'Efficient Retrieval from Vast Music Collections'
Authors
Jensen, Claus Åge ; Mungure, Ester Moses ; Sørensen, Kenneth Rand
Term
4. term
Education
Publication year
2006
Abstract
Efterhånden som musik bliver digitaliseret, opstår nye udfordringer i Musik Information Retrieval (MIR) – feltet, der handler om at finde og organisere musik i meget store samlinger. Vi præsenterer og evaluerer rammeværket Music On Demand, hvor sange forespørges som en kontinuerlig strøm. Lytteren kan løbende påvirke de næste numre ved at vælge at spille lignende sange, afspille tilfældige sange, springe over, begrænse samlingen eller angive en bestemt del af samlingen. For at understøtte dette definerer vi en generisk musikdatamodel og tilhørende søgefunktioner. Vi bruger bitmap-indekser til at indeksere metadata (fx kunstner eller genre) og mål for musikalsk lighed udledt direkte fra indholdet, så vi kan lave hurtige opslag med simple bitvise operationer. Selve søgningen kombinerer både metadata og indholdsbaseret lighed. I denne sammenhæng undersøger vi Word-Aligned Hybrid (WAH)-komprimering og tekniksen Attribute Value Decomposition til at repræsentere indholdsbaseret lighed. Eksperimentelle resultater viser, at vores implementering giver effektiv adgang til meget store musikbiblioteker, med kun en lille ekstra pladsomkostning ud over selve musikfilerne.
As music becomes digital, new challenges arise in Music Information Retrieval (MIR), the field that focuses on finding and organizing music in very large collections. We present and evaluate the Music On Demand framework, which treats song queries as a continuous stream. Listeners can steer upcoming tracks in real time by choosing to play similar songs, play random songs, skip songs, narrow the collection, or specify a particular subset. To support this, we define a generic music data model and associated query functions. We use bitmap indexes to store metadata (such as artist or genre) and measures of musical similarity derived from the audio content, enabling fast lookups with simple bit-wise operations. Retrieval combines both metadata and content-based similarity. In this context, we examine the Word-Aligned Hybrid (WAH) compression scheme and the Attribute Value Decomposition technique to represent content-based similarity. Experimental results show that our implementation provides efficient access to very large music libraries, with only a small additional storage cost beyond the music files.
[This abstract was generated with the help of AI]
Documents