Paper No: 68
Blind Vector Deconvolution: Convolutive Mixture Models in Short-time Fourier Transform Domain
Author(s): Atsuo Hiroe
Abstract
Short-time Fourier Transform (STFT) domain (or frequency domain) ICA often suffers
from a dilemma of STFT frame length; any shorter frame than reverberation
(short frame) generates incomplete instantaneous mixtures, while too long
frames may disturb the separation. To avoid the dilemma, the authors propose
a novel framework which accounts for STFT with short frames. In the framework,
time domain convolutive mixtures are transformed to STFT domain convolutive
mixtures, not instantaneous mixtures. For separating the mixtures, the authors
present an approach that applies another STFT so as to treat them as instantaneous
mixtures.
The authors experimentally confirm that this framework outperforms the conventional
STFT domain ICA, due to avoiding the dilemma. |
