Hacker News Reader

The algorithmic recommendation is broken, because it makes wrong assumptions. Every recommendation engine I've encountered appear to be based on the idea that I care about artists or genres, rather than the sound profile of individual tracks.

I have no idea, but my take is that streaming services are trying to find other artists within the same genre, or using the listing patterns of other users to match you up with new music.

This makes a weird assumption that I actually like everything a band makes, which is rare, or that because I like an artists or a few songs within a given genre, then I must like all music with in that genre. Mostly I listen to music from a wide number of artists, across generes. There a musicians where I like most of their work and some where I just like the sound of one particular track.

What I want, is button, when I press that, "The Almighty Algorithm" will analyse the sound, the beats per second, the vocals, the instruments, the lyrics, anything that affect the sound and locate other songs with similar profiles. Bonus points for letting me input stuff like: higher tempo, less bagpipes, the singer has a high pitch voice which hurt my ears.

The social stuff is easy and the streaming platforms can quickly implement something similar and put you out of business. Actually analyzing the sound profiles and using that to help you create playlists and discover new artists, that not something I've seen done and I bet it's because it's will require actual work.

Wasn't this the whole point of Pandora?

I think so. Pandora actually hired people to manually annotate their corpus with various sorts of traits. See https://www.pandora.com/corporate/mgp.shtml

Yeah, and as a result it's the one music service I pay for.

It doesn't give me any control of which attributes I care about, which frustrates me to no end, but it does fairly well at turning up things I haven't heard before and like if I can give it a decent set of seed tracks.

Agreed. Pandora is the only music service I've stayed with because of its discovery system. It's not perfect, but I've found countless artists and songs over the years at Pandora that I might not have known without.

I also occasionally go back to Slacker/LiveXLive for it's fine tune controls, DJs and news. But it's discovery system isn't as good and it's had plenty of bugs (like playing wrong songs/titles). So I stick with Pandora for that radio experience tailored to me.

"The algorithmic recommendation is broken, because it makes wrong assumptions. Every recommendation engine I've encountered appear to be based on the idea that I care about artists or genres, rather than the sound profile of individual tracks."

Agreed 100% on this. I find I like a certain BPM and "type" of sound to songs I often have on repeat.

Also agreed. I've had days where I was in a really productive mood and wanted some really high tempo music to go with it, and end up searching for 30 minutes and only finding things that slow me down. And other moods and modes have their own vibe, that has nothing to do with genre and a lot to do with tempo and timbre.

It's why DJ's are paid well for what they do. I wish there was an actually good A.I. DJ that can blend songs together well.

> Every recommendation engine I've encountered appear to be based on the idea that I care about artists or genres, rather than the sound profile of individual tracks.

And sometimes I don't even know why I like a particular song. This is especially true when it comes to the concept of catchiness. There is certainly some music theory around hooks, chord loops, etc. but I honestly haven't found any of it anywhere near as explanatory or predictive as more well-established classical western music theory concepts like harmonic function or voice leading. So often it really just feels like "that song is super catchy just because it is."

I believe we are very far from the technology necessary to really grok a person’s musical tastes. What currently probably works the best is the "people who liked x also liked y" scheme (generalized to vectors in song space or something).

However, that method has the property that the algorithm can’t judge any new music, it always requires existing ratings from members of the platform before it can recommend a song to other people.

Then there’s the additional issue that a person’s musical taste changes over time.

Do you want something that'll let you define a timespan (e.g. 2:45..3:12) and find tracks like that? I imagine that would be pretty fun from a discovery sense -- for EDM especially.

Every recommendation engine I've encountered appear to be based on the idea that I care about artists or genres, rather than the sound profile of individual tracks.

Spotify has the tech to do this, they've demonstrated it with a genre explorer tool, but they clearly don't use it in their radio stations and recommendations.

As several other comments have mentioned, music companies inevitably end up being redesigned to match the desires of rightsholders, and not users. The only way around this would probably be some kind of legally mandated compulsory licensing, so that any streaming service or end user can play any content if they pay the predetermined price, removing contract negotiations from the picture.

It feels like there’s a bunch of ML technology that exists that could plausibly be pieces of this solution: there are algorithms for computing embeddings, various audio processing models, recommender system networks, etc. I wonder what’s been tried in this space and how/why it’s been lacking…

> What I want, is button, when I press that, "The Almighty Algorithm" will analyse the sound, the beats per second, the vocals, the instruments, the lyrics, anything that affect the sound and locate other songs with similar profiles. Bonus points for letting me input stuff like: higher tempo, less bagpipes, the singer has a high pitch voice which hurt my ears.

This might be possible in theory but part of the issue is that the algorithms input parameters are not the same from person to person. The parameters you described are mostly technical attributes of the sound, but what about cultural/political affiliation? Do you care if the lyrics are in a different language (some people do, some people don't). Do you care if there is a political or religious message (and are you particular about which religion or political tribe)? Can Spotify take that into account somehow? Maybe, but it seems very difficult to tailor that to the user.

