Extracting speech signals by suppressing background noise using machine learning-based methods

U ovome radu, u prvome poglavlju napravljen je kratki uvod na tematiku, opisani su osnovni pojmovi i definicije. U drugome poglavlju detaljnije su opisani opći načini potiskivanja pozadinskih šumova, dan je pregled gotovih programskih rješenja za potiskivanje pozadinskih šumova te je dan kratki uvod u tradicionalne metode i metode temeljene na strojnom učenju. U trećem poglavlju, detaljnije su opisane tradicionalne metode i metode temeljene na strojnom učenju. Kod tradicionalnih metoda detaljno su opisani principi rada Wiener filtra te spectral gating pristupa – spectral gating je odabran kao tradicionalna metoda u eksperimentalnom dijelu. Kod metoda temeljenih na strojnom učenju detaljnije je opisan opći postupak rada metoda te su detaljno opisane povratne neuronske mreže i PoCoNet arhitektura koja je bila odabrana za metodu temeljenu na strojnom učenju u eksperimentalnom dijelu rada. Četvrto poglavlje daje pregled tehnologija danih u radu, tehnički opis uzoraka i skupa podataka, detaljan opis metrika za evaluaciju kako za objektivnu tako i za subjektivnu metriku, te su na kraju prikazani rezultati objektivne i subjektivne metode. U petome poglavlju dan je kratki uvod u mogućnost implementacije modela metode strojnog učenja na mobilnim uređajima.In this paper, in the first chapter, a short introduction to the topic is made, the basic terms and definitions used in the paper are described, as well as the presentation of sound recordings in both the time and frequency domains. In the second chapter, general methods of background noise suppression are described in more detail, an overview of ready-made software solutions for background noise suppression is given, and a brief introduction to traditional methods and methods based on machine learning is given. In the third chapter, traditional methods and methods based on machine learning are described in more detail. With traditional methods, the working principles of the Wiener filter and the spectral gating approach are described in detail - spectral gating was chosen as the traditional method in the experimental part. In the case of methods based on machine learning, the general procedure of the methods is described in more detail, details of recurrent neural networks and the PoCoNet architecture which was chosen for the method based on machine learning in the experimental part of the work. The fourth chapter provides an overview of the technologies presented in the paper, a technical description of the samples and data set, a detailed description of the evaluation metrics for both objective and subjective metrics, and finally the results of the objective and subjective methods are presented.. In the fifth chapter, a short introduction to the possibility of implementation on mobile devices is given