An analysis of the limitations of blind signal separation application with speech

Abstract

Blind Signal Separation (BSS) techniques are commonly employed in the separation of speech signals, using Independent Component Analysis (ICA) as the criterion for separation. This paper investigates the viability of employing ICA for real-time speech separation (where short frame sizes are the norm). The relationship between the statistics of speech and the assumption of statistical independence (at the core of ICA) is examined over a range of frame sizes. The investigation confirms that statistical independence is not a valid assumption for speech when divided into the short frames appropriate to real-time separation. This is primarily due to the quasi-stationary nature of speech over the temporal short term. We conclude that employing ICA for real-time speech separation will always result in limited performance due to a fundamental failure to meet the strict assumptions of ICA.