[Announce] Music Similarity DSTM mixer

[mruddo]

Thinking about this some more. Would it make sense to allow creation of 'Smart Mixes' that will select a number of tracks, allow this to be added to the queue, and then DSTM takes over adding new tracks. DSTM, however, would not know about the SmartMix, just that these would be the initial songs to create mixes from. If so creating another API that returns X songs based upon some attributes should be doable. It then requires the plugin to implement a JSONRPC and the creation of a UI to create these mixes. Obviously for me the UI would be created in Material. I know SlimBrowse would allow some sort of UI creation for this, but that's beyond my knowledge level at the mo - and as I only use Material it's not of major importance to me.

e.g. I would add a 'Smart Mixes' (or similar) entry to 'My Music' to list current mixes, allow editing, etc. The edit dialog would allow you to specify which attributes to use, number of tracks, and (perhaps) order of tracks (closest to attributes, random, etc.)

This sounds good to me. You could then use these Smart Mixes with/without DSTM taking over on completion.

I've been listening to MS DSTM mixes for a couple of weeks now, and I'm still enjoying the varied journey through some long since heard tracks in my library. No issues to report either.

[prabbit]

I've been listening to MS DSTM mixes for a couple of weeks now, and I'm still enjoying the varied journey through some long since heard tracks in my library. No issues to report either.

Out of curiosity, what are values for your settings?

- - - -

While it's perhaps to be expected, I've found that using the same seed song over and over plays the same group of songs (albeit in different orders), including songs I have rated low and would prefer not to hear too often, than occurs with Spicefly Sugarcube/MusicIP. This is true even though the playlist runs for hours. Sugarcube has a setting to exclude tracks below a certain rating, which "protects" me from hearing these lowly-rated tracks. My experience with Sugarcube using the same seed song repeatedly is different, in that it produces a wider range of playlists. I'm still playing with Similarity's settings, which is the genesis of my question.

Curiously, Heart thus far seems to be the (Six Degrees of) Kevin Bacon in my collection. I used Stevie Wonder's Superstition as a seed track and soon enough got to Heart. Used Iron Maiden's Can I Play with Madness? and shortly thereafter heard Heart. And I played some other seed song from the 2000s and landed on Heart within a dozen tracks. This is anecdotally interesting because I don't have that many Heart songs in my collection.

[mruddo]

Out of curiosity, what are values for your settings?

I did play with them initially, then think I put them back to the defaults... I have:

Don't repeat artist for x tracks: 15
Don't repeat album for x tracks: 25
Don't repeat track for x tracks: 100
Min/Max Duration: 0
Number of mixer tracks: 5
Restrict Genres: No
Genre no match adjustment: 15
Genre gorup match adjustment: 7
Max BPM difference: 50
Max loudness difference: 5
Filter on key: Yes
Filter on attributes: Yes
Similariy Weighting: 35

While it's perhaps to be expected, I've found that using the same seed song over and over plays the same group of songs (albeit in different orders), including songs I have rated low and would prefer not to hear too often, than occurs with Spicefly Sugarcube/MusicIP. This is true even though the playlist runs for hours. Sugarcube has a setting to exclude tracks below a certain rating, which "protects" me from hearing these lowly-rated tracks. My experience with Sugarcube using the same seed song repeatedly is different, in that it produces a wider range of playlists. I'm still playing with Similarity's settings, which is the genesis of my question.

Curiously, Heart thus far seems to be the (Six Degrees of) Kevin Bacon in my collection. I used Stevie Wonder's Superstition as a seed track and soon enough got to Heart. Used Iron Maiden's Can I Play with Madness? and shortly thereafter heard Heart. And I played some other seed song from the 2000s and landed on Heart within a dozen tracks. This is anecdotally interesting because I don't have that many Heart songs in my collection.

I've never used ratings on tracks - although I can see why many do, and it's something I've often considered.

I'm probably using it differently though, as I tend to think of a seed track, see where that takes me... then often listen to an album I've not heard recently that comes up... then seed from there etc. I do get what you're saying about frequently re-occuring items though. For me it seems to be Sting (I have one compilation, that's all), but I'm often seeing that appear in the playlist.

