Quantifying the Uncertainty of Blindly Estimated Room Embeddings Using a Dispersion-Calibrated Score

2026-07-01

AI Summary

Blind estimation of room embeddings is highly susceptible to interference from speech content and recording distortions, degrading downstream task performance. This work proposes a self-supervised framework that requires no downstream supervision, learning speech-content-invariant room representations and associated confidence estimates directly from single-channel reverberant speech. The approach leverages a structured embedding space derived from the room impulse response (RIR) latent manifold, a multi-positive contrastive learning mechanism, and a lightweight uncertainty head calibrated via embedding dispersion. By incorporating multi-view data construction, KL alignment, and joint waveform-spectrogram distortion modeling, the method yields uncertainty scores that closely align with embedding dispersion, enabling reliable selective prediction across diverse distortion conditions.

Abstract

Room embeddings derived from reverberant speech are often unreliable: speech content and recording degradation can alter the representation even when speaker, room, and source-receiver geometry remain unchanged, degrading downstream task performance. We propose a framework that learns room embeddings robust to speech-content variation and a representation-level uncertainty score from reverberant speech without downstream-task supervision. The embedding is anchored to a structured room impulse response (RIR) latent space and trained using a multi-view data structure with Kullback-Leibler (KL)-based alignment; a multi-positive contrastive term further refines robustness. A lightweight uncertainty head is calibrated using the dispersion of corruption-induced embeddings and optimized with a rank-based objective. Across waveform- and spectrogram-level corruptions, the score is consistent with representation dispersion and enables effective selective prediction while requiring only a single utterance at inference.

Problem

• room embeddings
• reverberant speech
• representation uncertainty
• speech-content variation
• downstream task performance

Innovation

• room embedding
• uncertainty quantification
• reverberant speech
• contrastive learning
• dispersion calibration