You should try Spotify's "Discover Weekly" personalised autogenerated playlist, it's been built by the Echo Nest team and learns directly from what you listen to.

https://press.spotify.com/uk/2015/07/20/introducing-discover...

I found them to be pretty bad. Let's say I listen to genre X exclusively for two weeks, and then for genres Y and Z during week 3. The genres are completely different (lets say X - rap, Y - kpop, Z - just pop music). On week 4, I'd expect to be recommended a mix of X, Y and Z, seems reasonable. However, it's only X because I listened to it a lot a while ago.

So it's pretty good if you are listening to the same music genre, however, it sucks if you like to change the music based on your mood, etc. I also noticed that it doesn't detect the language the music is in so I often get songs with lyrics in languages that I have no clue about.

When was the last time you used it? I've noticed that the discover weekly playlist has consistently featured four or five majorly different genres, all drawn from the music I've tended to listen to lately. It's been good about doing this since I first subscribed back in May.

Some of the songs have absolutely missed the mark, but there have been others I've found that make me go back and listen to the entire album they're from. Definitely have encountered different languages as well.

A week or two ago :) it was pretty in summer/earlier autumn when I was listening to one genre almost exclusively but I've been mixing up lately and my recent preferences got totally ignored.

I listen to multiple, vastly different genres. I find Spotify's Discover playlist may favor one particular genre one week, but it's usually pretty good.

And the behavior you described doesn't sound "pretty bad". If you haven't listened to a particular genre in a few weeks, I think it's reasonable you don't get suggestions for it for the upcoming week of listening. But if you do start listening to that genre again, your next playlist will have it.

The recommendations are usually on point, but to my knowledge, there's no way to say "I don't like this artist, stop recommending it". I have quite a few of those, that I should like based on my tastes but for some reason I don't.

I could try that. But I think I eventually want something more flexible. For example, when listening to a song, I want the system to automatically pick songs that are similar to that song. Or I'd like to pick songs based on instruments used, tempo, origin, language, genre, etc.

Spotify does that already too - right click a song and "Start Song Radio". This will automatically pick songs that are similar.

I know it seems like this is what you want, but for me I've never been satisfied with this simplistic approach to discovery. Taste is very complex, and for me saying to someone "I really love xylophone" and them recommending me a lounge jazz tune with xylophone will always fall flat. What most people don't realise is why they love certain instruments or tempos or genres etc. It's usually down to how they were introduced to these, which defines their archetypal music using these instruments for example. It's all contextual.

My context for liking xylophone goes something like this "I love xylophone, because in the late 90s I listened to a lot of Beastie Boys 'Hello Nasty' which has some really nice pieces of music with some great xylophone in them. Then hearing some Mulatu Astatke recently got me interested in xyolophone and ethio jazz in general, combined with a really cute reaction my girlfriend had to one of his songs."

The best recommendation engine I've seen is Spotify Weekly. It's really very good, but make sure you get in there and save a lot of artists and albums you like first

You can do this if you right click and do play song/artist radio.. but in my experience it sucks really bad, and has only gotten worse over time. Use to, it was kinda useful. Now it's complete garbage. I'm not sure what they changed in their algorithm but typically when I do this I get 5 artists that are only marginally related, and then it proceeds to play music from those 5 artists in a loop and the "dislike this" feature seems to do nothing. I've actually disliked a song and then it came on about 20 minutes later.

I have a playlist of a lot of songs I like (except I try to keep it to one or two songs per band).

The radio from that converges to endless Foo Fighters and The Prodigy -- over half the songs!

It's not that I dislike them, but I don't particularly love them either and there is just no diversity on that station.

I understand, suggestions still have issues. I listen to a lot of more obscurely specific genres, and sometimes will get just a couple of repeated songs if I make a radio based on a lesser-known artist.

That said, if you haven't given the Spotify Weekly Recommendations a shot -- you should. Maybe it's because it has more data to work with than song/artist/playlist-specific radios, but for me that feature has been extremely spot-on. In fact, I hadn't used spotify for a while because I kept hearing the same songs, but now I religiously keep on top of my discoveries each week.

I agree on the weekly discoveries, they are great. The radios, at least the playlist radios, are crap.

It seems like the radios point me towards the most common denominator and plays the most mainstream stuff that matches, while the weekly discovery lists come up with some very quirky stuff that I love.

I guess the radios can't have any "memory", it seems to generate close to one song at a time (hence a lot of repetitions), whereas the discovery list can be created 30 songs at a time and seems to go out of its way to find songs by bands that I've never listened to.

