Similarity (Essentia) DSTM mixer - call for help

[cpd73]

I’ve created a “Don’t Stop The Music” mixer that attempts to mix tracks based upon the similarity of certain characteristics. Tracks are analysed using the Essentia music analyser to extract certain characteristics of tracks (danceability, aggressiveness, happy, sad, etc). This data, along with some tag data, is stored in an SQLite database. The DSTM mixer calls a python app which queries the SQLite database for other tracks with characteristics similar to seed tracks. From these tracks a random selection of 5 is chosen and added to the queue.

Currently my ‘matching’ is very trivial. The Essentia analyser computes the probability that a track is (e.g.) danceable – in the range 0..1 (with 0 being not danceable). To determine if two tracks are similar I calculate the difference of each attribute, and then take the average difference. Tracks are sorted with the most similar (least difference) first. I’ve tried to give certain characteristics more ‘weight’, but this does not seem to make a great deal of difference.

If you are interested in this, you will need to checkout 3 git repositories:

1. https://github.com/CDrummond/essentia-analyzer Contains the python code to analyse tracks. This calls the ‘essentia_streaming_extractor_music’ binary to do the actual analysis, and stores the results in an SQLite database. The git repo has a pre-built 64-bit Linux version taken from Roland0’s LMS Essentia integration page. If you want to run analysis on a Windows, or macOS, machine you will need to re-build this with support for ‘Gaia’ (Perhaps Windows users can use WSL (Windows Subsystem for Linux) to use the prebuilt binary?) Please note I have not built Essentia myself, so will be of limited help in fixing any build issues. I do plan to look into this, but improving the matching code is of more importance (to me) for now.
2. https://github.com/CDrummond/essentia-api Contains a simple HTTP server that the DSTM mixer calls to get mixable tracks. This should be installed onto the same machine as LMS.
3. https://github.com/CDrummond/lms-musicsimilarity Contains the ‘Similarity’ DSTM mixer. This is a regular LMS plugin.

Please note that the analysis is quite slow. On a (6 year old) i7 laptop with SSD I averaged about 15 tracks per minute. Still, for me, this is about 3-4 times faster than MusicIP.

I’d be interested in help from others to tweak the similarity matching side. It seems to produce reasonable results for me, but I guess results will vary. This area is very new to me, so my matching code is very basic and I’m sure this could be vastly improved - which is why I'm asking for help.

Once you have analysed some tracks, and run essentia-api, you can use the following to ‘dump’ a list of similar tracks:

curl 'http://localhost:11002/api/dump?track=AC-DC%20-%20Who%20Made%20Who/01%20Who%20Made%20Who.mp3&format=text&count=50&weighting=1'

This will return the 50 most ‘similar’ tracks (along with similarity (lower value = higher similarity) and genre info). weighting=0 will return the average similarity with no weighting applied to any attributes, weighting=1 adds some weighting to try to make certain attributes more important than others. This call will iterate through all entries in the database. If you pass 'close=1' it will only look at tracks where each attribute is +/-0.65 of the seed track - this should speed up queries with large datasets. When looking for similar tracks, tracks by the same artist are excluded - I might, however, remove this. This 'dump' API is only intended to help with development, and as a quick way to query the DB. Track's passed to this API must be URL-encoded, and be the relative paths that are stored in the DB.

The DSTM mixer, however, uses a slightly different API. It calls the 'api/similar' API (passing a JSON object with the seed tracks, previous tracks, etc). This API will also look for 'close' tracks first, and only iterate all if there are not enough 'close' tracks returned.

[mherger]

> - https://github.com/CDrummond/essentia-api Contains a simple HTTP
> server that the DSTM mixer calls to get mixable tracks. This should be
> installed onto the same machine as LMS.

Are there technical limitations which require this to be an external
server (put aside the advantage of having it in a different process)?
Eg. is there a Python library, but none for Perl?

At first I thought as this was a SQLite database, you should be able to
query it from the LMS plugin.

