Scholarship of Discovery
A well-known signal processing issue is that of the “cocktail party problem”, which refers to the need to be able to separate speakers from a mixture of voices. A solution to this problem could provide insight into signal separation in a variety of signal processing fields. In this study, a method of vocal signal processing was examined to determine if principal component analysis of spectral data may be used to characterize differences between speakers and if these differences may be used to separate mixtures of vocal signals. Processing was done on a set of voice recordings from 30 different speakers to create a projection matrix which could be used by an algorithm to identify the source of an unknown recording from one of the 30 speakers. Two different identification algorithms were tested. The first had an average correct prediction rate of 15.69%, while the second had an average correct prediction rate of 10.47%. Additionally, one principal component derived from the processing provided a notable distinction between principal values for male and female speakers. Males tended to produce positive principal values, while females tended to produce negative values. The success of the algorithm could be improved by implementing differentiation between time segments of speech and segments of silence. The incorporation of this distinction into the signal processing method was recommended as a topic for future study.
Fitzgerald, Bradley, "Vocal Processing with Spectral Analysis" (2018). Honors Program Projects. 92.