It does that as well, and calls it "radios" for a given artist or song.

I've given up on discovering new music using technology and gone back to curated lists. No matter how hard I've tried to exploit technology, friends/bloggers who are obsessed with music are always the best sources.

Subjective tastes tend to diverge too rapidly with machine learning.

If (fingers crossed someone is doing this) there was a really easy and non-spammy way for people to build a playlist using _any_ source material into something like muxtape (or opentape) I would use it. Until then, I will just continue to ask them for youtube links and bandcamp profiles.

Also, songs in an individual artists library can vary WILDLY.

There are some artists from whom I truly only like one single song from their entire catalogue.

Others, I like every single song, but that is super rare.

And I dont have the time interest or energy to actually attempt to tag or classify music and songs that I do like -- so I just pretty much am stuck listening to the same thing or actually getting individual recommendations from real people.

You might want to look into http://www.last.fm

I've found those recommendations much better than Pandora or Google Music, and you can hook into it with many different services/players.

Last.FM used to have an amazingly accurate and helpful music discovery system, and great similar artist recommendations up until roughly 2 years ago where they either limited the amount of similar artists, or put some sort of blanket filter or whitelist for a smaller pool of "approved similar artists".

Last.FM was so ahead of the game it's mind boggling how they just seem to have fizzled out while Grooveshark, and Spotify came in.

Last.FM had a huge user base with great info on individual music tastes, had a functioning streaming service, a whole functioning ecosystem of music fans and musicians that just seemed to have layed dead in the water after it was bought out.

> Last.FM was so ahead of the game it's mind boggling how they just seem to have fizzled out while Grooveshark, and Spotify came in.

I used to think that same thing, but I've come to realize that most people don't care that much about music. It's just pleasant background noise to them.

you can also check this little "similar artists" spotify tracklist generator I've built, based on last.fm api: http://verdisco.yebsoft.com/

Shameless plug, but we are doing exactly that here at http://milq.com. We are attempting to organize the world's media into collaborative playlists, which you can then subscribe to for a feed of content uniquely tailored to your personal taste. Subscribing or dismissing a playlist helps us recommend better content to you. Give it a look! :)

Why do you have the youtube player defaulting to 240p?

It should set it to highest available, or let it be an option.

I like it otherwise!

It defaulted to auto for me.

I've run a music blog for the past 8 years and I've used a ton of different discovery methods. In the earlier years blogs provided me with most of my content. As of the past year Soundcloud has really shown me a whole new world of music. When you play a single track it goes straight to a recommended track, generally right in the same vein as the previous track. I've found hundreds of great tracks I would have otherwise never known existed.

On an unrelated but interesting note: I'm still trying to figure out their recommended algorithm, it seems to be a mix of tags, and cross-referencing your account with who's liked/played a song. What I don't understand is if I listen to a folk song, it will generally play another completely unrelated folk song but with no folk identifiers (tags). I've concluded it's magic.

I've occasionally discovered new music from the SoundCloud recommended tracks, but at this point I follow so many users that I barely have time to keep up with what they are posting/reposting. As with virtually all social networks, the users are the content creators. Not only do they actually create the content (they might not), the bulk of what you get from them is their likes and repost.

If I catch up on my SoundCloud stream (posts/reposts) I'll switch over to somebody's list of likes. Some of my favorite producers are buried in their passion for music 24/7 and have upwards of 1000 likes to listen through.

Recommended tracks are nice, and it's something to fill the void if you are low on content from people you follow, but in my opinion it's not that much better than going out and seeking new music and artists in other ways. I'm sure as SoundCloud improves the algorithm though it could automate some of the behaviors I'm doing manually (e.g. playing content from and following users whose content is liked/reposted by the users I follow). It might be doing this already, if maybe only indirectly, I guess I just like having control. One of the most annoying things I'm afraid it could do is promote sponsored music. I love listening to music that people are passionate about, not music that is funded by large amounts of money.

Similar experience here.

Re: recommendations from blogs, did Hype Machine do anything for you?

That's what last.fm does. You don't need a paid account for that, just add a last.fm scrobbling plugin to your music stack.

I used to use Pandora before they cut off the UK due to licencing. I kept discovering great new songs using Pandora that I just haven't been able to recreate in any other service (I use Spotify regularly).

I've no idea if that would still be the case as we're looking at about 5+ years ago now when I used it.

I had the same experience and only started using Spotify when Pandora left Europe.