--

Michael

[Roland0]

[*] https://github.com/CDrummond/essentia-analyser Contains the python code to analyse tracks.

should be https://github.com/CDrummond/essentia-analyzer

This calls the ‘essentia_streaming_extractor_music’ binary to do the actual analysis, and stores the results in an SQLite database. The git repo has a pre-built 64-bit Linux version taken from Roland’s LMS Essentia integration page. If you want to run analysis on a Windows, or macOS, machine you will need to re-build this with support for ‘Gaia’ (Perhaps Windows users can use WSL (Windows Subsystem for Linux) to use the prebuilt binary?) Please note I have not built Essentia myself

build instructions are here (Gaia/SVM)

[cpd73]

> - https://github.com/CDrummond/essentia-api Contains a simple HTTP
> server that the DSTM mixer calls to get mixable tracks. This should be
> installed onto the same machine as LMS.

Are there technical limitations which require this to be an external
server (put aside the advantage of having it in a different process)?
Eg. is there a Python library, but none for Perl?

At first I thought as this was a SQLite database, you should be able to
query it from the LMS plugin.

Not really - but I find python much easier to code in than Perl. The main reason its in Python is that its based upon a previous mixer using the Musly library, and that was in python - as I borrowed some code from Roland0 to integrate the Musly C library with Python. Also, if the similarity code needs to iterate all tracks it can be slow, so running in a different process stops LMS from stalling. I find Perl to be very odd in some scenarios - passing arrays, $, @, @$, \@$ ??, no parameter names in function definitions, etc.

The code could be re-written in Perl, so removing one app. But, then main reason for this posting was to ask for help with improving the similarity code - or even to get validation that it is OK.

[cpd73]

should be https://github.com/CDrummond/essentia-analyzer

D'oh! I did a search and replace for consistency, and I guess that changed the URL too! Will edit my post - thanks.

[Roland0]

I’d be interested in help from others to tweak the similarity matching side. It seems to produce reasonable results for me, but I guess results will vary.

Finally did some testing (had to write an import script for my existing analysis data first).
I submitted 33 tracks (with default params) and evaluated the suggestions by checking each of the suggested 10 tracks for similarity / fit.

Here's the result:

bad:      ###########
middling: ################
good:     ######

same tracks for musly:

bad:      #####
middling: #########
good:     ##################

Next, I'll run the same tracks with different parameters (weighting, no same artist exclusion)

[cpd73]

Finally did some testing (had to write an import script for my existing analysis data first).
I submitted 33 tracks (with default params) and evaluated the suggestions by checking each of the suggested 10 tracks for similarity / fit.

Here's the result:

bad:      ###########
middling: ################
good:     ######

same tracks for musly:

bad:      #####
middling: #########
good:     ##################

Next, I'll run the same tracks with different parameters (weighting, no same artist exclusion)

Thanks for the feedback! I assume this means 'Essentia' had more bad matches than good? How do you evaluate good/bad? My 'similarity' checks are very basic, and need a LOT of improvement. Perhaps I should be comparing more parameters?

[Roland0]

Thanks for the feedback! I assume this means 'Essentia' had more bad matches than good?

Yes, by far. Musly performed much better.

How do you evaluate good/bad?

Basically, I assign each result track one of:
+ (good fit)
~ (acceptable, at least fits "in spirit" (related mood, similar energy))
- (noticably dissimilar mood and/or energy)

A good result should have significantly more + than ~, and a max of 1 minus
Everything with >=3 minuses is bad

Examples:
Essentia:

Pixies - Brixton Academy 05.06.2004/1 - 03 - Monkey Gone To Heaven.mp3
 + NoMeansNo - Why Do They Call Me Mr. Happy/07 - I Need You.mp3
 + A Place to Bury Strangers - Exploding Head/05 - Keep Slipping Away.mp3
 + Front 242 - Parted Tragedy/01 - No Shuffle.mp3
 - Editors - The Weight/02 - The Weight.mp3
 ~ VA - The Crow/01 - The Cure - Burn.mp3
 - Kristin Hersh - Speedbath/02 - Torque.flac
 ~ Doves - Kingdom Of Rust/03 - The Outsiders.mp3
 + Trans Champs - Double Exposure/01 - Give It to You.mp3
 - Emiliana Torrini - Rarities/04 - To Be Free.mp3
 + Sonic Youth - Sonic Nurse/04 - Mariah Carey + Arthur Doyle.mp3

