Improved stereo instrumental track recovery using median nearest-neighbour inpainting

Abstract

Several algorithms have been proposed for vocal removal which operate by finding the position of the vocals in the stereo field and removing the time-frequency bins associated with that position. However, in many cases, there will be other instruments such as drums and bass guitar present in the same position. These instruments will be removed along with the vocals and so adversely affect the sound quality of the recovered instrumental track. We present a method for estimating the missing information in the removed time-frequency bins, while still suppressing vocals, allowing recovery of an improved stereo instrumental track.