Discover Weekly is a very good replacement, though, even if it's not the same.

Have you tried http://www.gnoosic.com ? It is run by a HN user who does a whole lot of discovery research over at http://www.gnod.com

I am not sure whether this is a problem that can be solved entirely with software. Oftentimes it is not so obvious what's the relationship between what you have and what you'd like to discover.

In fact, one of the biggest issues I have with current methods is that they recommend me more of the same. When I look for new music I want - to some extend - something that is unlike what I've heard before.

Say you listen to Blue Lines by Massive Attack. Chances are, the recommendation engine will recommend you Portishead or Morcheeba. (User likes genre trip hop? let's play more trip hop). But maybe I'd like to explore different forms of UK rap music now...

> Say you listen to Blue Lines by Massive Attack. Chances are, the recommendation engine will recommend you Portishead or Morcheeba.

Last.fm [1] recommends more Massive Attack, or "Aftermath" by Tricky, followed by Portishead. Which is fine, as you said "chances are" that's what's wanted. If you wanted UK rap, you at least used to be able to search for the tags "uk" and "rap", but I can't see that on the new website...

This service[2] uses the Last.fm API to make a Spotify playlist based on a track.

[1] http://www.last.fm/music/Massive+Attack/_/Blue+Lines

[2] http://www.spotalike.com/#playlist/97600

> If you wanted UK rap, you at least used to be able to search for the tags "uk" and "rap", but I can't see that on the new website...

In case of my example such a search would work. However, I guess there are lots of cases where a user doesn't even know what s/he wants. That is: "I want more of the same but it should also be very different"

Now that's quite a task for an algorithm! Also, making suggestions based on empirical user data is rather hard as people are highly emotional, especially when listening to music. Thus, people (in general) are not very predictable when it comes to music.

Individuals might have some core principles when choosing music so that perhaps individual algorithms for individual users could work but coming up with an acceptable one size fits all solution is rather impossible, imo.

Some people are happy when the algorithm suggests another top 40 song they haven't heard yet. Others will loose it when they get more metal songs but none of them is teutonic thrash metal [1]

Now we could empower people by making apps like spotify scriptable so that everyone could refine their recommendation algorithm. But realistically that is not what your average user wants...

Perhaps one could improve the situation by asking the user a few questions when s/he first starts the application. That way the app can choose the best fitting recommendation algorithm from a range of algorithms in the background.

[1] Apparently, this is a thing: https://en.wikipedia.org/wiki/Heavy_metal_subgenres#Teutonic...

Totally agree. But, as usual with "nice things", this one doesn't align with Spotify's best interests. If it did, this would have been solved a long time ago - it's not like they haven't acquired the tech/skills to do so.

Those nice little lists Spotify helpfully shows are almost guaranteed paid-for inclusion (apart from the chart-style lists, to a degree, but that's another ball game). Labels/Record Companies/Promoters etc are almost certainly paying for inclusion there as it drives huge awareness and subsequent merchandise, ticket etc sales. Much like radio airplay, where the station controls the programming, this is where Spotify gets access to the same revenue stream.

So giving me perfect recommendations filtered through an exclusion list for which, let's be honest, the maths and data science for have long been sorted out via standard ad-tech and e-commerce recommendation engines, is simply not going to aid in the value proposition for a nicely curated list they can charge to be in.

I doubt that Spotify is charging money for inclusion in the Discover Weekly playlists. This would run counter to the goals of promoting Discover Weekly usage, which is to get more users thinking "What's new on Spotify this week?" rather than "Let me pull up the new Taylor Swift album." The more that users flock to Discover Weekly, the more leverage that Spotify gains against the labels, so my guess is that they're striving for a wide diversity of rightsholders in Discover Weekly so that any one label threatening to leave will not damage the playlist too badly.

This is like asking a chef to make new food based on what you like to eat.

You will never discover anything really new, your experience will be narrowcasted. Avoid this and let dj's, friends, musicians, people, and just surfing around the internet help you find new music.

Whenever I hear someone excited about automated music discovery a little voice goes off in my head going ..."ohh no".

There's very little difference between finding good stuff in automated discovery and finding it because a DJ chose it.

Robots have a hard time recommending unpopular items because they lack the context to differentiate between and good and bad. They'd have to be able to listen to music and make cross-referencing judgments about quality. Are the musicians playing well? Are the lyrics any good? Is the style original? Is the recording good? Is it cheesy? Does it try to imitate another song? Is it a good cover? Does it work in spite of a lot of negative indicators? Is it a nice complement to the songs before and after? etc.

