Audio foundation models are doing more than you think — and researchers now have receipts.
A team probing CLAP audio embeddings found that four low-level acoustic attributes — reverberation (RT60), loudness (LUFS), spectral centroid (SC), and relative pitch (RP) — are all reliably recoverable from frozen embeddings. The method trains probes of increasing complexity to predict each attribute across five datasets covering noise, speech, single musical notes, and full music mixes. Three of the four attributes — RT60, LUFS, and RP — decode cleanly with linear probes. Spectral centroid is the outlier: recovering it requires non-linear probes, meaning it is encoded in a curved, harder-to-untangle region of the embedding space.
The two-regime finding matters because it draws a practical boundary inside what is often treated as a uniform black box. If you are building a retrieval or classification system on top of CLAP, linear attributes are geometrically stable and composable; spectral centroid is not, and treating it as if it were will cost you. The result also holds across eight other audio foundation models — with one notable exception: amplitude-invariant architectures throw out loudness by design, so no probe recovers it regardless of complexity.
CLAP's cross-modal geometry gets a minor win here too — text embeddings of acoustic descriptors align with the RT60 feature direction, which is the kind of property multimodal retrieval depends on. The broader lesson is one the vision-model world learned a few years ago: knowing that a representation works is not the same as knowing what it encodes or how, and that gap tends to matter most when someone builds something production-grade on top of it.