Song similarity and automatic playlist generation

[NeptuneHub]

Hi,
I would like to have the possibility to automatically playlist creation and instant mix based on similar song. All of them based on the analysis of the song itself (someone call it sonic analysis).

I implemented a minimum viable product that you can find here:
https://github.com/NeptuneHub/AudioMuse-AI

It actually interact with Navidrome API (in this moment in :devel branch) to download and analyze song and then to create playlist.
It use Librosa and Tensorflow for the song analysis and it is written in python. It has a test front-end but the core are Flask API. For an easy test is all contenirezed: deployment.yaml and docker-compose.yaml example are present in the repo.

It will be great to have those functionality directly integrated in the Navidrome front-end instead of relay on an external front end. Especially because the “song similarity” feature could be used for instant mix if integrated instead of just create playlist.

The pro of doing song analysis over using external service API like lastfm is multiple:

• The privacy is higher, that for a self hosting is an important feature;
• It doesn’t need additional account, api key and similar;
• It is more powerful than just known a genre. It depict the song as an embedding vector and then enable to find similarity patterns that go behind of simple genre similarity.

I develop this testing product because till now I didn’t found any open source, self hostsble app with this feature. And I hope that this testing product could show the potentially of this song analysis feature to be integrated in more mature product like Navidrome.

[mrSaT13]

A great idea is an analog of spotify for creating playlists.that's fine, and you're right about privacy.

[danieldids]

AH! Nice to see more people looking into this thing. I'm also doing my own research about how to create spotify-like recommendation playlists, and I've found another kind of solution.

First of all, I've found this:
https://cosine.club/

This project basically has a large music library and they've made some sound similarity analysis that does not rely on AI - which is good because AI is usually computationally expensive.

Edit: cosine does use AI. Anyway, it still is an interesting project.

Then, I've looked further into this subject, and found this:
Deprecated but still informative:
https://github.com/CDrummond/music-similarity?tab=readme-ov-file

Current project:
https://github.com/CDrummond/bliss-analyser

It's a rust-based tool for analyzing audio files and creating data that can be used to compare them and suggest similar songs.

Yesterday I've analysed my collection and now I'm trying to create some playlist based on the information collected. I will do some evaluation (subjective, of course) of the results and intend to report back here.

The easiest way to try it is by running the executables provided here:
https://github.com/CDrummond/bliss-analyser/releases/tag/0.4.0

If this thing works well, I'd like to see it integrated into the current smart playlist feature that's being developed in navidrome.

@deluan @caiocotts I'm also a go developer. I'd like some directions of how I could implement this in a way that's easy to merge into navidrome, instead of creating an external tool that's used by it. bliss-analyser creates just one sqlite table and can use ffmpeg to decode audio files into PCM for analysis. The other dependencies are C and Rust based, I'm checking out how hard is to wrap then with go, or maybe do some binding (I've never worked with C binding in Go).

P.s.: English is not my native language, cheers from Brazil!

[deluan]

Current project: https://github.com/CDrummond/bliss-analyser

If this thing works well, I'd like to see it integrated into the current smart playlist feature that's being developed in navidrome.

@kgarner7 has been using bliss-rs (which is used by bliss-analyser) in a semi-integrated way in his Navidrome fork, with what seem to be promising results.

We’ve been discussing ways to allow analyzers and playlist generators like this to be implemented as plugins. Join the #dev channel on Discord to join the conversation. (Although I’m personally taking a one-month break before returning to Navidrome.)

@deluan @caiocotts I'm also a go developer. I'd like some directions of how I could implement this in a way that's easy to merge into navidrome, instead of creating an external tool that's used by it. bliss-analyser creates just one sqlite table and can use ffmpeg to decode audio files into PCM for analysis. The other dependencies are C and Rust based, I'm checking out how hard is to wrap then with go, or maybe do some binding (I've never worked with C binding in Go).

We generally want to avoid adding more non-Go dependencies to the codebase, since this complicates building and packaging across multiple platforms. The goal is to support plugins (written in Go or WASM, with support for other WASM-compatible languages in the future) that can invoke external tools via command line, enabling the use of tools like bliss-rs.

As mentioned, it’s best to join the Discord discussion to explore some ideas.

P.s.: English is not my native language, cheers from Brazil!

Opa, mais um conterrâneo!