Musly:

Pixies - Brixton Academy 05.06.2004/1 - 03 - Monkey Gone To Heaven.mp3
 + Pixies - Brixton Academy 02.06.2004/1 - 08 - Monkey Gone To Heaven.mp3
 + Pixies - Coachella 2004/11 - Monkey Gone To Heaven.mp3
 + Pixies - Brixton Academy 04.06.2004/1 - 13 - No 13 Baby.mp3
 + Pixies - Brixton Academy 02.06.2004/1 - 11 - No. 13 Baby.mp3
 + Pixies - Eugene 28-04-2004/13 - Monkey Gone To Heaven.mp3
 + Pixies - Brixton Academy 03.06.2004/2 - 10 - No 13 Baby - Caribou.mp3
 + Pixies - Brixton Academy 05.06.2004/2 - 02 - Where Is My Mind.mp3
 + Pixies - Brixton Academy 05.06.2004/1 - 14 - Gouge Away.mp3
 + Pixies - Regina, SK, Canada/1 - 11 - Where Is My Mind.mp3
 + Pixies - Brixton Academy 05.06.2004/1 - 11 - I Bleed.mp3

Essentia:

Ignis Fatuus - The Futility Goddess/01 - Communion.flac
 + Dirty Three - Horse Stories/03 - Hope.mp3
 + Alaska In Winter - Dance Party In The Balkans/09 - Staring At The Sun.flac
 ~ The Black Heart Procession - Six/01 - When You Finish Me.flac
 + Clair Obscur - Antigone/20 - Irlande.flac
 + Jessica Bailiff - At The Down-turned Jagged Rim Of The Sky/05 - Violets & Roses.flac
 - Cat Power - Moon Pix/05 - Metal Heart.flac
 + Nils Frahm - Wintermusik/01 - Ambre.flac
 ~ The Magnetic Fields - The Charm Of The Highway Strip/06 - Crowd Of Drifters.mp3
 + Philip Glass - Glassworks/01 - Philip Glass , Opening.flac
 - The Breeders - Pod/12 - Metal Man.mp3

Musly:

Ignis Fatuus - The Futility Goddess/01 - Communion.flac
 + The Legendary Pink Dots - Any Day Now/01 - Casting the Runes.flac
 + Rachel's - The Sea And The Bells/07 - Cypress Branches.flac
 + The Moon Lay Hidden Beneath A Cloud - A New Soldier Follows The Path Of A New King/14 - A New Soldier Follows The Path Of A New King Track 14.flac
 + The Moon Lay Hidden Beneath A Cloud - The Smell Of Blood But Victory/13 - The Smell Of Blood But Victory Track 13.flac
 + Nature And Organisation - Beauty Reaps The Blood Of Solitude/11 - Bonewhiteglory.flac
 + Rachel's - Selenography/01 - A French Galleasse.flac
 ~ Kristin Hersh - Sunny Border Blue/07 - Trouble.mp3
 + Current 93 - All The Pretty Little Horses/06 - The Blood Bells Chime.flac
 ~ Current 93 - Black Ships Ate the Sky/02 - Sunset (The Death of Thumbelina).mp3
 ~ Nigel Kennedy And Kroke - East Meets East/04 - Eden.mp3

Essentia:

Philip Glass - Koyaanisqatsi/03 - Cloudscape.flac
 + Julia Kent - Delay/16 - Tempelhof.mp3
 - Der Blutharsch - The Pleasures Received In Pain/06 - Track VI.mp3
 ~ GYBE - Yanqui U.X.O./02 - Rockets Fall on Rocket Falls.mp3
 + Max Richter - The Blue Notebooks/02 - On The Nature Of Daylight.flac
 - VA - Pi Soundtrack/01 - Clint Mansell - Pi R Squared.mp3
 + Set Fire To Flames - Sings Reign Rebuilder/03 - Steal Compass-Drive North-Disappear.mp3
 + Clint Mansell - Requiem For A Dream/01 - Summer Summer Overture.mp3
 + A Challenge Of Honour - Spartan Victories/04 - Bow Belle.mp3
 - Portishead - Third/01 - Silence.flac
 + The Dead Texan - The Dead Texan/07 - When I See Scissors, I Can't Help But Think Of You.mp3

