NDF+: Joint Neural Directional Filtering and Diffuse Sound Extraction

A team led by Weilong Huang has published a new paper on arXiv titled "NDF+: Joint Neural Directional Filtering and Diffuse Sound Extraction." The work builds on the previously introduced neural directional filtering (NDF) approach for reconstructing a virtual directional microphone (VDM) with a desired directivity pattern. NDF+ reformulates VDM estimation into two coupled subtasks: first, recovering a dereverberated VDM signal, and second, extracting the diffuse sound component. This separation gives the model an additional degree of freedom to control the diffuse part of the final reconstructed VDM output.

In experiments under reverberant conditions, NDF+ consistently outperformed representative conventional baselines on both subtasks—dereverberated VDM reconstruction and diffuse sound extraction—while maintaining overall VDM reconstruction quality comparable to the original single-task NDF. Crucially, the method allows practical control over the inter-channel level difference between left and right channels in a stereo recording application by adjusting the estimated diffuse component. This opens up new possibilities for spatial audio capture and post-processing in immersive audio, VR/AR, and professional sound engineering.