Application of Two Time-Domain Convolutive Blind Source Separation Algorithms to the 2008 Signal Separation Evaluation Campaign (SiSEC) Data

Abstract

This paper describes the application of two time-domain convolutive blind source separation algorithms – the scaled natural gradient algorithm [1] and the spatio-temporal FastICA algorithm with symmetric orthogonality constraints [2] – to a portion of the determined and overdetermined acoustic data sets created for the 2008 Signal Separation Evaluation Campaign (SiSEC). As the 2008 SiSEC competition provides no ground truth data and thus no a priori method for numerical performance calculation, our approach to determining overall performance is a decoding of the contents of the recorded sources used to create the data through the two algorithms used. Information about the sources themselves, such as the instrumentation and structure of the musical selections chosen, the qualities of the voices and written transcripts of what is spoken, and additional information about the signals extracted, are provided without our ever having heard the sources in isolation. A qualitative performance comparison of the two approaches is also provided.