Project Description:

In this work we explore the task of distance-based speaker separation performed on the Poincaré Ball model. The recent advent in audio-based task exploiting non-Euclidean manifolds has shown great promises towards effectively modeling some of the semantics found in the audio mixtures projected onto these type of space.

We first formulate the separation task by assuming two hierarchical levels in the input mixture; the parents (i.e. near and far fields) and the children (i.e. the single speaker pertaining to each of the parent classes). Beside offering a consierable improvement over its Euclidean counterpart, we show that the Poincaré model successfully model some of the distance-based semantics pertainign to the taks at hand, such has speaker-to-microphone, speaker-to-speaker distances, or speaker density. Checkout our ICASSP paper linked above.