A single transport vector between clean and degraded latent centroids of neural audio codecs matches large diffusion models on music bandwidth extension

- Koops, Tan & Quinton (2026) compare state-of-the-art conditional latent generative methods (diffusion, Schrodinger Bridges, Flow Matching) against a simple arithmetic baseline for music bandwidth extension in the latent spaces of four open neural audio codecs (Stable Audio Open VAE, CodiCodec, DAC, EnCodec). [koops-etal-2026-music-bandwidth-extension-audio-codec-geometry] - The baseline estimates a single transport vector between the clean and bandwidth-degraded latent centroids on a reference set and simply adds it to a degraded latent; this arithmetic vector-addition operation, requiring no learned conditional model, yields restoration performance competitive with large diffusion/Schrodinger-bridge models on some metrics and some datasets (e.g. on MTD at 4 kHz, lower LSD than AudioSR and A2SB without partitioning for VAE, CodiCodec and DAC), while large models remain stronger on other metrics such as SiSpec and on MAESTRO. The per-sample restoration displacements align with the mean-shift vector (average cosine 0.985-0.986 for VAE, CodiCodec, DAC but 0.481 for EnCodec on MTD), the direction is governed more by cutoff frequency than by dataset, and the same probe is much less aligned for denoising, declipping and dereverberation (cosine 0.13-0.63). [koops-etal-2026-music-bandwidth-extension-audio-codec-geometry] - The authors suggest that "some neural codec latent spaces exhibit structure aligned with audio bandwidth" and that in such cases complex conditional models "may offer only limited gains over a simple vector addition"; they describe the mean-shift vector as "the average linear direction that represents bandwidth extension in the latent space". [koops-etal-2026-music-bandwidth-extension-audio-codec-geometry]