Application of independent component analysis for identification of speakers from speech recordings

U ovom radu je na snimke više govornika primijenjena metoda nezavisnih komponenata (FastICA algoritam u Matlabu) u svrhu izdvajanja izvornih signala. Kada bi se izvorni signali mogli uspješno izdvojiti iz bilo koje mješavine signala to bi mogli primijeniti na niz različitih problema, npr. omogućili bi slušnim aparatima da raspoznaju različite govornike i fokusiraju se na određenog govornika, što je poznato pod nazivom 'cocktail party problem'. Dobiveni rezultati pokazuju da ova metoda funkcionira na snimkama gdje je broj govornika jednak broju mikrofona uz pretpostavku da su izlazne vrijednosti mikrofona linearne mješavine nezavisnih komponenti. Ljudski govor je sam po sebi jako kompleksan, a to posebno dolazi do izražaja kad istovremeno govori više govornika. Otkrivanje mehanizma selektivne pozornosti u mozgu je ključ za rješavanje 'cocktail party problem'.In this paper was applied independent component analysis (FastICA algorithm in Matlab) on the records of more speakers for the purpose of extracting the original signals. If the source signals could be successfully extracted from any signal mixture they could apply to a variety of different problems, for example, allow hearing aids to discern different speakers and focus on a particular speaker, known as the 'cocktail party problem'. The results obtained show that this method works on experimental recordings where the number of speakers is equal to the number of microphones assuming that the output values of the microphone are linear mixtures of independent components. Human speech is very complex by itself, and this is particularly pronounced when multiple speakers speak simultaneously. Detecting a selective attention mechanism in the brain is the key to solving the 'cocktail party problem'