Speech enhancement with variational autoencoders and alpha-stable distributions
Résumé
his paper focuses on single-channel semi-supervised speech en-hancement. We learn a speaker-independent deep generative speechmodel using the framework of variational autoencoders. The noisemodel remains unsupervised because we do not assume prior knowl-edge of the noisy recording environment. In this context, our con-tribution is to propose a noise model based on alpha-stable distribu-tions, instead of the more conventional Gaussian non-negative ma-trix factorization approach found in previous studies. We develop aMonte Carlo expectation-maximization algorithm for estimating themodel parameters at test time. Experimental results show the supe-riority of the proposed approach both in terms of perceptual qualityand intelligibility of the enhanced speech signal.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|
Loading...
