I rarely use the built in features like Stations or Guest DJ because it’s so incredibly far off that I’d get a better mix just by putting my entire library on shuffle.
I get that it’s “an algorithm” but there shouldn’t be any instance where you’re listening to '90s alternative rock and you get Sonny Rollins, Amy Grant, and James Taylor in the mix. Even Bell & Sebastian is questionable. It’s quite a jarring experience.
Would anyone be able to share some insight as to what’s going on with Sonic Analysis (beyond the published article) and if there’s anything we can do to improve upon the results it generates?
I’m on Version 1.32.5.7328 and have Sonic Analysis enabled on a library of over 70,000 songs.
Unfortunately, this was something I started to take note of a couple weeks ago and failed to note which of the DJs I was using. My reasonably confident guess would be DJ Gemini.
I’m including screenshots of the existing playlist (is there some other way to export a playlist?).
Sorry, I cannot deduce anything from these screenshots.
Please take attention next time you are activating one of the DJs.
The different DJs do different things.
Sometimes relatively suprising things.
Because they don’t know anything about the cultural aspects of the artists or the surrounding sociopolitical events during which an album or track was created.
All they do know is the release year and some parameters about the music itself.
Which may be surprising to the end user, if it turns out that some culturally different tracks are actually quite similar musically.
P.S: not all of the Radios use sonic data. Some are just based on release year (of the album) or “Similar Artists” or “Genre” or “Mood” etc.
I’ve been listening to this playlist for the past 45 minutes with DJ Gemini. There were a handful of tracks that weren’t ideal (The Cramps) but not entirely out of place.
I’ve now switched to DJ Stretch - “relatively surprising” seems accurate. Massive Attack > Roy Orbison with the Real Philharmonic > Electric Light Orchestra > R.E.M. > Pearl Jam.
Well Massive Attack have at least one track with some strings in there, so pairing that with other string-based tracks is not too far-fetched.
Of course it also depends on the content of your library. Maybe there weren’t any tracks available with more similar properties.
Read the descriptions, Gemini doubles up a playlist so 50% of the content is yours and 50% (every other track) is the twin. There’s no indicator of what’s going on with your playlist, nor what the source was that you activated Gemini on.
DJ Gemini: Inserts the most sonically similar track after each track, when possible.
I presume that means it’s adding a track after the current track that sounds similar to the current track. Is there some other way to interpret this?
When I just turn DJ Gemini on, it inserted Nirvana - Very Ape after Weezer - Buddy Holly. These songs are in no way similar, sonically.
I have over 70,000 songs in my library. I assure you there is no shortage of songs that sound more similar to Buddy Holly than Very Ape. Though an actual Buddy Holly song may sound more sonically similar that Very Ape, that song also shouldn’t be in contention. And following Massive Attack with Roy Orbison just shouldn’t ever happen regardless of what any algorithm was designed to do.
So, my question here is what is the algorithm that’s making these decisions and, being that it’s so poor, how can I actively improve upon it?
Use the “Show Similar Tracks” to show what tracks it considers sonically similar.
I see the same rec over here. I wouldn’t say it’s dissimilar, you might be expecting too much. Behind the scenes, a neural network is assessing each track in snippets, and categorizing in terms of 50 different dimensions (e.g. aggressive, happy, etc…)
It’s a neural network, and you can’t, beyond adding more music so there are potentially better matches.
We don’t get a lot of negative feedback around Sonic, so it seems to work well for most people. However, there are certainly genres and types of music with which it works better.
