Re: [AUDITORY] Feedback on features for music similarity

[Paul Arzelier <paul.arzelier@xxxxxxx>]

Hi!

Thanks for your answer, and appreciate the kind words! You definitely 
make some good points:

1. I did NOT think about windowing the songs in "meaningful" moments, 
but the beginning and the end make a lot of sense since we wanna make it 
smooth. I'll explore and see what that yield.
3. I'm also thinking of letting users choose themselves, and see what 
they like more.
5. It's super cool, I didn't know they exposed that! I'll definitely use 
it to check that it matches at least e.g. the key the song is in.

@Pierre - do you have a hunch of what they might say? One problem is 
that musicians are not really concerned with the harsh reality of 
digital signal processing :D

Best Regards,
Paul

Le 30/07/2024 à 17:54, Hem, Charles a écrit :
> Hi Paul,
>
> Very cool project!  I'm a PhD student at Harvard studying cochlear 
> implants on the human perception side of things, so I'm no audio 
> specialist myself.  Trying to study music perception for cochlear 
> implanted recipients brings me back to featurization of music 
> frequently, though, so I have a couple non-expert thoughts for bullets 
> 1, 3, and 5:
>
> 1:  Maybe it's adding too much complexity, but I wonder if 
> splitting your feature analysis into more windows could be helpful if 
> you're most concerned about seamless transitions between songs.  For 
> example, "start of song" (first 20 seconds, or something near that), 
> "whole song", and "end of song" (last 20 seconds) could be used to 
> give some extra weight to making sure that the "end of song" from the 
> first song matches the "start of song" of the following song. Mainly 
> for songs that are highly variable or have a drastically different intro.
>
> 3:  I haven't read the chroma paper you cited, but gut reaction feels 
> like these features shouldn't outweigh tempo and loudness so much.  I 
> think your four class division could be a good idea.
>
> 5:  Spotify song features are pretty easy to get using this: 
> https://developer.spotify.com/documentation/web-api/reference/get-audio-features. 
> The features are much more abstract than your features, but in my 
> experience, Spotify's smart playlists do a pretty good job.  I'm 
> guessing these features are partially used in that, so could be a nice 
> reference to compare your results to.
>
> Most of this probably isn't super helpful, but I'll keep an eye out 
> for any updates about the project!  It could be a cool research tool 
> to quantify people's music preferences one day, which would be an 
> interesting use for me.
>
> Best,
> Charles Hem