Music to My Ears: Melody, Memory and Mathematics

Some more recent tools, like Last.fm and Pandora, try to be a bit more sophisticated in devising their recommendations. Pandora is based on the Music Genome Project, which analyzes music according to its attributes:

Pandora’s music analysts check off which of 400 attributes each individual song has, and then, as you give thumbs up or down to various songs in its library that you listen to, it recommends other songs that have the same or similar attributes. Pandora  is available only to Americans, and I am not prepared to invent a US zip code to cheat the system (though it would be easy to do so), so I can’t really say how well it works, but before it forced me to register (and then told me, as a Canadian, I couldn’t register) its recommendations were just awful. What’s worse, you can only give a quick ‘thumbs down’ to so many songs before it forces you to listen to some from start to finish, in accordance, apparently, with its licenses. Extremely frustrating, hugely labour intensive (every song they add has to be ‘expertly’ analyzed and tagged), and it fails to recognize that we love songs as much for the context in which we first and most often heard them as for their analytical attributes like “hard rock features, acoustic rhythm piano, varying tempo and time signatures, mixed acoustic and electric instrumentation and paired vocal harmony”.

Last.fm relies more on its members than musicologists, and because it isn’t limited to licensed music, it can recommend artists and songs that your ‘neighbours’ (people whose song lists correlate most closely with yours) like, even if they’re obscure. And it can be programmed to monitor your iTunes listening, so you don’t have to tell it what you like and what you’re listening to — if you listen to more than half of a song, it assumes you like it, and logs it. The sheer number of members and songs it correlates allows it to use neural rather than analytical logic to make recommendations, but it relies heavily on the frequency of listening to artists rather than to specific songs, and most artists produce a wide variety of songs. It does have an intriguing more popular/more obscure ‘slider’ that you can tweak to get rid of the Beatles, U2 and Madonna.

So far it has logged 1650 songs I have listened to, and recommended about 500 artists I might like. Some of these came from ‘groups’ I joined of people who ostensibly tend to like the same types of music I do — these recommendations were useless, zero for 150, to the point I unjoined several groups so I wouldn’t have to wade through any more. The rest of the recommendations came from correlations either with other members or with ‘tags’ — members can tag the genres of music they listen to, so the recommender can get a sense of what tags you tend to prefer, and weight its recommendations towards artists similarly tagged. These tags are autopoietic (folksonomic) — they evolve as members pick new tags and abandon old ones. So for example last.fm tells me I tend to like music tagged as ‘shoegaze’ genre.

Of the 350 artists recommended to me so far using these neural heuristics, I have liked at least one song of sixteen of them:

• Female Singer-Songwriters: Imogen Heap (UK), Anna Nalick (US), KT Tunstall (UK), Inara George (US), Lucie Silvas (UK), Zero 7 (UK), K’s Choice (Belgium), Nina Gordon (US), Alex Parks (UK), Lene Marlin (Norway), Nerina Pallot UK)
• Folk Singers: Nickel Creek (US)
• New Age/Instrumental/Acoustic/Soundtrack Artists: William Ellwood (US), Thomas Newman (US)
• ‘Supergroups’ that I didn’t know about that include artists on my playlist: Frou Frou (UK) and The Wreckers (US)

The recommendations also include a few African, Latin, and Classical artists (all disproportionately represented on my playlist), but none of them worth a second listen. None of these genres is well represented in last.fm’s massive database. Very few Canadian artists have been recommended either, possibly because with 80 Canadian musicians (two thirds of them women) in my playlist already, there may not be a lot more out there to recommend — though I’m dubious of this.

In five of the 16 cases, last.fm has been able to play me 30-second cuts or even (through its Recommendations Radio station selecting music randomly from artists it has recommended to me) full-length samples of these artists’ work. For the other 11, I had to Google to find the artists’ home pages and listen to samples there, or use the more extensive MySpace artist page samples, or (perfectly legal in Canada, per our Supreme Court) file-sharing services, to hear these artists’ music.

For someone with tastes as peculiar and fussy as mine, 16 out of 350 is pretty good — My MP3 player has only 800 songs by 300 different artists on it, spread over about 40 years of listening, so to have found 16 new artists I like in just three months is remarkable, despite the investment required in finding music of and listening to 334 crappy artists to find the 16. The recommender has ‘learned’ to send me a heavy weekly dose of ‘female singer-songwriters’ (a tag connected with much of my playlist), and a disproportionate number of UK & European artists (perhaps because last.fm is UK-based and has so many members there).