[NeptuneHub]

Just to add that AudioMuse-AI not only analyze tempo and energy of the song but use Tensorflow for deep learning and extract the embedding, a 200 dimensional vector rappresentation. Than with it use similarity index like Voyager (from Spotify) to search for similar playlist and if integrated in navidrome it can be used to create instant mix on the fly.

The bast way to run this analysis is on a side container, more than directly integrated in the media server this for different reason:

• don’t be limited from the technology of the media server itself;
• It an high computational work and having a container implementation do it easy to scale by just deploy more container.

My suggestion, if it possible, is some kind of front-end plugin that than enable to talk with AudioMuse-AI API directly integrated in the Navidrome fronted.

For jellyfin I developer the plugin, I don’t know if there is possibility to develop plugin also for Navidrome.

[deluan]

Just to add that AudioMuse-AI not only analyze tempo and energy of the song but use Tensorflow for deep learning and extract the embedding, a 200 dimensional vector rappresentation. Than with it use similarity index like Voyager (from Spotify) to search for similar playlist and if integrated in navidrome it can be used to create instant mix on the fly.

Navidrome plugin system is still in its infancy, and for now it does not allow changes to the UI. But the goal is to allow plugins to integrate with external tools like yours as well.

[danieldids]

Well, I've explored bliss-rs a little bit more.

As of now, I've got an executable file which does the actual media file analysis, and I'm wrapping that with golang (exec.Command for the rescue). It works, it's not the best solution of course.

Each track analysis takes around 2 (two) seconds (in my old and faithful i5-2500). The analysis results are saved into a SQLite table, and they reference a media file ID as it is stored in navidrome's database.

Track similarity is done with a simple Euclidean distance measurement on the 20 attributes provided by bliss-rs analyser.

Now I'm assessing how good is this similarity measurement.

It does not look that good for "small" (~5600 tracks in my collection) collections. :-D

I'm checking what's going on, but I'm confident this is going to work. Maybe I need to add in some weights to the similarity measurement.

[danieldids]

I was wrong. It looks good. Most of the time.

The picture bellow is a sample of one of the playlists I've generated. The first track is what I'm calling reference track. It's the one I want to be used as "seed" for a playlist. All the other tracks are ordered by similarity, from the most to the least similar.

There are some duplicates. That's a good sign: in this test, the reference track (La Mer, by Claude Debussy) appears multiple times because I have multiple interpretations (?) - different orchestras and conductors - of it.

Out of 25 tracks (1 reference track + 24 similar ones), only two are not classical music: An inmates lullaby, by Gentle Giant, and Farolim de Felgueiras, by Khruangbin. Given that my classical music library is really small (25 albums only), and this playlist generation was allowed to scan all collections, the result looks good.

[danieldids]

Another sample, 25 tracks in total:

The track in bold was the reference track of a group from Cuba - Buena Vista Social Club. The selected tracks (15, in blue) are Brazilian artists. So there is quite a lot Latino similarity in this playlist. :-D

For the other artists/tracks, I would say there's a similar sound texture for all of them. Unfortunately, I can't prove this feeling here. But it's a good playlist anyway.

[danieldids]

Ok, I think it's official: I'm going to create a separate web application so people can test this tool, experiment with it, and give me feedback.

It will work for playlists creation only, no player functionality. Also, I'd like to make it work together with navidrome. m3u files should be available too.

This baby has no name for now. Suggestions are welcome. navidrome-bliss maybe?

[danieldids]

But why?

I mean, having different alternatives is always good, but what did you think is missing from AudioMuse-AI and why you think there is not the possibility to improve it?

Something else: based on your comment, I've got the feeling that you think I'm helping with developing navidrome. No, I'm not in any means related to navidrome development, I'm just another guy that likes self hosting sutff (including jellyfin too).

Cheers!

[NeptuneHub]

But why?
I mean, having different alternatives is always good, but what did you think is missing from AudioMuse-AI and why you think there is not the possibility to improve it?

Something else: based on your comment, I've got the feeling that you think I'm helping with developing navidrome. No, I'm not in any means related to navidrome development, I'm just another guy that likes self hosting sutff (including jellyfin too).

Cheers!

