Post by Orchisama Das

We’ve been tackling a major hurdle in AR audio: how to learn the acoustics of complex spaces from limited RIRs and render efficient, realistic 6DoF late reverberation without the massive computational cost. Our results are real-time compatible and significantly faster than traditional Neural Acoustical Fields. 📄 𝐖𝐀𝐒𝐏𝐀𝐀: Learning common slope parameters with DNNs, fast inference via shaped white noise convolution.[https://lnkd.in/e5Rv9uAD] 📄 𝐓𝐀𝐒𝐋𝐏: Learning Differentiable FDN parameters for efficient, real-time rendering as the listener moves. [https://lnkd.in/edfgdmNg] 📄 𝐈𝐂𝐀𝐒𝐒𝐏: Learning and rendering anisotropic late reverberation with Differentiable FDNs.[https://lnkd.in/eapaYYUt] 💻 Codebase: https://lnkd.in/enG6_vBQ This was the outcome of a productive collaboration with Gloria Dal Santo, Sebastian Schlecht, Vesa Välimäki and my PI, Zoran Cvetkovic and builds on prior work by Georg Götz.