VeRe-Flow: Guiding Flow Matching toward Clean Speech via Velocity Contrastive and Representation Regularization for Noise-Robust Bandwidth Expansion
Abstract
Noise-robust bandwidth expansion (NR-BWE) aims to reconstruct high-fidelity wideband speech from noisy low-resolution inputs. While flow matching has shown strong performance for speech generation, accurately recovering clean speech under noisy conditions remains challenging due to ambiguous velocity estimation. In this work, we propose VeRe-Flow, a clean-guided flow matching that introduces multi-level clean supervision to guide the generative process toward clean speech. At the velocity level, we introduce velocity contrastive regularization, which attracts the predicted velocity toward the clean trajectory while repelling it from noisy manifold. At the representation level, we incorporate a representation alignment objective that aligns intermediate features with clean self-supervised speech representations. Experimental results demonstrate that the proposed method achieves the lowest LSD, the highest DNSMOS OVRL and MOS among NR-BWE baselines. Audio samples are available.
This work has been submitted to Interspeech 2026 for review.
Figure 1. Overview of the proposed system