Musly:

Philip Glass - Koyaanisqatsi/03 - Cloudscape.flac
 + Max Richter - The Blue Notebooks/02 - On The Nature Of Daylight.flac
 + The Dead Texan - The Dead Texan/08 - Girth Rides A (Horse) +.mp3
 + The Dead Texan - The Dead Texan/02 - Glen's Goo.mp3
 + OST - Europa/04  - Dinner At Hartmanns.mp3
 + SPK - Zamia Lehmanni (Songs Of Byzantine Flowers)/05 - In Flagrante Delicto.flac
 + Rachel's - The Sea And The Bells/04 - Lloyd's Register.flac
 + Philip Glass - Koyaanisqatsi OST/03 - Clouds.flac
 + Alex Somers - Dawson City - Frozen Time/21 - Track 21.flac
 + Amber Asylum - Frozen in Amber/04 - Heckle and Jeckle.flac
 + Rachels - Handwriting/05 - Seratonin.mp3

Essentia:

Current 93 - Judas As Black Moth/26 - The Death Of The Corn.mp3
 ~ Julee Cruise - Twin Peaks CDS/01 - Falling.mp3
 + Isobel Campbell - Milkwhite Sheets/02 - Willow's Song.mp3
 + VA - Looking For Europe/1 - 05 - Willows Song.mp3
 - Throwing Muses - Purgatory Paradise/19 - Terra Nova.flac
 + Helium Vola - Liod/11 - Frauenklage.mp3
 + Revolutionary Army Of The Infant Jesus - Mirror/07 - Psalm.mp3
 - Cat Power - What Would The Community Think/04 - Nude As The News.flac
 - Shannon Wright - Maps Of Tacit/06 - Dirty Facade.mp3
 ~ Kristin Hersh - Strings/04 - Me and My Charms (String Version).mp3
 ~ OST - Europa/14 - The Aria.mp3

Musly:

Current 93 - Judas As Black Moth/26 - The Death Of The Corn.mp3
 + Current 93 - Horsey/02 - The Death of the Corn.flac
 + Current 93 & Hilmar Oern Hilmarsson - Island/08 - Oh Merry-Go-Round.flac
 + Current 93 & Hilmar Oern Hilmarsson - Island/04 - Fields Of Rape (Sightless Return).flac
 + Current 93 & Hilmar Oern Hilmarsson - Island/02 - The Dream Of A Shadow Of Smoke.flac
 + Current 93 & Hilmar Oern Hilmarsson - Island/12 - Fields Of Rape And Smoke.flac
 + Backworld - Isles Of The Blest/11 - Beautiful Dream.flac
 + The Legendary Pink Dots - Any Day Now/01 - Casting the Runes.flac
 - Marc Almond - Mother Fist And Her Five Daughters/11 - Ruby Red.mp3
 ~ Broken Bells - Broken Bells/07 - Citizen.mp3
 ~ An Pierle & White Velvet - st/01 - Jupiter.mp3

Essentia:

Cyclobe - Luminous Darkness/06 - You're Not Alone You're Dreaming.mp3
 + Chapelier Fou - 613/09 - Grahamophone.mp3
 + Labradford - Mi Media Naranja/01 - S.flac
 ~ Nobody - Transparent Seasons/03 - Listen.flac
 + Mogwai - Peel Sessions/03 - Killing All The Flies .mp3
 + Shinjuku Thief - The Scribbler/01 - The Assessor.flac
 + Blind Cave Salamander - Troglobite/07 - Used To Be Last.mp3
 ~ Dirty Three - Whatever You Love You Are/01 - Some summers they drop like flys.mp3
 ~ Clock DVA - Thirst/08 - Moments.flac
 + John Murphy - Sunshine/19 - The Surface Of The Sun.mp3
 - Rome - To Die Among Strangers CDM/02 - Reversion (Edit).mp3

Musly:

