Citation Context

... decreasing exponentials, as those on Fig. 4 (H), and adaptive Gex j . G ex j The two settings with Rj = 1 and 2, and unconstrained Eex j and Pex j correspond to the state-of-the-art methods [13] and =-=[17]-=-, respectively (see Section III-F), while the remaining six settings are new. In line with [13], parameter estimation via GEM is sensitive to initialization for all the settings we consider. To provid...

Citation Context

...ed to find some common formulation for methods we would like to generalize. Several recently proposed methods for source separation and/or characterization [12], [1], [7], [6], [5], [11], [10], [13], =-=[3]-=- (see also [14] and references therein) are based on the same zero-mean Gaussian model describing both the properties of the sources and of the mixing process, and only the global structure of Gaussia...

Citation Context

...n receiving attention with a view of realizing sound source localization and separation with multiple microphones, and have been studied for instantaneous mixtures [3, 4] and for convolutive mixtures =-=[5, 6]-=-. In a convolutive case, the phase difference information on timefrequency observations between different microphones is crucial for distinguishing/identifying source directions. Therefore, complexval...

Citation Context

...m and time-dependent gain. In the NMF separation framework the spatial properties of the sources can be modeled using a spatial covariance matrix (SCM) for each source at each STFT frequency bin [18]–=-=[22]-=-. Such extensions are hereafter referred to as complex-valued NMF (CNMF). The SCM denotes the mixing of the sources by magnitude and phase differences between the recorded channels, and is not depende...

Citation Context

...eighted combination of DoA kernels which are derived similarly to array manifold vectors towards a certain look direction as in the field of beamforming [10]. A benefit of the model over ones used in =-=[4, 5, 11]-=- is that the proposed structure of the SCMs is constrained by geometrically possible source directions by knowing the array geometry and phase difference caused by each DoA. Additionally, parameterizi...

Citation Context

...ling the source spectrograms (i.e. the magnitude of the source TF coefficients) with a low-rank matrix has not been used directly, but indirectly via the non-negative matrix factorization (NMF) [10], =-=[11]-=- which also assumes the non-negativity of the factors. While this idea has been quite successful in audio BSS, it remains that the NMF approximation has some important limitations: its solution is non...

Citation Context

...lutive source separation [27, 28]. Existing work pointing out the limitations of the narrowband approximation includes the statistical models of time-frequency source images introduced by Duong et al =-=[29, 30]-=-, and the convex optimization approaches for source estimation with known mixing filters of Kowalski et al [31]. Kowalski et al. show theoretically and experimentally that a wideband approach outperfo...

Citation Context

...ed to find some common formulation for methods we would like to generalize. Several recently proposed methods for source separation and/or characterization [12], [1], [7], [6], [5], [11], [10], [13], =-=[3]-=- (see also [14] and references therein) are based on the same zero-mean Gaussian model describing both the properties of the sources and of the mixing process, and only the global structure of Gaussia...