I realised recently that most media recommenders don't work for me because they are based on genre. If you like The Watchmen film, you're sure to like Marvel films because they're also action, and based on comics. We see you listened to Pearl Jam a lot, you're going to get a kick out of Nirvana.
That's just not how I work. I like _good_ stuff, I don't confine myself to genres, I'll go anywhere if it's half decent. I want a service that says hey, you listened to Benny Goodman's Sing Sing Sing, you should check out The Thirteenth Floor Elevators' first album.
The only recommendation that kinda works for me is based on an aggregation of what other people listened a lot to. Sure on small data sets that is probably going to be genre biased, but on large sets it can start working and break out of that.
I am like this also. It's tricky though because I think people hear 'quality' in different ways. Some people focus on lyrics, some like cultural identity, some people are probably responding at a more 'primal' level, i.e. just to the pure energy / emotion. But I do think it is different for everybody and does bear some relation to the individual's level of musical awareness and how much time and effort they invest in cultivating their own taste. I once worked with someone who when asked "What kind of music are you into?" answered "I don't really listen to music". Shocking!
The best way I have found to find new stuff is just to listen to lots and lots of music, and follow different leads. It's hardy a chore doing so!
"I don't really listen to music" == "I don't breathe much".
Way back I had friends over, one of them was looking through my CD collection which was minimal at the time, maybe 60 CDs. "Wow, you have a lot of CDs, do you listen to all of them?"
> "I don't really listen to music" == "I don't breathe much".
Could you elaborate?
I definitely used to be a person who just didn't listen to music - I found it quite boring, really. I've since found genres that I like that I listen to on occasion, but definitely still not as frequently as the average person.
This wasn’t meant as a judgement whatsoever. Others would say the same about, say football, which I just don’t relate to.
The comment was relating my experience with music, it’s like I can barely live without it.
That's why "quality" radio can be so powerful: it's not uncommon for DJs to curate a set based on "songs with the same musician" or "original song and remakes/samples from the original" and the like.
Not that it's guaranteed to be "good" stuff but at least unique and more memorable in my experience
Any recommendations? I listen to BBC 6 Music a lot. Tried various others but nothing really 'stuck'. Always fun to have a fly around on https://radio.garden also. KEXP have some great performances on youtube, I should try listening to the radio station probably. Although saying that, neither 6 Music or KEXP (or others I have tried) seem to go quite obscure enough for me. i like pop / rock / indie but also experimental rock, lo-fi, noise, etc
I made a 6 Music 'now playing' app (https://radio-tracker.geoffclayton.com/) so I could easily keep track of things I liked. I kept hearing things then forgetting... I never _quite_ finished it, so never really advertised its existence, but over the course of year it has collected some interesting stats. It's there and it works, anyway! Mainly the Spotify integration part is not quite finished.
I used to listen to bbc 6 music a lot. These days I listen to fip.fr it’s a French national radio, that pans all genres, and is very well curated. There is no ads, and very few people talking.
I'm a big fan of Chile based Radio Isla Negra (available in Slow beat and upbeat as well as the regular eclectic selection on a well curated playlist.)
No Ads, No DJs, just an occasional jingle. Much of it is just "new to me" rather than strictly "new", but it's one of the new I can enjoy leaving on for hours (or days) at a time without finding it irritating enough to change stations.
Yeah, a bare minimum for this is probably to give at _least_ two songs and not just one. What is "good" is still subjective but given a few more data points and the gist of those enjoying all those should also be able to better provide what else these users enjoy.
Back when humans still did work, the Music Genome Project had people listen to songs and try to break them down into little "genes" of metadata, everything from "vocal harmonies" to "minor chords" to "strong lyrical presence". It worked way better than the automated systems, IMO, and gave rise to Pandora: https://en.wikipedia.org/wiki/Music_Genome_Project?wprov=sfl...
I wish other music apps could license their recommendation system. Still nothing seems to come close.
In my personal experience I gave up on Pandora. It was regularly suggesting me music that wasn't at all similar to the song I started on. In fact it was consistently worse at finding me similar music than just plain Spotify radio (or playlist enhancer) and/or YouTube Music radio.
