Musical signature identification and plagiarism detection through feature extraction and analysis

Seow, Yi Xuan (2025) Musical signature identification and plagiarism detection through feature extraction and analysis. Final Year Project, UTAR.

PDF Download (2128Kb)

Abstract

Music information retrival (MIR) systems are introduced to counter musical borrowing and unintentional plagiarism in the present music industry. Music plagiarism detection is a complex task, as similarity may arise not only from melodic lines but also from rhythm, harmony, and timbre. This system introduces an approach that integrates multi-dimensional audio features with segment-based analysis to improve detection accuracy and interpretability. Extracted features include harmonic descriptors (CENS, CQT, tonnetz, harmonic n-grams), rhythmic descriptors (tempogram, onset autocorrelation), timbral descriptors (MFCC), and spectral texture measures (spectral contrast, centroid, rolloff, bandwidth, flatness). Database songs are segmented into overlapping windows of varying lengths, while query tracks are segmented consistently, allowing robust local-to-local comparisons. A weighted similarity model balances contributions from all features and normalizes results even when certain descriptors are unavailable. Potential plagiarism is flagged when segment similarities exceed adaptive thresholds, and is presented in a ranked list. To support expert evaluation, the system employs large language models (LLMs) to generate textual analyses of the top matches, highlighting how melodic, rhythmic, harmonic, and timbral evidence supports or weakens plagiarism claims. The aim is to provide music experts with clear, evidence-driven insights, enabling more efficient and transparent decision-making.

Actions (login required)