Multi-stage representation learning for blind Room-Acoustic parameter estimation with uncertainty quantification

The ability to infer a general representation of the acoustic environment from a reverberant recording is a key objective in numerous applications. We propose a multi-stage approach that integrates task-agnostic representation learning with uncertainty quantification. Leveraging the conformal prediction framework, our method models the error incurred in the estimation of the acoustic environment embedded in a reverberant recording, which reflects the ambiguity inherent in distinguishing between an unknown source signal and the induced reverberation. Although our approach is flexible and agnostic to specific downstream objectives, experiments on real-world data demonstrate competitive performance on established parameter estimation tasks when compared to baselines trained end-to-end or with contrastive losses. Furthermore, a latent disentanglement analysis reveals the interpretability of the learned representations, which effectively capture distinct factors of variation within the acoustic environment.