June 2024
·
66 Reads
Recent advances in music source separation through deep learning algorithms have enabled the isolation of vocals from recordings where separate tracks are not available. We have used Demucs (v4) Music Source Separation (Défossez, 2021; Rouard et al., 2023) to obtain isolated vocal parts from song recordings by Vasilis Tsitsanis, a prominent songwriter of the urban popular Greek idiom Rebetiko. We then used the Matlab library, Timbre Toolbox (Kazazis et al., 2021) to extract a set of audio features (e.g., spectral centroid, Tristimulus values, odd even ratio, inharmonicity, etc.) from manually edited and retained vocal phrases and sustained vowels. The sustained vowels were additionally analysed in terms of their vibrato characteristics and formant values using the voice analysis software Praat (Boersma and Weenink, 1992-2024). This poster comparatively examines the audio features extracted from four iconic voices of Rebetiko, the vocally related male singers Prodromos Tsaousakis and Stelios Kazantzidis and the quite distinct female voices of Marika Ninou and Sotiria Bellou. We demonstrate how such features can be used to quantify vocal qualities and highlight differences and similarities, not only between separate singers but also across various stages of the same vocalist's career.