[FEATURE] Improve multi-artist handling

[spcfox]

Describe the bug
AudioMuse interprets the artist metadata field as a single literal string rather than a list. Consequently, any track featuring multiple artists (e.g., "Artist A; Artist B") is treated by AudioMuse as a single, unique artist entity. This results in the creation of numerous combined artist entries instead of associating the track with the individual artists involved.

To Reproduce
Steps to reproduce the behavior:

1. Navigate to the Artist Similarity search bar in AudioMuse
2. Search for an artist that frequently collaborates with others
3. Observe the search results dropdown

Expected behavior
AudioMuse should parse the artist metadata field and handle multiple artists as a list. Tracks should be correctly associated with each individual artist name internally, ensuring that the "Artist Similarity" search returns a single, unified entry for each artist regardless of their collaborations.

Screenshots

Environment (please complete the following information):

• OS: Windows
• Deployment: Docker Compose
• AudioMuse-AI Version: v0.9.3
• Jellyfin/Navidrome Version: Navidrome 0.60.3 (34c6f12a)

Additional context
It is important to note that Navidrome and compatible players properly handle multiple-artist strings.

[spcfox]

It seems that Navidrome returns a list of artists within the track information. While this might not strictly be part of the base Subsonic standard, it would be fantastic to have support for this structure.

Example:

{
  "artist": "aikko • INSPACE",
  "artists": [
    { "id": "3kb7JkgNlwiKjvc27kN2x2", "name": "aikko" },
    { "id": "2Bj7t3gaR7Pd130Aqw0oN0", "name": "INSPACE" }
  ],
  "displayArtist": "aikko • INSPACE",
  "albumArtists": [
    { "id": "3kb7JkgNlwiKjvc27kN2x2", "name": "aikko" },
    { "id": "2Bj7t3gaR7Pd130Aqw0oN0", "name": "INSPACE" },
    { "id": "3xbHzbfXH4EAgLGXZxwEWx", "name": "katanacss" }
  ],
  "displayAlbumArtist": "aikko • INSPACE • katanacss"
}

[spcfox]

Oh, it is part of OpenSubonic
navidrome/navidrome#2695

[NeptuneHub]

Hi and thanks for your suggestion. Before starting I would highlight that AudioMuse-AI support multiple media server so any change should be shared against all of them when possible (Jellyfin, Navidrome, Emby and Lyrion).

Looking at navidrome implementation we save in our db "Author" following this prioritization present in mediaserver_navidrome.py:

def _select_best_artist(song_item, title="Unknown"):
    """
    Selects the best artist field from Navidrome song item, prioritizing track artists over album artists.
    This helps avoid "Various Artists" issues in compilation albums.
    Returns tuple: (artist_name, artist_id)
    """
    # Priority: artist (track artist) > albumArtist > fallback
    # Navidrome provides artistId and albumArtistId
    if song_item.get('artist'):
        track_artist = song_item['artist']
        artist_id = song_item.get('artistId')
        used_field = 'artist'
    elif song_item.get('albumArtist'):
        track_artist = song_item['albumArtist']
        artist_id = song_item.get('albumArtistId')
        used_field = 'albumArtist'
    else:
        track_artist = 'Unknown Artist'
        artist_id = None
        used_field = 'fallback'
    
    return track_artist, artist_id

And this off course match what you share in your example where you visualize "artist": "aikko • INSPACE".

Now before going forward on the technical implementation, let's look the logic: here we are saing that:

• Artist1 => sound in a way;
• Artist2 => sound in another way
• Artist 1 WITH Artist2 on the same track => sound in yet another way.

This is something that you can agree or not, but is not a bug, is a (maybe easy?) implementation choice.

If I understood correctly you're suggesting to store for each song the full list of Artist, is that correct?
This will means that when I do artist similarity, I'm saying that the above song will be used two time, one with Artist1 and one with Artist2. This will be an assumption that when Artist1 sound like Artist1+Artist2 and the same for artisst2.

If this is whay you're suggesting is not something that I agree.

Maybe you can say that the first artist is the primary (are we sure?) and that is this artist that influence the trait of the song. But even in this way I think we run the risk to lose information of "two artist when sound together are different then when sound alone".

And till now we are only talked of the logical side. The technical side means that I need to save multiple artist from the different music server and add additional logic that require important non regression test. Nothing of impossible, off course.

Can you better explain your point of view on this?

[spcfox]

Thank you for taking the time to provide such a detailed response.

Before starting I would highlight that AudioMuse-AI support multiple media server so any change should be shared against all of them when possible (Jellyfin, Navidrome, Emby and Lyrion).

As I understand it, Jellyfin and Emby support lists of artists within their APIs, so this change would be beneficial for them as well.

I realise that this is complex both technically and from a design perspective. I would like to outline the problems this creates and offer some suggestions. I do not have a readymade solution, and some of my ideas may not work in practice:

1. Duplicates: Currently, "Artist A; Artist B" and "Artist B; Artist A" are treated as different artists. As a first step, I think it would be helpful to compare artists regardless of their order. This would solve the issue of duplicating the same set of artists and help build the infrastructure for further multi-artist processing.
2. Sonic Profiles: You made a good point that collaborative tracks can create a unique sound, distinct from each artist individually. However, I believe this is typically because each artist brings their own contribution to the track. I think this is information that could be used fruitfully for better recommendations. I am not sure which mathematical model would be best; perhaps it is worth weighting each artist's contribution as 1 / (number of artists)?
3. Song Alchemy: This also affects how we work with artists in Song Alchemy. If I add an artist to an "exclude" list, do I want to see their collaborations? Maybe, but I think they should be less preferable.

I understand that this may not be a high-priority task, but I believe that treating every unique combination of artists as a new individual entity is a poor approach.

[spcfox]

If I understood correctly you're suggesting to store for each song the full list of Artist, is that correct? This will means that when I do artist similarity, I'm saying that the above song will be used two time, one with Artist1 and one with Artist2. This will be an assumption that when Artist1 sound like Artist1+Artist2 and the same for artisst2.

You are correct in your understanding. As I said, I do not have a robust mathematical model. However, I believe that linking both artists to the song is the good approach. While I am not certain which model would yield the best results, I hope this issue can serve as a platform for further discussion.

[cy1der]

I'm interested in this, I would like to chime in with some thoughts:

2. I am not sure which mathematical model would be best; perhaps it is worth weighting each artist's contribution as 1 / (number of artists)?

Weighing each artist's contribution to the track would be a good way to model this but your formula assumed that each artist in the song contributes an equal amount. For example, let's say Artist A sang 80% of the track and Artist B sang 20%, then the weights would adjust accordingly. Ideally, some way to automatically identify different artists in the song by splitting vocals and somehow do the magic there. An alternative, possibly better method is to use the lyrics of the song if available to calculate per-artist contribution, but that's not always the case. This also wouldn't work for hidden features (for that, maybe build a fingerprint of each artist's voice in order to identify and map them, not sure if that behavior is desired though) and potentially intrusive adlibs and background vocals. I'm no expert on this, just my 2 cents :)

3. If I add an artist to an "exclude" list, do I want to see their collaborations? Maybe, but I think they should be less preferable.

How about instead of excluding them fully, make it an adjustable threshold? For example, if the artist I don't want contributes more than 30% in a song, exclude it.