Blind Source Separation for musical signals is an active research area and is currently evaluation often using only the Signal to Distortion Ratio. However, he metric has been critised in literature for not correlating with listeners rating scores for separation models. Hence, this thesis aims to document the creation of a new metric called audioConv, a deep learning perceptually inspired metric. AudioConv is Convolutional Neural Network using additional features based upon well established models. Analysis of audioConv is through the correlation with listener rating score and shows that while there is a potential, further work is needed improve the metric. The work focuses on machine learning techniques for audio and the need for quality data for these models.