A connectome is a comprehensive map of neural connections in the brain, and may be thought of as its “wiring diagram”… A connectome is constructed by… mapping where neurons are connected through synapses…
Hmm, what about mapping connections between music, songs, artists, even albums?
In music, as in synapses:
There are strong natural constraints on which neurons or neural populations can interact, or how strong or direct their interactions are.
In music these constraints are tonal, mappings to scales, desirable chord progressions, etc. Obviously you can go from any note to any other, but if you do that planlessly enough, you’ll lose your audience pretty quickly. The rules of what (generally) sounds good are well enough understood, and the entire field of “Music Theory” exists for that purpose. But simply learning music theory doesn’t let you make good music, and indeed “good” is also highly subjective, so clearly something more is needed to understand not only what is good, but what is good for each person.
What better way to approach this than to ask them?.. Sort of.
In short, this is an attempt to create a better recommendation model for music based on the combination of more granular and specific feedback mechanisms, and correlation of that feedback across a large database of music. The basic way this could work is well-exemplified by SoundCloud, which lets you “react” and comment on specific parts of music by time signature. So rather than a coarse, song-level understanding of who likes a given song, you can potentially understand who likes individual “elements” of a song. The hope is that this then would let you determine more complex relationships between different songs, artists, etc.
A simple example of this in action would be if the chorus of one song gets a lot of Likes and/or positive Comments from people with otherwise disparate musical taste, then perhaps the chorus of other songs that each of those people individually Like could be recommended to the others in that group who haven’t already listened to those songs. This could potentially uncover commonalities between songs that are otherwise difficult to identify and articulate.
This is a project with (theoretically) similar goals to the Music Genome Project which gave birth to Pandora. Essentially, to “take apart” music and figure out what makes it work, what makes people like a certain song, artist, band, etc. However, instead of the “experts” approach of the MGP, this would be based on crowd sourcing. Although crowd sourcing is not an all-purpose tool, when it comes to determining what people like and, potentially, why, a good place to start is asking people what they like.
The Music Connectome Project (MCP) would seek to collect information about people’s likes and dislikes, especially for specific songs and their components (e.g. vocal, beat, etc.), and potentially collect similar relevant information about bands, albums, etc. This would be accomplished through a free-to-access website with a song rating and comment system.
This project is somewhat inspired by the current popularity of music “suggestion” services, from the early examples like Pandora, Last.fm, etc. to more modern and popular options like Spotify and its “Discovery Weekly”, etc. From my experience so far all the currently available options fail to make good suggestions for more complex, nuanced, or specific starting criteria. For example, let’s say you like one or two Beyonce songs, but you don’t like all Beyonce songs, nor songs by Destiny’s Child (of which Beyonce is a former member). How does Spotify determine what songs to play if you enter Beyonce’s “Halo” as a playlist seed, given this unique preference? Spotify does not allow you to specify the things you like or dislike about a particular song or artist, so it will naturally assume you like the artist, which in many cases is actually not true. Musical taste is more complex than this, especially with the focus on singles releases, popularity charts, etc. The ability to better communicate one’s preferences and musical taste is also critical when considering more complex or varied artists, e.g. Pink Floyd or The Kinks, or when considering “One Hit Wonders”, e.g. The Flys (“Got You Where I Want You”).
I should say that Spotify now does a better job of recommendation than Pandora did/does. But it still feels like it could be notably improved upon. In particular I think a better understanding of why I like a song could help me avoid “negative” recommendations, i.e. songs I definitely don’t like, which still turn up often in my Discovery and other Spotify recommendations.
Putting something like this together could be quite challenging, requiring a large user base and data set to actually generate useful, novel recommendations. Ideally there would be a starting point with a lot of existing data. Rights issues are also a massive challenge across the media playback space, which makes it basically infeasible to start a new, stand-alone service without large amounts of startup capital.
Leveraging existing platforms and licensing deals is one potential way around this. Primarily if there’s an SDK, which Spotify does have:
I’d need to look into whether I could create a “comment on point in track” functionality with it.
Also for testing the idea, using an existing database of time-correlated comments would be fantastic. It just so happens SoundCloud has that, and an API that lets you access them.
This is an older list from 2017 which may need to be updated for my current thinking about this concept.
- Browsable catalog of artists and songs
- Links to where songs or song samples can be played
- Each song has groups of components that can be rated
- Ratable components are separated into categories, e.g. Song Writing (lyrics, instrumentation), Vocal Performance (vocal quality, “soul”, etc.), Instrumental Performance, Arrangement, Production
- Each category would have items within it as above, e.g. lyrics and music are sub-items of the Song Writing category
- When a category only has one item perhaps it should not show the category, only the item; may need to prefix item with category name for clarity anyway
- Items that can be rated start simply and can be added to by visitors
- Begin by presenting raters with only a basic set of elements (e.g. vocal quality, lyrics)
- Allow raters to add items, both from a list of pre-defined items, as well as their own custom items (e.g. “synth hook”)
- Pre-defined items that would not show by default would be e.g. “arpeggio”
- Should have a complete library of basic musical analysis terms to be added as desired by raters
- Custom items would show up initially for e.g. 1 month but would then be removed if they did not get enough votes as this would demonstrate the item is not relevant to the majority of raters; these could be shown in a “show all items” option perhaps
- Users would be able to add description/explanation to their items, e.g. “the saw-type synth hook during the chorus right before the vocal begins”, and perhaps links could be possible to specific pre-authorized sites, e.g. wikipedia (to prevent link spam)
- Ideally a SoundCloud-like time-specific comment/note system. If this could work with a plugin/back end of a licensed or ad-supported music service, that would be awesome!
- Embedding sound clips would be nice, but has potential copyright issues
- Ratings should be simple, like or dislike radio buttons, with possibly a “neutral” option
- Site requires login to rate
- Users could add an overall comment to a song after they rate it; the comment would be for the whole song and they could include references to each of the points they rate if they choose, but they would not be able to comment on each rated item (at least initially)
- Ratings would be aggregated and anonymized and any song with more than 5 ratings on an item would allow for viewing of aggregate rating information
- Facebook connect
- Lots of info about musicians/songs on item pages?
- Allow import from other music suggestion/rating sites that you may have an account with, e.g. iLike, that way users can get a quick start with their music ratings