PRISM proposes a test-time adaptation method for Audio-Text Foundation Models under severe acoustic noise, based on the Affine Noise Hypothesis that the distortion is a low-rank affine shift in the multimodal latent space, with over 90% of the distortion energy in the leading 60 principal components. The method estimates and reverses this shift with frozen text prototypes as anchors and a single static projection matrix derived by Affine Bias Regression.1
Context
Existing adaptation strategies either use gradient-based test-time adaptation, which can reinforce the noise, or prompt tuning requiring privileged noise annotations. PRISM avoids both by making no calls to the training data and using only an unlabeled target batch, which removes the need for gradient updates or supplied noise prompts at inference.1
Assessment
On UrbanSound8K, PRISM reports a 12.94 percentage-point improvement over the zero-shot baseline and surpasses an oracle-assisted TTA baseline by 9.41 percentage points. It also identifies the Polyphonic Trap, a failure mode of subspace deflation for broadband classes, and introduces Confidence-Aware Regression to recover up to 8.16 percentage points for the worst-affected class.1
Caveats
The abstract provides no architecture details, experimental protocol, or formal definition of the Polyphonic Trap. It also does not state how the static projection matrix depends on batch composition or whether the low-rank affine assumption holds across real-world, non-synthetic noise. These omissions limit current verification; the result would be more compelling with error bars, more datasets, and code.1
