Frequency domain blind source separation in a noisy environment

Abstract

A prototype system for blind source separation (BSS) in a reverberant and noisy environment is presented. Our system uses a small three‐dimensional array with eight microphones and realizes BSS in various situations such as the separation of many speech signals located in a three‐dimensional space, and the extraction of primary sound sources surrounded by many background interferences. The mixed signals observed by the microphone array are processed by independent component analysis (ICA) in the frequency domain. The permutation problem is solved by normalized basis vector clustering, which is a generalized version of directions of arrival (DOA) clustering and has several advantages to the DOA based method. The system estimates the DOA of the source signals as a by‐product of the separation process. Moreover, the system has the ability to distinguish primary target signals and ambient noise. Our system performs an on‐the‐spot BSS of live recorded signals. Live demonstration will be performed at the meeting.