Haha, same. I used to work in a music equipment rental warehouse that was staffed by musicians and all day we would be curating and training our favorite Pandora stations. If a bad song came on the person closest to the computer would scan the warehouse to see how many of us were giving that song a thumbs-up or a thumbs-down and train the algorithm according to our votes.
The trouble was, despite our diligent training and curation of many Pandora stations, any station remotely related to rock would eventually converge on Morrissey. It got so bad that whenever a Morrissey song would play you could hear shouts of protest from across the warehouse and it would be immediately downvoted with extreme prejudice.
Nevertheless it would keep feeding us Morrissey so the joke among us became “All stations lead to Morrissey”.
I feel like many (most?) people who used Pandora extensively have a similar story, although curiously it never seems to be the same band. My fixed point was Radiohead's Karma Police, specifically. I wonder if your personal Pandora fixed point says something about you...
To be very fair, there's nothing plain whatsoever about Spotify's recommendations. Not only do they have the largest corpus of training data, their models are hard to beat.
I wish they had a way for me to say I'm not interested in something as well. For years there's been a few of the same albums recommended to me every day. Like, I've ignored them for this long and it's clear I'm not interested but the space for things I might actually listen to is still being wasted.
What's worse is that they're not even obscure albums. Like, if I was ever going to get into the Eagles I would have already done so, I don't need to "discover" them.
It was actually the other way around. It would play me songs technically within the same broad genre (Hip Hop, RnB, Pop) but wildly different sound.
It felt less like a music genome project and more like this typical nonsense streaming platforms do, "people who listen to song A, also tend to listen to song B".
Yes, so amazingly labor-intensive. Now I have found some programs which boast of a feeble seventy percent accuracy in identifying the key of a song (if you were going to use the Camelot System to build a mix).
Hm... I never thought that would be extremely useful but I think I have that working close to perfection on piano (I need it for audio->midi->score conversion), I should probably try it on other styles of music as well.
So, I am told -- and could be wrong -- that the subtle art of DJing requires many little tricks in tandem. Some are obvious, like similar tempos, similar "energy" (I have never heard a definition of this I could understand). Also, from one song to another, you're only supposed to move either major to minor, minor to major, or change just one letter. That's known as the Camelot method and it may be a lot of hot air.
Another trick I heard was taught in schools was "don't tickle me with a feather and then hit me with a brick."
Wow! Music recommenders come up every once in awhile on HN and they're almost alway unusable for underground music. Aside from not recognizing some of the names I tried, this one does quite well. It just needs quick links to listen to examples and it'll already be more useful than Spotify, Apple Music or Last.fm.
Some test cases I used: Logic1000, Tessela, Piezo, TSVI, Two Shell, Caterina Barbieri, Schacke, Amnesia Scanner, Doss, Celyn June [fail], DJ Plead [fail], DVRTN [fail], JakoJako [poor results], LDS [needs disambiguation]
This may already be a better approach, but has anyone built a music recommender that's based on published data (stuff you can find on Discogs) like record labels, aliases, collaborators, remixed-by, etc? As a listener mainly to dance, experimental, and contemporary classical music, I suspect a system based on simple metadata associations would easily beat other approaches.
That's exactly what I've been working on. Choose it based on the people someone chose to work with and those networks.
It's very much just a single user system that I haven't generalized because that's a way harder problem, the rules aren't generalizable. Classical music for instance, may have 15 names on the credits, a jazz record may have like 4 labels it gets placed on.
Let's take an EP by a popular artist. It may have remixes by popular DJs. Those links are poor quality. Now if it's by an unknown artist, those links usually become high quality.
So do you follow the network of the guy playing the oboe? Maybe? Sometimes weird connections like the album artist is the strong link, sometimes it's compilation albums that one of the songs is placed on. Take say the 1992 release Trancemaster 1: https://www.discogs.com/release/54719-Various-Trancemaster-V... the clustering of those artists is a very strong high quality link. And then there's the "that's what I call music" type compilations where they're worthless.
