説明
Sound Scene Understanding Team (https://aip.riken.jp/labs/goalorient_tech/sound_scene_understand/) at RIKEN AIP
Speaker 1: Kazuyoshi Yoshii (45 min)
Title: The Unified Theory of Blind Source Separation Based on Independence, Nonnegativity, and Low-rankness
Abstract: We comprehensively review blind source separation (BSS) methods from a unified theoretical point of view. Independence, nonnegativity, and low-rankness inherent in sound sources, which are the three major clues of BSS, have been used for formulating a probabilistic model of observed mixture signals consisting of a source model representing the time-frequency structures of sources and a spatial model representing the inter-channel covariance structures of the sources. Nonnegative matrix factorization (NMF) is the most basic technique used for single-channel BSS and has successfully been extended to deal with the time, frequency, and/or spatial covariance structures of sources for single- and multi-channel BSS. We organize existing BSS methods in terms of covariance complexities (full-rank, rank-1, and jointly diagonalizable models) over the time, frequency, and channel dimensions. Among them, FastMNMF is considered as the state-of-the-art versatile BSS method with excellent separation performance and computational efficiency in practice.
Speaker 2: Kouhei Sekiguchi (15 min)
Title: Joint Blind Source Separation and Dereverberation with ARMA Models
Abstract: We explain an extension of FastMNMF (ARMA-FastMNMF) for joint blind source separation and dereverberation. The probabilistic model of ARMA-FastMNMF is obtained by integrating the source and spatial models of FastMNMF with AR and MA models representing the early reflections and late reverberations of sources, respectively.
Speaker 3: Mathieu Fontaine (15 min)
Title: Robust Blind Source Separation with Heavy-Tailed Models
Abstract: We comprehensively review heavy-tailed extensions of FastMNMF from a unified theoretical point of view.
Speaker 4: Yoshiaki Bando (15 min)
Title: Semi-supervised Source Separation with Deep Source Models
Abstract: We explain a DNN-based extension of the source model for semi-supervised source separation.
Speaker 5: Aditya Arie Nugraha (15 min)
Title: Unsupervised Source Separation with Deep Spatial Models
Abstract: We explain a DNN-based extension of the spatial model for unsupervised BSS.
