作者: Clément Laroche , Gael Richard , Hélène Papadopoulos , Matthieu Kowalski
DOI:
关键词:
摘要: Dans cet article, nous proposons une methode structuree de decomposition en matrices non-negatives visant a utiliser la structure multi-couche des signaux audio. Les audio peuvent etre vus comme superposition deux couches : couche tonale (modelisee par sommes sinuso¨des evoluant lentement frequence et temps) transitoire (les sons percussifs, ´ ev enements courtes durees etales frequence). Notre decompose partie du signal composantes orthogonales parcimonieuses, bien adaptees pour l'extraction tandis que est representee bases classiques. resultats separation sources obtenus sur reels musique ont montre notre approche obtient similaires ceux l'´ etat l'art. Abstract – In this paper, we propose new unconstrained nonnegative matrix factorization method designed to utilize the multilayer of signals improve quality source separation. The tonal layer is sparse in frequency and temporally stable, while transient composed short term broadband sounds. Our has part well suited for extraction which decomposes orthogonal components, represented by regular decomposition. Experiments on real music data context show that such suitable signal. Compared with three state-of-the-art harmonic/percussive algorithms, proposed shows competitive performances.