Speech Separation with Microphone Arrays using the Mean Shift Algorithm

摘要

Microphone arrays provide spatial resolution which is useful for speech source separation, due to the fact that sources located in different positions cause different time and level differences in the elements of the array. This features can be combined with time-frequency masking in order to separate speech mixtures, by means of clustering techniques, such as the so-called DUET algorithm, which uses only two microphones. However, there are applications where larger arrays are available, and the separation can be performed using all these microphones. An speech separation algorithm based on mean shift clustering technique has been recently proposed, using only two microphones. In this work, the aforementioned algorithm is generalized for arrays of any number of microphones, testing its performance with echoic speech mixtures. The results obtained show that the generalized mean shift algorithm notably outperforms the results obtained by the original DUET algorithm.