'Novel Approaches in Audio Similarity - NCD; Amplitude Pattern Mining'
Authors
Adegboye, Adesola ; Li, Ruobing ; Zhou, Ruoran
Term
10. Term
Education
Publication year
2006
Abstract
Efterhånden som digitale musiksamlinger vokser, har vi brug for værktøjer til at finde sange, der lyder ens. Denne afhandling undersøger to måder at måle lydlig lighed. Den første er Normalized Compression Distance (NCD), en metode uden håndlavede træk, der sammenligner, hvor godt to lydfiler kan komprimeres sammen i forhold til hver for sig. Den anden er Amplitude Pattern Mining (APM), som ser på ændringer i lydstyrke over tid for at finde tilbagevendende mønstre og sammenligne numre. Resultaterne viser, at NCD ikke er praktisk til musik, fordi almindelige kompressorer ikke pålideligt fanger musikalsk lighed. Til gengæld er amplitude et lovende signaltræk, og APM klarer sig godt. Med disse indsigter bygger afhandlingen modeller, der kan generere playlister ved at gruppere lignende sange.
As digital music collections grow, people need tools to find songs that sound alike. This thesis explores two ways to measure audio similarity. The first is Normalized Compression Distance (NCD), a non-feature method that compares how well two audio files compress together versus separately. The second is Amplitude Pattern Mining (APM), which looks at changes in loudness over time to find recurring patterns and compare tracks. Findings show that NCD is not practical for music because common compressors cannot reliably capture musical similarity. In contrast, amplitude is a promising signal feature, and APM performs well. Using these insights, the thesis builds models that generate playlists by grouping similar songs.
[This abstract was generated with the help of AI]
Documents