Individual violin sound identification using audio features and machine learning

Identifying individual instruments of the same type from their recordings is a rarely addressed challenge.This paper explores violin sound identification using two datasets of recordings, featuring multiple violinists playing multiple violins.Several long-term audio features were compared, and their performance in violin classification was evaluated using classical machine learning algorithms.Among these features, long-term MFCCs demonstrated the ability to distinguish individual violins while taking into account player-induced variability, enabling reliable violin recognition.Additionally, the influence of key parameters -such as the number of violinists and choice of musical excerpts -on classification performance was analyzed.These findings offer guidelines for optimizing future data collection aimed at capturing a violin's unique sound signature.