I can do the music I like because I can narrow the ruleset but a general application is basically a winograd schema challenge because there's a large body of intuition required to weight the network. This task is certainly a nontrivial neural network problem.
Doing it manually with human discretion works. I've got tools for doing that and large labeled data sets I've been working on for 4 years. It just doesn't generalize.
That makes sense. I threw in my favorite (used to be underground) artist and saw every side-project they had pop up (and they have around a dozen). OTOH, I just threw in Nujabes and saw a number of people I do not recognize, which I find interesting.
From my tests, the suggestions make sense. It seems to be better than the spotify "I will play things that send you into an echo chamber" algorithm.
There's two approaches: musicologist and popularity.
The musicologists do a serious study and have sophisticated tools I can't pretend to understand but their results sound similar: so similar that the serendipity and adventure is sucked out of it.
Also they suffer from the generality problem as well. Qualities that matter in one genre, such as airy female vocals in a minor key, or whatever, are absolutely irrelevant for another genre. Take for instance, Irish folk music where it may be signal and say, Acapella, where it may be noise. Thus considering them is both right and wrong and we're back to our problem.
If you want to categorize the music to "fix" the problem, you run into binning issues. Let's say early 90s hardcore; they can have trance, breakbeat, dnb, house and jazz sections in a single song. Good luck trying to use your genre based contextual mapping on an unsupervised model.
The popularity approach, which ignores the content, is deceptive because it appears in many forms: people who listened to X also listened to Y or Y is trending or any of a number of variations where some magnitude of humans or temporal delta is used as signal.
These all tend towards the not long end of the long tail and so you eventually get the same mediocre experience - you start with your obscure prog rock group from the 1960s and 20 songs later you're at Cream or Hendrix. The "solution" is to tamper the drift via clustering but it will tend the same directions.
In practice though, these approaches service the majority of tastes, that is familiarity, so they're fit for purpose.
Nice! Glad you're doing this work. Sounds like something that would be best integrated into a community with strong contributor culture that could do the tagging, but that's another task entirely.
That would have to be a specialized genre tracker though, most of your suggestions would fall flat for metal, and even label would not always be helpful.
FWIW, the site does not work for my underground bands at all, out of 5 tries 3 didn’t show up at all (Remember Twilight, Der Rest, Jessica's Crime), and 1 (Scythia) has barely any related bands. Only "Die Streuner" got decent results. Some of the bands it’s missing even have wikipedia pages.
could this be as simple as...a python script which takes an artist - using python discogs_client and just scours every label mate of that artist and feeds itt to you - made more intelligent with acoustic fingerprinting...and a tiny bit of human intervention? :)
Throughout all these years, I got convinced that human recommendations (those from a close circle of friends I trust) are the best source of music discovery.
[Shameless plug] That's why I created https://digs.fm, which is kinda like Goodreads but for music.
All manually user-generated by people contributing <= 3 names through https://www.gnoosic.com/, then voting on other names suggested in the same neighborhood of the graph.
It would be cool data to get access to. But it's a nice singleton in the mean time. Like Netflix's challenge dataset, it would be interesting to have another layer of associations to compare with more formal genre lists. I wonder what kind of data Spotify releases.
Whenever I’ve had Peter Gabriel in a playlist I automatically get a lot of Genesis (Phil Collins era) mixed in. I’m aware PG was a founding member of Genesis, but when they split so did the music. They just aren’t similar in my mind.
On this map they are right next to each other. Maybe the problem is me?
Kate Bush <-> Maggie Reilly <-> Mike Oldfield would be good links, but she's not even on the page and Mike Oldfield doesn't show up on Reilly's page either.
I’ve been wanting something like this for a long time. My tastes are a little offbeat sometimes.
Wax Tailor, one of my favorite artists, uses sampled dialogue from films, educational programs, corporate videos and whatnot to create quite interesting sonic creations that I just love.
I wish more we’re creating similar types of music (I just found Biosphere’s “Petrified Forest” which is themed to that old movie - and again I love it). I’ve had very little luck except an occasional one-off from DJ Shadow, or Aim.
