Hold Me Tight : Stable Encoder-Decoder Design for Speech Enhancement - Archive ouverte HAL
Communication Dans Un Congrès Année : 2024

Hold Me Tight : Stable Encoder-Decoder Design for Speech Enhancement

Résumé

Convolutional layers with 1-D filters are often used as frontend to encode audio signals. Unlike fixed time-frequency representations, they can adapt to the local characteristics of input data. However, 1-D filters on raw audio are hard to train and often suffer from instabilities. In this paper, we address these problems with hybrid solutions, i.e., combining theory-driven and datadriven approaches. First, we preprocess the audio signals via a auditory filterbank, guaranteeing good frequency localization for the learned encoder. Second, we use results from frame theory to define an unsupervised learning objective that encourages energy conservation and perfect reconstruction. Third, we adapt mixed compressed spectral norms as learning objectives to the encoder coefficients. Using these solutions in a low-complexity encoder-mask-decoder model significantly improves the perceptual evaluation of speech quality (PESQ) in speech enhancement.
Fichier principal
Vignette du fichier
haider2024interspeech.pdf (1.61 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04708073 , version 1 (24-09-2024)

Identifiants

  • HAL Id : hal-04708073 , version 1

Citer

Daniel Haider, Felix Perfler, Vincent Lostanlen, Martin Ehler, Peter Balazs. Hold Me Tight : Stable Encoder-Decoder Design for Speech Enhancement. INTERSPEECH, International Speech Communication Association (ISCA), Sep 2024, Kos, Greece. ⟨hal-04708073⟩
0 Consultations
0 Téléchargements

Partager

More