However, when I compare the mixes to LastMix, where whatever I'd started with, that always seemed to tend towards a ninties-indie mix (not least because I have a reasonable number of such albums in my collection)... With Music Similarlty, the mixes tend to head off in different directions which I'm far more appreciative of.

I only have the one Stevie Wonder album (and no Superstition), and not a single Heart track... so I can't test to see if there's any in-built bias at play!

[cpd73]

While it's perhaps to be expected, I've found that using the same seed song over and over plays the same group of songs (albeit in different orders), including songs I have rated low and would prefer not to hear too often, than occurs with Spicefly Sugarcube/MusicIP. This is true even though the playlist runs for hours. Sugarcube has a setting to exclude tracks below a certain rating, which "protects" me from hearing these lowly-rated tracks. My experience with Sugarcube using the same seed song repeatedly is different, in that it produces a wider range of playlists. I'm still playing with Similarity's settings, which is the genesis of my question.

Yeah, I agree the randomness of selected tracks could do with improving - especially if there is only one seed track. I must admit I tend to have DSTM kick in after playing an album.

Curiously, Heart thus far seems to be the (Six Degrees of) Kevin Bacon in my collection. I used Stevie Wonder's Superstition as a seed track and soon enough got to Heart. Used Iron Maiden's Can I Play with Madness? and shortly thereafter heard Heart. And I played some other seed song from the 2000s and landed on Heart within a dozen tracks. This is anecdotally interesting because I don't have that many Heart songs in my collection.

Odd, I have Heart songs - but when I choose 'Can I Play With Madness', I don't see Heart being selected. However, different library, etc. Have you tested your analysis results, as stated on the music-similarity github homepage? Musly does have a known issue where it assigns the same similarity to all songs. If everything is OK with that, perhaps set all the filters to 0, so that just Musly is used.

Seeing as you have analysed your tracks with Essentia high-level, you could try setting 'Similarity weighting' to 100. This will cause only Essentia to be used. The similarity score is the calculated on how close each track is to the seed track's Essentia high-level attributes. Would be interesting to see if you see an improvement there. Or even try at 50, so that both are used, etc?

[prabbit]

Have you tested your analysis results, as stated on the music-similarity github homepage? Musly does have a known issue where it assigns the same similarity to all songs. If everything is OK with that, perhaps set all the filters to 0, so that just Musly is used.

Database looks clean to me. In fact, if you'd like a second, similarly-sized database to your own collection to use when testing, I can share mine with you. PM if interested.

Seeing as you have analysed your tracks with Essentia high-level, you could try setting 'Similarity weighting' to 100. This will cause only Essentia to be used. The similarity score is the calculated on how close each track is to the seed track's Essentia high-level attributes. Would be interesting to see if you see an improvement there. Or even try at 50, so that both are used, etc?

Actually, I have been using the Similarity weight at 100, just to focus on the high-level attributes. I haven't played much with the Musly side of things. Still on my To Do list. I don't feel I have enough experience to write with any confidence about my experience, even the Heart example could have been circumstantial. In Essentia scoring, I would put my confidence around 0.08425123, give or take. (Aside for any German-speaking readers, has anyone else noticed the loose connection between Musly and Müsli, the breakfast cereal made of oats, grains, fruits, and nuts? I couldn't find any link from the Musly creator, D. Schnitzer, describing such a connection, but it does seem inevitable, or a happy coincidence at the very least.)

[cpd73]

Database looks clean to me. In fact, if you'd like a second, similarly-sized database to your own collection to use when testing, I can share mine with you. PM if interested.

You can't tell if Musly is OK by looking at the DB, its only after creating the jukebox. But seeing as you are not using Musly, its a mute point.
With regards to your DB, without the music I'd have no idea what to expect.

Actually, I have been using the Similarity weight at 100, just to focus on the high-level attributes. I haven't played much with the Musly side of things. Still on my To Do list. I don't feel I have enough experience to write with any confidence about my experience, even the Heart example could have been circumstantial. In Essentia scoring, I would put my confidence around 0.08425123, give or take.

