[Updated on Apr. 18] Integration of CiNii Articles into CiNii Research


HANDLE Web Site Open Access

Bibliographic Information

Other Title
  • タイスウ スペクトル ジゼン ブンプ オ モチイタ MAP スペクトル スイテイ ニ モトズク レツケッテイオン ゲン ブンリ
  • Under-Determined Audio Source Separation Based on MAP Spectral Estimation Using Log-Spectral Prior

Search this article


Assuming speech to be non-stationary Gaussian process, maximum likelihood spectral estimation has been studied as an effective speech enhancement approach. Recently, to improve the estimation accuracy of this approach, we have proposed an extention of it, namely a maximum a posterior (MAP) estimation approach using pre-trained log-spectral priors, and showed its effectiveness. This paper newly applies this extention to a multi-channel Wiener filtering based undetermined blind source separation (BSS) technique proposed by Duong et al. This conventional method adopts the likelihood maximization approach for estimating the source spectra and the spatial correlation matrices for the Wiener filtering. The proposed method extends it by introducing the MAP estimation approach for estimating the source spectra, and improves the accuracy of the Wiener filtering.

IEICE Technical Report;EA2012-114



Report a problem

Back to top