I’m not familiar with some of the artists that show up in relation to Wax Tailor on this map - but the ones I do know are not the same vibe.
.. these are the folk that later collaborated to form M|A|R|R|S of Pump Up the Volume fame
There's an interesting French bit of old film digital cutting together with some samples and sountrack from Polo & Pan: https://www.youtube.com/watch?v=hVW63Z_8deE that might strike a chord - they appear not so far from Wax Tailor when entered.
I have tried various approaches to search for something to Wax Tailor's "Que Sera" and have found only Barry Adamson's "Something Wicked This Way Comes" (you will recall from Lost Highway) and Venetian Snares' "Gloomy Sunday."
I tend to spend a few months at a time in each music rabbit-hole. The most I can remember from that general area are not all the same vibe, but close enough to keep me around: Pretty Lights, The Avalanches, Nightmares On Wax, RJD2, Little People, Gramatik, Quantic, Emancipator. Kid Koala also belongs on that list, particularly after Baby Driver.
Most of my exploration routes were just Pandora or Spotify recommendations. Digging through the graph of "Similar Artists" is my main path - and basically the same quest as OP.
What I would really like to do is zoom out on that graph, and interact with it a little less linearly. What if you could grab a handful of favorite nodes, stretch it out, and see what lies in between them? What if you could remove the ones you already know?
Listened to a few Wax Tailor songs, reminds me a lot of Pretty Lights--at least the vibe. Pretty Lights doesn't do as much of the voice samples. Check out Pretty Lights' album Taking Up Your Precious Time, you might dig it.
wax tailor really dosent have that many peers in the sense, he does what he does in full albums, whereas other artists may do his style in one or two tracks amongst their catalogue. those peers lean more towards the instrumentalization, and then towards jumbling rap lyric clips throughout. also dosent help hes european, which makes a difference believe it or not as far as peerage (i can hear the difference, dont know how to verbalize it, much in the same way aim is too)
broadly? start scouring with dustedwax.org its dusted wax kingdom label. then theres artists im sure you know cut chemist, j rocc, d-styles, rob swift, roc raida (all dudes w/ west coast style) - though not the narrative type youre looking for and more scratch focused. theres prince paul & his related projects...ruckus roboticus...adlib, and madlib...mr scruff... also it seems the late 90's were the 'golden age' of this type of sound production, for places to start looking.
more narrowly? shoulder shrug emoji. heres a couple
- mr scruff - fish (or the keep it unreal alb)
- stuntdouble & tenshun - drunktro
- professor brian oblivion - feel the funk
- tack-fu - how long
- main sequence - mannhandler
- cut chemist - motivational speaker
- 1200 hobos - the illuminati
- born talent - the avids affection
- pablo - due praises
- public enemy - dark side of the wall 2000
- zaire black - experiments with the truth
- roc raida - ill kick your ass/who you fuckin wit
- jurassic 5 - contact, react
- colossus - interlude 1
- insight - outro (crooked needle alb)
- wax tailor - que sera
- d-styles - wont you be my neighbor
- dj js1 - rule 4080
- centz - insult 2 watch
- detane - luv interlude
- dj revolution - head 2 head, rhythm control
- rob swift - the ghetto
- adlib - grovers eerie story, introduction to rhythm
- ruckus roboticus - here we go, never play with scratches
- time machine - the mekster
- pugslee atomz - michael's
- epstein - tape 1
- dj day - close your eyes, what planet what station
- malcolm and martin - do it again
- captain murphy - hovercrafts and cows
- louis mackey - mc-ide
- ribbonmouthrabbit - not such a clear day
- trillionaires - recession proof
- madlib - smoke theme for dankery harv
- koncept - understanding
....and one of the earliest tracks to use 'clips' style in the 80s: willesden dodgers - not this president
Yes, that's a good point. For instance 'We do what we can' by Sheryl Crow is one of those. It stands out so much for me that I find it hard to recognize it as her work compared to the other tracks.