AudioMuse-AI is lightweight and runs smoothly on modest hardware — I’ve tested it on a Raspberry Pi 5 (ARM) and an Intel i5-6500, and it should work on even older systems.
Song analysis time depends on your hardware and the number of tracks, but it’s only required during the first run.

The analysis uses Librosa and TensorFlow. From this, you can either run clustering for automatic playlist creation or use the Spotify Voyager Index for similar-song discovery and instant mixes. AI is optional — it powers an experimental feature where you can describe what you want, and it generates the query for you. AudioMuse works perfectly without it.

It integrates with both Navidrome (Subsonic API) and Jellyfin. For Jellyfin, I’ve developed a plugin for direct use in the Media Server front end; I’d like to see something similar for Navidrome.

Deployment is straightforward with Kubernetes (tested on K3S) or Docker Compose. The project is open source, so there’s always room for contributions and improvements.

For current progress, see the “Support in test” issues:
https://github.com/NeptuneHub/AudioMuse-AI/issues
The results aren’t broken down song-by-song (since we process thousands), but they give a good overview.

This feature request aims to improve Navidrome integration so users can enjoy a unified front end. Algorithmically, the results are already strong — and with community help, they can get even better.

Just give a try, then if you think you’re proposing something alternative why not, is always good to have alternatives. But if we are doing “the same things” maybe could be useful work on the same project.

[danieldids]

I'm trying it right now, with a remote navidrome instance. It's going to take quite a long time to finish processing my whole library. What I can already say is: this is a completely different beast (in a good way). :-D

From what I'm seeing, AudioMuse and my future app do similar things differently. Mine is probably not going to give the best results (just good enough), but works using very few resources (memory, CPU, disk space), while yours is feature rich while requiring more resources and time.

I'm posting a sample analysis bellow just because I find its results are amazing!

audiomuse-ai-worker-instance  | [INFO]-[11-08-2025 18-43-00]-[AlbumTask-97c74b03-6d47-4af4-a617-52d0c12e49c1-Our Love Is Here to Stay: Ella & Louis Sing Gershwin] Analyzing track: Things Are Looking Up by Ella Fitzgerald & Louis Armstrong (7/16)
audiomuse-ai-worker-instance  | [INFO]-[11-08-2025 18-43-01]-Downloaded 'Things Are Looking Up' to '/app/temp_audio/S3O886HuJ9zXAqahPiNNL9.mp3'
audiomuse-ai-worker-instance  | [INFO]-[11-08-2025 18-43-01]-Starting analysis for: S3O886HuJ9zXAqahPiNNL9.mp3
audiomuse-ai-worker-instance  | [INFO]-[11-08-2025 18-43-06]-SUCCESSFULLY ANALYZED 'Things Are Looking Up by Ella Fitzgerald & Louis Armstrong' (ID: S3O886HuJ9zXAqahPiNNL9):
audiomuse-ai-worker-instance  | [INFO]-[11-08-2025 18-43-06]-  - Tempo: 133.93, Energy: 0.1035, Key: D minor
audiomuse-ai-worker-instance  | [INFO]-[11-08-2025 18-43-06]-  - Top Moods: {'jazz': 0.48041975498199463, 'oldies': 0.2108180969953537, 'female vocalists': 0.18554438650608063, 'soul': 0.07545589655637741, 'pop': 0.07332172989845276}
audiomuse-ai-worker-instance  | [INFO]-[11-08-2025 18-43-06]-  - Other Features: danceable:0.15,aggressive:0.01,happy:0.56,party:0.22,relaxed:0.78,sad:0.82

[NeptuneHub]

Remember that what you showed are only the human readable information. It also capture for each song Musicnn embedding that are 200 elements vector used for find similar song and instant mix. You can really tap on a song and find the similar one that are really similar (not only same tempo or energy).

The analysis is CPU intensive but run also on old cpu, it only takes time. But analysis need to be done only do the first time, and then you just run time to time for the new added song.

Also this run on Kubernetes, for example I have a small cluster of 4 old computer, and audiomuse parallelize the work among all of them so what it usually can takes 1 weeks you do in 3 days if you have more old computer to use.

I really looking forward for now about your feedback and if something need to be improved, why don’t improve it together ?

[danieldids]

I got to say: I'm sorry. Just now I've realized that this is your feature request. :-) I got excited about the whole thing being easier than I thought - in the sense that there are already many tools that can be used.

