Audio Signal Denoising with Reduced Gaussian Distributions
Résumé
Probabilistic time-frequency representations, though highly promising for signal processing applications, particularly in audio, have been largely overlooked due to their computational complexity and the difficulty in interpreting the underlying models. However, recent research has revealed that these representations can be modeled using Spectral Mixtures Gaussian Processes (SM-GP), which offers a solution to these challenges and enables more tractable and interpretable probabilistic analysis. In this paper, we extend this work by introducing a significant improvement in terms of computational efficiency, flexibility, and real-world applicability. Our approach leverages a recent covariance approximation technique to accelerate SM-GP inference, alongside a frequency-domain method for hyperparameter learning.
Domaines
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |