Penalized nonnegative representations for speech separation

RUBEN D. SPIES; LEANDRO E. DI PERSIA; FRANCISCO J. IBARROLA

Río Cuarto, Córdoba

Congreso; VII Congreso de Matemática Aplicada, Computacional e Industrial; 2019

Asociación Argentina de Matemática Aplicada, Computacional e Industrial, ASAMACI

In this work we address the problem of supervised audio source separation within a reverberant environment. We make use of a nonnegative representation in order to model the mixture along with reverberation. This kind of models often pose the problem that the number of variables to learn is large with respect to the data, which is to say there are many possible choices of the elements that result in the same approximation of the mixture. We use a probabilistic approach in order to derive a penalized cost function that aims to overcome this issue by inducing a certain structure over the representation elements. Preliminary results account for a considerable improvement in restoration quality with the introduction of penalizers.