Sound object separation based on the principle of sparsity

• Grant number: 19/22795-1
• Support Opportunities: Scholarships in Brazil - Doctorate (Direct)
• Start date: March 01, 2020
• End date: February 28, 2023
• Field of knowledge: Engineering - Electrical Engineering - Telecommunications
• Principal Investigator: Bruno Sanches Masiero
• Grantee: Arthur Nicholas dos Santos
• Host Institution: Faculdade de Engenharia Elétrica e de Computação (FEEC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Associated research grant: 17/08120-6 - 3D-Audio. Sampling, encoding and reproduction of spatial audio, AP.JP

Abstract

The determination of the amount and direction of sound sources can be made, at first, by using spatial filters to estimate the sound level reaching an array of microphones from different directions, similar to the estimation of acoustic images. This technique, however, usually has low resolution due to the relatively small number of microphones that make up the arrangements. Several methods have been proposed to increase the quality of acoustic image estimation without increasing the number of sensors, which apply deconvolution techniques to eliminate the effect of the array spread function. A more recent alternative is the use of sparsity-promoting regularization, which uses a 1 norm regularization to solve the problem of determining sound objects. These algorithms work by giving greater weight to sparse scenes (which is often a valid assumption for scenes composed of a small number of sound sources), promising a better separation of the present objects. Now, if we apply these sparse-promoting algorithms to a sound scene made up by a few sources but recorded in a very reverberating environment, the results will not be so encouraging, as for these algorithms reverberation will reduce the sparse sound scene. To improve the separation of sound objects under these conditions we propose the use of an enhancement step, which discards superfluous information for our perception, in this case, reverberation. The proposed way to perform this enhancement step is to use the psychoacoustic concept of masking, to discard all information that will not be perceived by the listener, performing a kind of "de-reverberation" to assist sparsity-promoting algorithms in their task of separating the sound objects. (AU)