Cyclobe - Luminous Darkness/06 - You're Not Alone You're Dreaming.mp3
 + VA - World Serpent MM Compilation/05 - Cyclobe - You're Not Alone, You're Dreaming.mp3
 + Michael Nyman - Prosperos Books/05 - Miranda.mp3
 + Michael Nyman - The Composer's Cut Series, Vol. 1- The Draughtsman's Contract/02 - The Garden is Becoming a Robe Room.mp3
 + Michael Nyman - The Composer's Cut Series, Vol. 1- The Draughtsman's Contract/06 - Queen of the Night.mp3
 + Michael Nyman - The Composer's Cut Series, Vol. 2- Nyman - Greenaway Revisited/02 - Queen Of The Night.mp3
 + Michael Nyman - The Composer's Cut Series, Vol. 2- Nyman - Greenaway Revisited/06 - Vermeer's Wife.mp3
 + Michael Nyman - The Composer's Cut Series, Vol. 1- The Draughtsman's Contract/04 - A Watery Death.mp3
 + Carter Burwell - Millers Crossing/15 - End Titles.mp3
 + Michael Nyman - The Composer's Cut Series, Vol. 1- The Draughtsman's Contract/05 - The Disposition of the Linen.mp3
 + Nature And Organisation - A Dozen Summers Against The World/03 - Blood of Solitude II.mp3

My 'similarity' checks are very basic, and need a LOT of improvement. Perhaps I should be comparing more parameters?

tbh, I only had a (cursory) look at the analysis code in order to write an importer for existing data. I'll look at the similarity part next.
As I said, I'm still unsure if this approach has a chance of surpassing musly, but it's certainly an interesting experiment.

[cpd73]

tbh, I only had a (cursory) look at the analysis code in order to write an importer for existing data. I'll look at the similarity part next.
As I said, I'm still unsure if this approach has a chance of surpassing musly, but it's certainly an interesting experiment.

Yeah, I'm 100% certain just taking the average of the differences between params is too simplistic. Which is why I tried adding weighting, to give certain params more significance than others - but I totally failed at this so far! I also think I should use some of the other essentia models as well. The ones I chose seemed to be the easiest.

Perhaps looking at two tracks that Musly thinks are very similar, and then looking at the difference in their essentia values will help.

I have been using this mixer with ~20k tracks for the last few weeks, and the mixes have been reasonable for me.

[Roland0]

Perhaps looking at two tracks that Musly thinks are very similar, and then looking at the difference in their essentia values will help.

Hmmm:

* seed: /data/music/Pixies - Brixton Academy 05.06.2004/1 - 03 - Monkey Gone To Heaven.mp3
+------------------------------------------------------------------------------------+-------------+--------------+--------------+------------+---------+---------+-----------+-------+--------+---------+---------+-------+
| file                                                                               |   danceable |   aggressive |   electronic |   acoustic |   happy |   party |   relaxed |   sad |   dark |   tonal |   voice |   bpm |
|------------------------------------------------------------------------------------+-------------+--------------+--------------+------------+---------+---------+-----------+-------+--------+---------+---------+-------|
| /data/music/Pixies - Brixton Academy 05.06.2004/1 - 03 - Monkey Gone To Heaven.mp3 |       0.146 |        0.574 |        0.661 |      0.037 |   0.348 |   0.281 |     0.634 | 0.487 |  0.750 |   0.151 |   0.836 |   119 |
| /data/music/Pixies - Brixton Academy 02.06.2004/1 - 08 - Monkey Gone To Heaven.mp3 |       0.044 |        0.000 |        0.595 |      0.024 |   0.570 |   0.000 |     0.809 | 0.480 |  0.915 |   0.544 |   0.809 |   121 |
| /data/music/Pixies - Coachella 2004/11 - Monkey Gone To Heaven.mp3                 |       0.569 |        0.326 |        0.536 |      0.186 |   0.247 |   0.110 |     0.590 | 0.692 |  0.775 |   0.809 |   0.999 |   124 |
| /data/music/Pixies - Brixton Academy 04.06.2004/1 - 13 - No 13 Baby.mp3            |       0.445 |        0.028 |        0.721 |      0.059 |   0.450 |   0.136 |     0.818 | 0.245 |  0.660 |   0.892 |   0.578 |   123 |
| /data/music/Pixies - Brixton Academy 02.06.2004/1 - 11 - No. 13 Baby.mp3           |       0.000 |        0.000 |        0.833 |      0.017 |   0.322 |   0.000 |     0.809 | 0.399 |  0.728 |   0.944 |   0.713 |   127 |
| /data/music/Pixies - Eugene 28-04-2004/13 - Monkey Gone To Heaven.mp3              |       0.102 |        0.605 |        0.086 |      0.005 |   0.227 |   0.427 |     0.316 | 0.165 |  0.225 |   0.163 |   0.954 |   119 |
| /data/music/Pixies - Brixton Academy 03.06.2004/2 - 10 - No 13 Baby - Caribou.mp3  |       0.316 |        0.009 |        0.836 |      0.082 |   0.065 |   0.046 |     0.887 | 0.405 |  0.959 |   0.319 |   0.068 |   126 |
| /data/music/Pixies - Brixton Academy 05.06.2004/2 - 02 - Where Is My Mind.mp3      |       0.353 |        0.081 |        0.276 |      0.312 |   0.900 |   0.209 |     0.337 | 0.410 |  0.111 |   0.969 |   0.815 |   162 |
| /data/music/Pixies - Brixton Academy 05.06.2004/1 - 14 - Gouge Away.mp3            |       0.616 |        0.429 |        0.244 |      0.060 |   0.658 |   0.473 |     0.035 | 0.643 |  0.500 |   0.881 |   0.508 |   124 |
| /data/music/Pixies - Regina, SK, Canada/1 - 11 - Where Is My Mind.mp3              |       0.457 |        0.217 |        0.159 |      0.025 |   0.601 |   0.184 |     0.196 | 0.305 |  0.023 |   0.995 |   0.514 |   155 |
| /data/music/Pixies - Brixton Academy 05.06.2004/1 - 11 - I Bleed.mp3               |       0.430 |        0.824 |        0.422 |      0.045 |   0.794 |   0.370 |     0.390 | 0.167 |  0.943 |   0.091 |   0.987 |    92 |
+------------------------------------------------------------------------------------+-------------+--------------+--------------+------------+---------+---------+-----------+-------+--------+---------+---------+-------+
*** stats
+--------+-------------+--------------+--------------+------------+---------+---------+-----------+-------+--------+---------+---------+---------+
| func   |   danceable |   aggressive |   electronic |   acoustic |   happy |   party |   relaxed |   sad |   dark |   tonal |   voice |     bpm |
|--------+-------------+--------------+--------------+------------+---------+---------+-----------+-------+--------+---------+---------+---------|
| min    |       0.000 |        0.000 |        0.086 |      0.005 |   0.065 |   0.000 |     0.035 | 0.165 |  0.023 |   0.091 |   0.068 |  92.000 |
| max    |       0.616 |        0.824 |        0.836 |      0.312 |   0.900 |   0.473 |     0.887 | 0.692 |  0.959 |   0.995 |   0.999 | 162.000 |
| mean   |       0.333 |        0.252 |        0.471 |      0.082 |   0.484 |   0.195 |     0.519 | 0.391 |  0.584 |   0.661 |   0.695 | 127.300 |
| median |       0.392 |        0.149 |        0.479 |      0.052 |   0.510 |   0.160 |     0.490 | 0.402 |  0.694 |   0.845 |   0.761 | 124.000 |