Everything but the girl is one of those. Every time they'd ship a new album I'd run out and buy it only to think 'bleh'. But four weeks later after it had grown on me I'd usually consider it their best work. So much for consistency :)
I have found this to be terrifically difficult. I went to the trouble of going through Tom Waits' discography to find something similar to "Alice," without luck. Some of my best mixes take years to create.
The problem, to me, breaks down into three separate difficulties, perhaps three and a half:
First, if the Music Genome people are correct, the number of variables (if you will) is quite large, creating an enormous search space. I grant an extra half to the concept that portions of that search space would be sparsely populated. We are getting pretty good at searching these kinds of spaces (something I would like to know more about).
Second, each person has a different idea of what similarity entails. For some it is instrumentation. Others, production value. Some are looking for lyrical subject similarity. And so forth. And most people are very much "I'll know it when I hear it," unable to really qualify that similarity function. It requires quite a bit of introspection, meta-cognition, and a good vocabulary to really describe what that target is. Right now, we're bad at that and have a bad habit of just throwing everything in a bucket. As an example, I might like some music with a fast tempo for some mood, but I might also like "darkjazz" (a microgenre which more or less collapsed about seven or so years back). Most recommender systems would then pick out some fast, "hot" jazz, which I absolutely do not want, but it has the tempo of some things I do like and the genre of something else I like.
Third, the process of assigning values to said variables was quite labor intensive when they were doing it, as well as subjective. Highly trained musicians spending half an hour on each song (which means that, with that many variables, while the time seems long, the examination must be cursory). It would really be something to make progress in this area and I think it might be the most fruitful. However, I think any real progress in this area might mean you would need a corpus of at least a hundred thousand songs human-rated, each by multiple evaluators, then programs written for each of the, say 450 genes (if we are to use the Music Genome Project). Finally, the programs would be tweaked until they aligned with the ratings given by humans.
It's one of those things I would fund if I were an eccentric billionaire.
350 genes for rap music? LOL please. what would that even entail? music genome project is the equivalent to the AI images with eyes everywhere in the photo from years ago. ancient tech that just does not work.
How do y'all define 'similar music' anyway. I haven't found the perfect playlist creator or music recommender. Spotify's radio has a pretty low hit rate for me.
In some ways, I think there's a market for a multidimensional music similarity search tool or classifier (where playlists are basically classifications).
A few examples. I have a 'rock concert' playlist that maps on style, artist, era/decade, but not tempo (since rock concerts sometimes break up intense songs with less intense ones) and a 'slow lounge' playlist that maps onto instruments, average tempo, tempo variance within song, etc.
What I really want is a way to assign a feel or purpose across a few axes which are not just the typical 'genre'. Something actually objectively measurable like tempo, volume, etc.
Edit: As a side note, I found that country (like Alan Jackson) seems to work better in the car than some other song types. I think it might actually be a frequency thing where country is Darwinism optimized to be audible and enjoyable in old trucks?
This website has fallen in the exact same "artist names are unique" trap that Last.fm did, making every single name that's used by two or more performers completely useless.
There's also no name normalisation. I got a result where I saw at least five duplicated artist names because people can't agree on how to spell them.
Does it say anywhere what this is based on? Is it based on some index of 'genres' associated with artists or something? Some spotify/apple index or something they absorbed?
If you listen to any kind of underground dance/dj-related music, I never saw the need for any of these recommendation engines/databases. Radio, recorded sets, dj mixes are always the unbeatable curation. It's kind of the whole point.
Cool concept and the UI is well executed. The results themselves were somewhat hit and miss for me - Purity Ring was a nice discovery related to Grimes, but the Lorde results were all somewhat "Eh". I wonder how an Album based version would fare - i.e. focus on Solar Power, not Melodrama.
I'm a huge fan of Purity Ring, and most of the artists I looked at seemed to a solid gathering of familiar neighbors. Unfortunately, that means this graph is pretty similar to the Spotify one I usually wander around. I'll definitely have to spend some more serious time with it, though.
