PERFORMANCE EVALUATION OF SPARSE SOURCE SEPARATION AND DOA ESTIMATION WITH OBSERVATION VECTOR CLUSTERING IN REVERBERANT ENVIRONMENTS
Shoko Araki, Hiroshi Sawada, Ryo Mukai, Shoji Makino · 2006
This paper investigates the effects of real world acoustic envi-ronments on sparse source separation and direction of arrival (DOA) estimation performance. The time-frequency mask tech-nique is widely studied as an approach for sparse source separa-tion and DOA estimation. The approach relies on source sparse-ness, which can easily be affected by, for example, reverbera-tion. In fact, most proposed approaches assume an anechoic condition, which is difficult to maintain in a real acoustic en-vironment. We investigate how the performance of such meth-ods is affected when the problem does not meet the assumed conditions. We show that strong reverberation and a large dis-tance between the sources and sensors degrade the separation performance, however, the DOA estimation performance is not so severely affected. 1.