In contrast, the now-defunct Rock Chicks Radio introduced me to only 4 new artists I really like, after a similar investment in listening time. But I’ll miss RCR — they played a better mix of music than any off-air, Internet or satellite radio station out there, even though I knew almost all of the artists they played.

Last.fm is far from perfect, however. I suspect because most members are young, listening to about 20% 1960s music quickly stereotyped me, and my identified ‘neighbours’ soon included lots of Canadians in their 50s, even though most of what I was listening to was contemporary. My ‘neighbours’ seem mostly rooted in the past, and just a few oldies got me lumped in with them, and produced recommendations of some truly dreadful bands of that era. I’ve found it more fruitful to listen to ‘tag radio’ — music tagged with ‘female singer-songwriters’, ‘acoustic’, or even ‘shoegaze’. I’m disappointed at the lack of depth of African, Latin, Classical, Folk, New Age and Soundtrack selections and recommendations. Yahoo Launchcast seems to be pretty good at letting you listen to music of selected genres, though I think it’s outrageous it doesn’t work with Firefox.

But none of this gets at the Three M’s that, I think, really determine whether you will like a song or not: Melody, Memory and Mathematics. The composition of a song goes far beyond ‘attributes’ — just like a bird’s songs, each song is unique, and will appeal to different people, or not, depending on how it resonates, emotionally and intellectually, with the listener, and largely independent of who sings it or what instruments are used to play it. I like both the Doors’ original version and Jose Feliciano’s cover version of Light My Fire — utterly different, but still essentially the same song, the same composition. I think we are ‘programmed’ to just like certain songs — it’s an evolutionary thing. I also think we are ‘programmed’ to like certain voices — which doesn’t mean we like everything we hear in those voices, but rather that we are predisposed to like songs by certain artists because the tonal quality resonates with something inside us.

Memory also plays an enormous role in whether we like a song or not. A song is essentially a story, and we will like songs that we remember in a positive context (e.g. initially or most often heard while we were doing something we loved) far beyond their musical merits. If the lyrics, or even the name of a song or a singer, evoke a certain memory, our like or dislike of the song will be tainted by that memory.

And the mathematics of a song also speak to us, and are either consonant or dissonant with the mathematics of our brains and bodies. Certain chords and harmonies (major sevenths, minor ninths, suspended seconds) strike me, for example, as poignant, fraught with meaning, simply because of the way the overtones of the notes hit my eardrum together and rumble along the neural pathways to my brain. Likewise rhythms and the pacing of songs (the staggeringly complex rhythms of African Zoukous music, the soothing flow of samba, the teasing offbeat of merengue, or the halting pace of the adagio of Ravel’s Concerto in G) either fit, or don’t fit, with the syncopation of our souls, and that, rather than intellectual discernment, will determine whether we love a song with those ‘time qualities’ or loathe it.

So, composition, tonal quality, memory, chords & harmonies, rhythm & pacing, all determine how a song will be received by our musical ‘taste buds’. And how these things all ‘work together’, the result of effort of the producer more than the artist (the song’s ‘production values’) is also important. It is all about the senses, and about chemistry. Whether we like a song or not is not really our choice, as much as we would like to credit or fault our intellectual appreciation for our judgement. There are those who, like wine ‘connoisseurs’, will proclaim one selection ‘superior’ to another, and tell you why, but they are merely proclaiming their own filters, frames and prejudices, telling you not about the singer or the song but about them, the critic. Sheer vanity.

This is all about complexity — there are too many variables to analyze or predict in any useful or reliable way. Using neural approaches, as last.fm has done, is the right way to try to tackle the problem, but it’s like doing surgery with a spatula: the instrument is really not up to the subtlety of the task. I can sigh and say that finding 16 needles in a haystack of 350 is at least a vast improvement over the one-in-a-thousand I was batting scanning through the radio spectrum. But, just as I know the cure for some horrific diseases lies buried in some tiny plants in the dwindling rainforest, so I know there are songs out there that would define me, change me, transport me, pull me out of the darkness, or show me thetruth of the universe — if only I could find them.

Image: Representation of a major seventh chord by Stewart Butterfield, co-founder of Flickr.