+--------+-------------+--------------+--------------+------------+---------+---------+-----------+-------+--------+---------+---------+---------+
*** delta (track - seed track)
+------------------------------------------------------------------------------------+-------------+--------------+--------------+------------+---------+---------+-----------+--------+--------+---------+---------+-------+
| file                                                                               |   danceable |   aggressive |   electronic |   acoustic |   happy |   party |   relaxed |    sad |   dark |   tonal |   voice |   bpm |
|------------------------------------------------------------------------------------+-------------+--------------+--------------+------------+---------+---------+-----------+--------+--------+---------+---------+-------|
| /data/music/Pixies - Brixton Academy 02.06.2004/1 - 08 - Monkey Gone To Heaven.mp3 |      -0.102 |       -0.574 |       -0.066 |     -0.012 |   0.222 |  -0.281 |     0.174 | -0.007 |  0.165 |   0.394 |  -0.026 |     2 |
| /data/music/Pixies - Coachella 2004/11 - Monkey Gone To Heaven.mp3                 |       0.424 |       -0.247 |       -0.125 |      0.149 |  -0.101 |  -0.171 |    -0.045 |  0.206 |  0.025 |   0.659 |   0.164 |     5 |
| /data/music/Pixies - Brixton Academy 04.06.2004/1 - 13 - No 13 Baby.mp3            |       0.299 |       -0.545 |        0.060 |      0.022 |   0.102 |  -0.145 |     0.184 | -0.242 | -0.090 |   0.741 |  -0.257 |     4 |
| /data/music/Pixies - Brixton Academy 02.06.2004/1 - 11 - No. 13 Baby.mp3           |      -0.146 |       -0.574 |        0.172 |     -0.020 |  -0.026 |  -0.281 |     0.174 | -0.087 | -0.023 |   0.793 |  -0.122 |     8 |
| /data/music/Pixies - Eugene 28-04-2004/13 - Monkey Gone To Heaven.mp3              |      -0.044 |        0.031 |       -0.575 |     -0.031 |  -0.121 |   0.146 |    -0.319 | -0.322 | -0.525 |   0.012 |   0.119 |     0 |
| /data/music/Pixies - Brixton Academy 03.06.2004/2 - 10 - No 13 Baby - Caribou.mp3  |       0.171 |       -0.565 |        0.175 |      0.046 |  -0.283 |  -0.235 |     0.252 | -0.082 |  0.209 |   0.169 |  -0.768 |     7 |
| /data/music/Pixies - Brixton Academy 05.06.2004/2 - 02 - Where Is My Mind.mp3      |       0.207 |       -0.493 |       -0.385 |      0.275 |   0.552 |  -0.072 |    -0.297 | -0.077 | -0.639 |   0.819 |  -0.020 |    43 |
| /data/music/Pixies - Brixton Academy 05.06.2004/1 - 14 - Gouge Away.mp3            |       0.470 |       -0.145 |       -0.417 |      0.024 |   0.310 |   0.192 |    -0.600 |  0.156 | -0.250 |   0.730 |  -0.327 |     5 |
| /data/music/Pixies - Regina, SK, Canada/1 - 11 - Where Is My Mind.mp3              |       0.311 |       -0.357 |       -0.502 |     -0.012 |   0.252 |  -0.097 |    -0.438 | -0.182 | -0.727 |   0.844 |  -0.322 |    36 |
| /data/music/Pixies - Brixton Academy 05.06.2004/1 - 11 - I Bleed.mp3               |       0.285 |        0.251 |       -0.239 |      0.008 |   0.446 |   0.089 |    -0.245 | -0.320 |  0.193 |  -0.060 |   0.152 |   -27 |
+------------------------------------------------------------------------------------+-------------+--------------+--------------+------------+---------+---------+-----------+--------+--------+---------+---------+-------+
*** stats (delta)
+--------+-------------+--------------+--------------+------------+---------+---------+-----------+--------+--------+---------+---------+---------+
| func   |   danceable |   aggressive |   electronic |   acoustic |   happy |   party |   relaxed |    sad |   dark |   tonal |   voice |     bpm |
|--------+-------------+--------------+--------------+------------+---------+---------+-----------+--------+--------+---------+---------+---------|
| min    |      -0.146 |       -0.574 |       -0.575 |     -0.031 |  -0.283 |  -0.281 |    -0.600 | -0.322 | -0.727 |  -0.060 |  -0.768 | -27.000 |
| max    |       0.470 |        0.251 |        0.175 |      0.275 |   0.552 |   0.192 |     0.252 |  0.206 |  0.209 |   0.844 |   0.164 |  43.000 |
| mean   |       0.188 |       -0.322 |       -0.190 |      0.045 |   0.135 |  -0.086 |    -0.116 | -0.096 | -0.166 |   0.510 |  -0.141 |   8.300 |
| median |       0.246 |       -0.425 |       -0.182 |      0.015 |   0.162 |  -0.121 |    -0.145 | -0.085 | -0.056 |   0.694 |  -0.074 |   5.000 |
+--------+-------------+--------------+--------------+------------+---------+---------+-----------+--------+--------+---------+---------+---------+