There is a lot of ongoing research in the field of music retrieval aiming to improve retrieval results for queries posed as text, sung, hummed or by example as well as to automatically tag and categorize songs. All these efforts facilitate scenarios where the user is able to somehow formulate a query – either by describing the song or by giving examples. But what if the user cannot pose a query because the search goal is not clearly defined?
Sometimes, well rather often, when I browse my iTunes, I cannot explicitly state what I’m looking for.
I start of with this one song stuck in my ear, maybe Rollerskater by Mathias Aguayo, and what next? Ah, Dave Aju fits after this, and then, maybe another vocal-beatbox tune, …
This is where the usual music retrieval systems fail.
They are developed to browse a collection in order to find a specific song, or to get offers of songs that are connected to my initial starting point in most obvious ways like same artist, same genre, same tag.
Some come with a more elaborated surface, or more tags or some random variability of the results presented.
The problem is, that all search engines assume that you know what you want, and rely on a limited set of obvious tags. Thus they offer a very limited set of result options based on comparing basic tags only.
So even the most elaborate retrieval systems are just as good as the meta data of the song.
Bottom line is we still scroll through endless lists.
What would be appreciated in order to finding other similarities would request what interface designers call ‘similarity-preserving projection of objects (tracks, albums or artists) onto the (typically two-dimensional) display space’.
Users are very different in their approach – a musician might be looking for similar instrumentations, someone else for music for a certain mood or scenario, others want to have playlists NOT based on similarity at all, like John Peel did play Glen Gould with Bad Brains in one show, and it went well together, without the slightest match in mata data…
A music retrieval system should be able to incorporate this subjectiveness in order to better comply with the individual needs of its users.
Otherwise manual handling is far easier and better.