Discovering Directions of Uncertainty in Speech Inpainting

Speech inpainting aims to restore missing segments in audio signals. While both classical signal processing and deep learning approaches have been developed for this task, they typically generate a single output, despite the existence of multiple plausible reconstructions. In this paper, we adapt Neural Posterior Principal Component (NPPC) to recover the principal directions of variation in the posterior distribution of the original signal given its masked version. Our empirical results demonstrate that these directions capture diverse and meaningful variations in both speech content and style, while more precisely capturing the predictive error compared to a more costly Bayesian deep learning approach.