Prototype-Rectified Iterative Self-supervised Manifold Denoising under Severe Acoustic Shift

2026-08-15

AI Summary

This study addresses the performance degradation of audio-text models under severe noise and the reliance of existing adaptation methods on gradients or labeled data. We propose PRISM, a training-free, source-free test-time adaptation framework grounded in an affine noise assumption. By leveraging frozen text prototypes as geometric anchors, PRISM compiles a static projection matrix via closed-form correction and bias regression, reducing adaptation to a single matrix multiplication that effectively circumvents polyphonic traps. Requiring only 0.0009ms for inference, PRISM achieves efficient manifold denoising without source data. On UrbanSound8K, it outperforms zero-shot baselines by 12.94% and surpasses oracle TTA baselines by 9.41%, demonstrating superior robustness and computational efficiency in noisy environments.

Abstract

Audio-Text Foundation Models (ATMs) fail catastrophically under severe acoustic noise, yet existing adaptation strategies either rely on gradient-based Test-Time Adaptation (TTA), which reinforces noise rather than signal, or on prompt tuning that requires privileged noise annotations unavailable at inference. We address these failures with PRISM (Prototype-Rectified Iterative Self-supervised Manifold Denoising), a training-free, source-free TTA framework grounded in the Affine Noise Hypothesis: severe acoustic noise induces a low-rank affine shift in the multimodal latent space, with more than 90% of distortion energy confined to the leading 60 principal components. PRISM estimates and reverses this distortion from an unlabeled target batch using frozen text prototypes as geometric anchors via three closed-form geometric corrections compiled into a single static projection matrix by Affine Bias Regression. At inference, adaptation reduces to one matrix-vector multiplication in 0.0009 ms, making it substantially faster than gradient-based TTA while requiring no additional training. On UrbanSound8K, PRISM improves over the zero-shot baseline by 12.94 percentage points and surpasses an oracle-assisted TTA baseline by 9.41 percentage points, despite never observing its privileged augmented noise prompts. We further identify the Polyphonic Trap, a principled failure mode of subspace deflation for broadband classes, and resolve it via Confidence-Aware Regression (CAR), recovering up to 8.16 percentage points for the worst-affected class.

Problem

• Audio-Text Foundation Models
• Severe Acoustic Noise
• Test-Time Adaptation
• Manifold Denoising
• Polyphonic Trap

Innovation

• Test-Time Adaptation
• Affine Noise Hypothesis
• Manifold Denoising
• Polyphonic Trap
• Confidence-Aware Regression