estimating the phase with deep neural networks are introduced, and strategies for jointly optimizing magnitude and phase estimation are proposed. Objective measures and subjective listening experiments confirm the effectiveness of the proposed approaches, underlining their relevance for the next generation of speech enhancement systems.

Autor*in

Lars-Hendrik Thieling

Themen in »Phase-Aware Spectral Speech Enhancement Using Deep Learning Techniques«

Speech Enhancement Deep Learning Phase Estimation

Stimmen zu »Phase-Aware Spectral Speech Enhancement Using Deep Learning Techniques«

Details

Link teilen