Speech Separation for Aerodrome Traffic Control Recordings by Auxiliary function based Independent Vector Analysis

Abstract

Voice recordings of aerodrome traffic control include vocal control directive and vocal duty shift speech which are spoken by aerodrome traffic controllers. These recordings are significant for safety investigation and shall be clearly recorded according to management requirements. The microphones of recorder are installed on each working position so that original recordings are usually mixture of all the voice sources in the control room (voice of all controllers and background noise). This scenario represents a typical “Cocktail party problem” with multi-source and multi-sensor. This paper introduces an algorithm to separate the voice of each controller by only using the observation of system and supposing the mixing process unknown. Simulation with the room impulse response considered shows a practical performance and improvement compared to traditional noise cancel method or blind source separation method.