Speaking personally, music is more than just an aural experience for me. At times it's about connecting with my social group, or cultural traditions, or a spiritual experience. At other times it's just the music. When I listen to mainstream stuff, Spotify is ok. When I get deeper into edge-case territory, Spotify doesn't really know what I'm looking for and often makes suggestions that appear to demonstrate a shallow understanding the cultural meaning of the music I'm listening to.

So many of the things in this comment section are things that were solved by Zune very well (e.g. the social features were great, if you knew anyone else who used Zune - https://en.m.wikipedia.org/wiki/Zune_Social)

Zune had the Channels concept - essentially what Spotify-owned playlists are nowadays - but they partnered with radio stations and Billboard (among others) to provide content.

https://m.youtube.com/watch?v=UmUU3R-Y21I

(Notice how 10 years later that UI is still largely fresh and modern; I miss Zune)

Was an original Zune adopter from the earliest days (30GB Brown Fat Zune). I ran into two people on my college campus with Zune's and each time we shared a song to each other. After a month I never used the song sharing feature ever again and never really had anyone bring it up.

I really loved my Zune, the desktop software was some of the worst ever created. It helped pushed me towards eventually just getting an iPod to flash with Rockbox so I could just use common lightweight sane software on my desktop to manage the library on my devices.

Things are so much better now that I can just load up 200GB of music on my android phone, manage it via a normal file explorer on desktop, and use the VLC app to listen and create/manage playlists.

Please help if you can - I seem to remember reading that the preferred verb for the song sharing on zunes, per MS marketing, was "squirt," as in, "to squirt a song."

I doubt myself in this, of course, because it's patently idiotic. Is this anything you have the slightest recollection of?

https://www.newsweek.com/zune-should-go-beyond-squirting-107...

I really didn't remember it being called that, but reading this makes me vaguely remember my assistant manager making fun of the feature name because she also had a Zune.

Thank you. I didn't mean to make you play LMGTFY, but I appreciate it.

I remember that too, FWIW.

Great points. I'd even go a step further and suggest that "discovery" is not really a problem that can or should be solved in the manner that tech companies would like. I ran a radio show in college on my college's station, and from my experience most of that crowd is uninterested in further tech penetration of music.

I'm a bit late to the discussion but I like the way Bandcamp does discovery. They have a carousel that you can very easily listen to 20-30 songs in quick succession, and they have these curated blog posts where you can read and listen to the music on a topic.

I really get the feeling that the folks and Bandcamp are actually into music.

Everything else I've tried just seems to make it very hard to break out of the bubble they decided to put me in.

I’d add Radio Paradise to the list of good discovery tools.

I see a lot of replies about discovery not being a solved problem here, which it isn't, but that also isn't what the OP said.

> Discovery is not a viable product

I'm parsing "viable" as "commercially viable". You may solve the problem (for some definition of "solve" and "problem"), but can you make money doing so?

I don't know, I have no experience in this field, but OP seems convinced that the answer is "no".

Radio stations solved the problem and they make money doing it, so I say yes discovery is a viable product - but not in a way tech companies have been able to automate.

I have no idea how much radio stations make, but perhaps the implication was viable [tech] product where returns likely have to be much higher to get an ROI for a venture-backed startup.

> They do have a dedicated “discover” tab on the app, but you can’t stream the audio from it to a network speaker so it’s basically useless for me.

What discover tab are you referencing, I've never found a place in the app where you cannot directly play from and all their apps support cross device management.

Huh... so for the past couple weeks there was a top level "discover" tab on the iPhone app (so Home/Discover/Search/Your Library), and when you clicked it it would just a random song in a weird full-app video playback UI, with the only exposed controls being "go to this album" and "skip", no output device selector.

I went back to check today to confirm that even if I had an existing network device selected it'd disconnect that and only play locally, but clicking on the tab just showed an error page. I force quit the app and relaunched, and now the tab is gone entirely.

Guessing I was in some A/B experiment of a pre-release feature, it'll probably have more functionality in the full release (assuming they didn't can it)

> In my opinion, the best source for music discovery continues to be local radio

In my region(s) of France, they either suck or only play very big names. I envy people who live in range of a good local radio.

You might like this website then: http://radio.garden/

I'm curious why you pay for two music subscriptions. What's the process for deciding which one to listen to?

Got my family on a family account for Apple Music a while back, then at some point picked up a personal Spotify account to test the waters for their generated playlists, support for networked devices (chromecast/smart tv's), and API's (jkbx.fm). I can't migrate my family away from Apple Music and I don't want to give up Spotify's feature set, so I keep both.

Discovery is not a viable product, but Payola is - and Spotify is a repeat offender

do you just Shazam the songs you like from the local college radio?

That or I check the online playlist.