Reference-Based Speech Enhancement via Feature Alignment and Fusion Network

Speech enhancement aims at recovering a clean speech from noisy input, which can be classified into single and personalized enhancement. Personalized usually utilizes the speaker identity extracted itself (or reference speech) as global embedding to guide process. Different them, we observe that speeches of same are correlated in terms frame-level short-time Fourier Transform (STFT) spectrogram. Therefore, propose reference-based via feature alignment fusion network (FAF-Net). Given spoken by speaker, first level strategy warp with frame level. Then, fuse similarity-based strategy. Finally, fused features skipped connected decoder, generates enhanced results. Experimental results demonstrate performance proposed FAF-Net is close state-of-the-art methods on both DNS Voice Bank+DEMAND datasets. Our code available https://github.com/HieDean/FAF-Net.