Modeling the Short Time Fourier Transform Ratio and Application to Underdetermined Audio Source Separation

摘要

This paper presents the theoretical background for the Model Based Underdetermined Source Separation presented in [5]. We show that for a given frequency band, in contrast to customary assumption, the observed Short-Time Fourier Transform (STFT) ratio coming from one source is not constant in time, but is a random variable whose distribution we have obtained. Using this distribution and the Time-Frequency (TF) "disjoint" assumption of sources, we are able to obtain promising results in separating four audio sources from two microphones in a real reverberant room.