Adaptive separation of acoustic sources for anechoic conditions: A constrained frequency domain approach

摘要

Blind source separation represents a signal processing technique with a large potential for noise reduction. However, its application in modern digital hearing aids poses high demands with respect to computational efficiency and speed of adaptation towards the desired solution. In this paper, an algorithm is presented which fulfills these goals under the idealized assumption that the superposition of sources in rooms can be approximated as a superposition under anechoic conditions. Specifically, attenuation, the signals' finite propagation speed, and diffuse noise are accounted for, whereas reflections and reverberation are considered as negligible effects. This approximation is referred to as the Tree field' assumption. Starting from a general blind source separation algorithm for Fourier transformed speech signals, the free field assumption is incorporated into the framework, yielding a simple, fast and adaptive algorithm that is able to track moving sources. Implementation details are given which were found to be indispensable for fast and robust signal separation. Performance is evaluated both by simulations and experimentally, including separation of a moving and a fixed speaker in a recorded real anechoic environment. The potential benefits and shortcomings of this algorithm are discussed with regard to its inclusion into the signal processing framework of digital hearing aids for real reverberant acoustic situations.