Perceptual Constraints inSystems forAutomatic MusicInformation Retrieval

Perceptual modelsanddataprovide use- fulguidelines inthedevelopment ofsystems formusic information retrieval. Weshowthatformusictempoex- traction, similarity evaluation, andstructure extraction, perceptual dataarenotonlyuseful intheevaluation of automatic systems butalso intheir design. I.INTRODUCTION Recent advances inaudio compression technology andin- creases instorage capacities andbroadcast bandwidths al- lowusers toaccess vast (andgrowing) amounts ofmusic au- dio. Withthis capability comesanincreased needfortools tonavigate, browseandsearch formusic.Traditionally, suchtools havebeenbasedonannotated metadata, such asgenreandartist, butrecent developments inautomatic musicinformation retrieval (MIR)aremaking itpossible to extract sometypes ofmetadata directly fromthemusicau- diosignal itself (1). Theseadditional metadata, combined withtraditional annotated metadata, provide amuchricher setuponwhichsophisticated systems formusicnavigation canbebuilt. Whilemanyaspects ofalgorithms forMIRcanbecon- strained by,ordefined intermsof,musicological, socio- logical, and/or computational factors, wehavefoundthat perceptual criteria areoften valuable guides. Thisistrue notonlyfortheevaluation of,butalso forthedevelopment andconstruction ofMIRsystems. Whiletheadvantage of using perceptually-based constraints depends somewhat on theapplication context, wepresent here three cases inwhich perceptual knowledge clearly benefits orisneeded forthe development ofalgorithms. Morespecifically, wehaveex- amined automatic extraction ofmusictempo, musicsimi- larity ratings, andstructural boundaries inpopular music.