If your confidence in the Essentia scoring is so low, why use it? (I get the number is made up, but it does imply a very low confidence)

[prabbit]

If your confidence in the Essentia scoring is so low, why use it? (I get the number is made up, but it does imply a very low confidence)

It's not fair to Essentia/Similarity for me to say my confidence is low. That was perhaps awkwardly written. What I meant was that my sample size (i.e., the number of mixes I have created) is not large enough for me to be confident in my confidence of Similarity/Essentia, if that makes sense. And since I don't have a musical ear, I couldn't tell you whether a song was written in B# or E♭ nor do I know exactly what sections of each song Essentia analyzed to create its own values so I could listen to that same section and score it myself (a task I, frankly, am not interested in completing). Currently, I have no reason to doubt the quality of the data in my database. I just don't have enough experience with it; ergo, my low confidence. I expect my confidence will rise with more experience, which will inform what values I choose to create mixes that I consider interesting. I consider this normal, since I am a new user and as a new user I just don't know how the system works or reacts when I pull various levers. It's not a critique of the system or the data; it's of me. When I run "Similar tracks" or "Create similarity mix" and then compare any individual result with the seed track by looking in the database, the high-level attributes are indeed similar, explaining to me why the song was selected. And in some ways that's exactly why I use Sugarcube or SmartMix and now Similarity. I want to hear songs near each other that I wouldn't otherwise have thought of.

[slartibartfast]

It's not fair to Essentia/Similarity for me to say my confidence is low. That was perhaps awkwardly written. What I meant was that my sample size (i.e., the number of mixes I have created) is not large enough for me to be confident in my confidence of Similarity/Essentia, if that makes sense. And since I don't have a musical ear, I couldn't tell you whether a song was written in B# or E♭ nor do I know exactly why sections of each song Essentia analyzed to create its own values so I could listen to that same section and score it myself (a task I, frankly, am not interested in completing). Currently, I have no reason to doubt the quality of the data in my database. I just don't have enough experience with it; ergo, my low confidence. I expect my confidence will rise with more experience, which will inform what values I choose to create mixes that I consider interesting. I consider this normal, since I am a new user and as a new user I just don't know how the system works or reacts when I pull various levers. It's not a critique of the system or the data; it's of me. When I run "Similar tracks" or "Create similarity mix" and then compare any individual result with the seed track by looking in the database, the high-level attributes are indeed similar, explaining to me why the song was selected. And in some ways that's exactly why I use Sugarcube or SmartMix and now Similarity. I want to hear songs near each other that I wouldn't otherwise have thought of.

Have you tried Craig's MusicIP mixer plugin? It also works with "Don't Stop the Music".

Sent from my Pixel 3a using Tapatalk

[prabbit]

Have you tried Craig's MusicIP mixer plugin? It also works with "Don't Stop the Music".

I haven't. I don't see what advantage it would offer over Spicefly Sugarcube that I've been using for years, given that it would pull from the same MusicIP database but has fewer options than Sugarcube, especially TrackStat-related. I could see it potentially being interesting for anyone who doesn't have or doesn't want to purchase a Sugarcube license.

[cpd73]

Imagine I want to query the music-similarity database for all songs with these values and produce a playlist—that was SmartMix and that's what I think the purpose of those values would or could be.

• happy: > 0.9
• aggressive: 0.4 - 0.6
• danceable: >0.75
• electronic: >0.8
• dark: >0.5

I have started work on this. Very early stage, but if you are interested you need to use the '5-smart-mixes' branch from your git checkout of lms-musicsimilarity. (Not sure about the name 'Smart Mix' will probably change). No way to create, edit, or delete these at the moment. But the plugin will list all ".attrmix" (e.g. "Sad songs.attrmix") files in LMS's playlists folder. You can then click on a "smart mix" and the new API is called - creating a mix of tracks. The file format is a key=value", with the keys detailed in https://github.com/CDrummond/music-s...er/docs/API.md e.g.

minbpm=30
maxbpm=100
danceable=52
count=20
genre=Pop;Dance

Note, 'Smart Mix' entry is shown under 'My Music' in Material but does not work with Default skin. Also, LMS will query music-similarity at start-up to see if Essentia is enabled - so its proably best if music-similarity is running before LMS. The plugin does check every 15 seconds for up to 5 minutes to obtain the Essentia level - so should not be required to already be started, but might help.