On the use of the Relative Transfer Function for Source Separation using Two-channel Recordings

This paper investigates the use of the relative transfer function (ReTF) for source separation. ReTFs have been used to localize sound sources but have not been thoroughly investigated for the application of source separation especially when one of the sources is not speech. We propose two source separation algorithms using ReTFs. One of them is deterministic and enables the separation of two sources when one or both of their ReTFs are known. The other algorithm uses masking in the time-frequency domain and can be used for separating two or more sources. We also explore the limitations and assumptions of the ReTF and the proposed source separation algorithms.