But if you want to find hits in a genre + time period that you don't know very well, recommendation engines are pretty good, and you don't need to have DJs for that.

That's ridiculous and will be ridiculous until robots can replace humans.

For example, do you think someone like DJ shadow who has spent tens of thousands of hours digging in empty, old, dusty basements finding and curating music over a lifetime has the same insight as a dicovery algorithm?

It's not even close to a comparison, the algorithms are lame compared to actual experience and expertise.

I spent 10+ years in the music business as a "buyer" and was later part of a successful music start-up, I can tell you with a resounding "fuck that" to auto discovery unless you like water downed with a side of boring.

No, I don't think DJ Shadow can be automated, but live DJing is a music discovery channel for only a scant minority of fans. These fans are sometimes called "trainspotters."

Radio DJs, on the other hand, have any number of forces affecting their song choices (to the degree the role exists at all anymore), and all in all they're just as much a black box as an algorithm. Sure, we all had radio shows we liked for some time, but those DJs changed and we found something else (notwithstanding people who have been listening to "Renee and Crazy Pete in the morning" for 20 years).

Even with all of this, I suspect there's a discrete economic or mathematical reason that profit-oriented discovery engines regress to a mean and will always be trying to recommend the latest AdeleBieberNational. They may not, right away, but a few clicks deep in the sidebar and you find LL Cool J or Foo Fighters popping up.

I do most of my music discovery on Soundcloud these days. That plus Bandcamp covers most of my needs.

Many artists are starting collectives which publish albums with a variety of artists approximately monthly, which also is great for discovery.

Of course this only works if your preferred genres are there - if you like future funk, nightcore, and hip hop you're golden.

Check out The Echo Nest API for some inspiration - I'm pretty sure Spotify and some similar services use it for their recommendations. http://developer.echonest.com/docs/v4

Two things the 1990s got right and (in my opinion) have never really been replaced: Pointcast and Moodlogic.

Moodlogic worked on my library, I could choose any song, then it would craft a variable length playlist based on that song's mood and instruments. There was a crowd-sourced set of metadata, and I was able to fill in the gaps or correct anything I didn't agree with.

If anyone knows of anything along these lines, I'd be grateful to hear about it.

https://en.wikipedia.org/wiki/MoodLogic

Personally, I find music discovery a very interesting problem. So many variables to play with! If user listened to artist X, should we recommend more of it or would it be too much? How important are recently played songs vs long term trends in listening (e.g. listening to genre X most of the time over the last year, however, this week I'm listening to Y only)?

Unfortunately, it requires to be running at spotify-like scale to do something useful, have a large enough library and gather feedback quickly enough :( oh well. I will continue dreaming about it though.

Well, maybe the Spotify API could allow access to what you want. Though you don't get access to play count / history data, you could indirectly get it by scrobbling to last.fm.

I've found youtube's related videos list to be quite good at this (and I bet they must be already using some AI for recommendations?)... So you could look for the song title/id and get what youtube thinks are the related songs/artists:

https://developers.google.com/youtube/v3/docs/search/list#re...

I'm looking for a program that can look at my music library and group it up in a coherent way. It would make finding something to listen to so much easier, as genre is often woefully inadequate to reflect listening mood. I know there are some things like that (and iTunes has a feature) but they all miss the mark.

For new stuff, I'm fine with a combination of best of 20## lists, following some interesting soundcloud accounts, and idle searching the internet / various subreddits.

Have you tried last.fm?

For the last month or so, every Monday I open up Spotify's Discover Weekly playlist. I generally listen to it in the background all week long, giving the songs time to grow on me. I usually find 3-4 songs that I want to add to my "real" playlists, which I consider to be a pretty good ratio.

Last.fm worked well with me in terms of music discovery. They kept shooting themselves in the foot, removing features where I found value and terminated my subscription (living outside US/UK).

Not a recommendation engine, but for music discovery, I loved Twitter Music (ex: WeAreHunter) and it's down now.

As a replacement, I subscribed to popular blogs (popjustice, stereogum...) playlists on Spotify.

I've found Apple Music to be better at music discovery & recommendation but Spotify to still have a superior product (iTunes is a mess now)

I'd like to second this. After trying a lot of competing tools in the music streaming space, I've come back to Apple Music's playlists and suggestions more than any other. They're very good at recommending me things that I turn out to enjoy.

The UX is utter shit, but the suggestions are great.