I would recommend you listen to iamiwhoami, Crywolf, AWAY, Chrome Sparks, Of the Trees, Sam Gellaitry, Burial, James Blake (especially his early stuff), and SOPHIE; for a good variety of directions.
First artist I tried failed: Øystein Sevåg, also not in transliteration.
David Benoit fared a lot better, Neneh Cherry matched Carmel(?) but not Youssou N'Dour with whom she had her most popular song. Interesting concept though I will definitely be playing around with it.
Edit: long ago I did something like this but based on the individual band members so that you could trace them as they moved through their career, that might be another interesting angle to pursue as a weight on the links. But it's a boatload of work to track down all of the individuals that contributed to a track. Another angle is to go by producers and possibly even recording technician. Good luck untangling Alan Parsons and Trevor Horn from everything they touched, the page will be too crowded to read.
That's an absolutely amazing song. It is just incredible how much energy he pours into that. And Ramses Shaffy should be another easy link from Jacques Brel.
Angelique Kidjo ... her rendering of 'Summertime' is forever in my memory, live during the African Music Festival in Delft. She and Salif Keita stood out from the 10's of acts that were there.
You can't really describe the kind of power those two radiated. Off the scale in every sense.
I have now tried four different artists and all show very bad matches. Vaguely artists in the same genre, but still clustered by overly broad features.
Artist and music similarity is an incredibly nuanced area, I don't think any of the modern attempts worked sufficiently well (apart from the Music Genome Project, that's still in a different class to this day).
They other site in the network, Gnoosic, gave me some absolutely incredible recommendations. https://www.gnoosic.com Lots of new albums to listen to over the weekend :)
It is pretty bad for classical composers - groups together completely unrelated, very different people: I typed Monteverdi and the closest neighbors shown are Scriabin, Britten, Poulenc, Faure, Schubert. It does not make any sense.
I believe it's pretty bad for just about anything - at least, after a few tests, I haven't found a good example where it, indeed, provides meaningful and useful results. I wouldn't have an issue if it would say "check out what people also tend to listen to", but it's most certainly not "find similar music".
I strongly suspect that it has no idea about music whatsoever, and probably falls into same association fallacy as most recommendation systems, basing relationship on co-occurrence rather than any inherent similarity.
I tried Gerald Finzi. It shows Vaughn Williams with and without Ralph, mixes composers with performers, and adds rather unrelated composers. It also adds "Clocks and Clouds". This is either an instrumental trio that creates "epic" music, or a composition by Ligeti, neither of which is in the same category nor has any relation musically to Finzi. And it adds "Yak Bondy", which is inexplicable.
I guess it just does things when there's not much data, but that makes it practically useless for discovery.
This project was one of the major inspirations for my learning to code a decade ago, and I still turn to it sometimes when I'm looking to make sense of what kind of music a band plays
Cool. Created the same idea quite a few years ago. Also added the ability to traverse the network and view clips and info about the artists. The entire thing leverages one SQLite file.
Pretty much every music recommendation service tells me that Turin Brakes are almost identical to Gomez and I, having ears, can tell they are not.
I'd love to know how they have come to this conclusions. Is it because there aren't many Gomez lovers in the dataset and the few that are also happen to like Turin Brakes thus skewing the recommendation?
Tried 3, one gives me unrelated results, one good results, and one doesn't exist. Still very useful, but I hope the text won't float around which makes it hard to click, vertical list would be great.
I had good results with selecting all and then pasting into a text file. Not sure if the things at the top of the list ended up being "closer" or not, but at least I had a simple list of names to check out.
I think I broke it. I found Peewee's Big Adventure, clicked through a bunch of movies to TV shows, found Starcraft and finally SUS, Amongus, Barack Obama and Walter White.
That's just not how I work. I like _good_ stuff, I don't confine myself to genres, I'll go anywhere if it's half decent. I want a service that says hey, you listened to Benny Goodman's Sing Sing Sing, you should check out The Thirteenth Floor Elevators' first album.
The only recommendation that kinda works for me is based on an aggregation of what other people listened a lot to. Sure on small data sets that is probably going to be genre biased, but on large sets it can start working and break out of that.