I have one question: are you talking to navidrome using its subsonic API or something else? You say AudioMuse interacts with navidrome API but it's not clear to me what you mean.

[NeptuneHub]

Just to stay on AudioMuse-AI topic, do you have any feedback on it?
Is there any possibility to plan for his integration on Navidrome?

About the plugin integration which feature is possible to integrate going down this route ?
At the moment I would need to:

• run scheduled task automatically (daily/weekly) by calling an API;
• have the possibility, in some way, to click on a song and run an instant mix by calling an API that suggest similar song (so if it already exist as a functionality, maybe just ovverriding it).

Is it possibile to integrate this functionality as a plugin?
If not, is there any possibility to integrate this api directly in Navidrome as a next feature ?

[danieldids]

Just a snapshot of what I'm doing:

[danieldids]

Also:

[danieldids]

As soon as possible, I'll open this project for people to try.

[danieldids]

This screen is just a mock, but illustrates well the concept:

[githubdotcomuser]

This screen is just a mock, but illustrates well the concept:

Are you still working on this?

[danieldids]

Are you still working on this?

Unfortunately, no... I've go a lot of work lately. :-(

[Th3K1ngP1n]

Just dropping in to say that I have installed AudioMuse and it is working fantastically. Only thing missing for me is the automatic syncing and renewing of playlists to navidrome on a regular basis. A lot of people also use symfonium as a client on Android. The most recent beta should allow sonic analysis via jellyfin/audiomuse plugin. Would be awesome to have this integration with navidrome in the future.

TLDR; Thanks from me and looking forward to future updates 😀

[danieldids]

The tool I'm making will do that - automatic sync of tracks from multiple sources, analysis, and playlist generation.

I'm only targeting subsonic API compliant platforms - navidrome, subsonic, gonic, LMS etc. I don't know if jellyfin implements it. If so, it won't work for now.

[NeptuneHub]

The tool I'm making will do that - automatic sync of tracks from multiple sources, analysis, and playlist generation.

I'm only targeting subsonic API compliant platforms - navidrome, subsonic, gonic, LMS etc. I don't know if jellyfin implements it. If so, it won't work for now.

Nice, keep us posted!

[danieldids]

I will, don't worry :-)

[danieldids]

Just to keep you people anxious :-)

[cslamar]

Any movement on this?

[Th3K1ngP1n]

I am also super curious. @NeptuneHub has also been working hard on their plugin. Just pulled the latest version at the weekend and currently re-analyzing my tracks.

[danieldids]

Any movement on this?

Yes. I'm still working on it.

[isZYKerman]

Is there a github repo or something that we can try it out?

[danieldids]

Is there a github repo or something that we can try it out?

Not yet. Right now, I have my plate full of job stuff to do, and I'm not having the proper time to finish Jujubaz.

That said, I'm already using it personally. What is missing to release a very early MVP alpha version is user authentication. :-)

At first, it will only work on x86 linux machines.

[NeptuneHub]

From my point of view I would really like to ask to Navidrome developer to directly integrate AudioMuse-AI in their code base.

For direct integration I mean by API so:

• Navidrome backend call AudioMuse-AI api and expose this functionality directly in his backend;
• Navidrome front-end and any other external app then can directly integrate in the front-end the Sonic Analysis based functionality by just calling the Navidrome backend.

I know that not anyone will use AudioMuse-AI, but it could be definitely add as a functionality that are enabled only when audiomuse-ai is configured.

I’m asking this because Sonic Analysis functionality could be more than just an “instant mix”, they could really redesign the way on which the user listen to the music. For example by creating sonic analysis based radio, or by visual representing the song library and so on.

If navidrome developer want to go for a full integration route off course I’ll do all is possible for me to help and support them.

[Th3K1ngP1n]

@deluan What do you think.? I know you are working hard on many things. But would this be a possibility?

[daniele-athome]

I believe a plugin system is being designed somewhere: this might fit as a plugin. IMHO I wouldn't link the two products so tightly together (i.e. think about different implementations and also the role of OpenSubsonic API in this).

EDIT: not a Navidrome dev, just my two cents as an interested Navidrome user.

[CustardFlan]

I'm interested in full integration of AudioMuse-AI with Navidrome too, i don't want another UI for something that can be done in the server, i use Symfonium and this can be useful for the radio mixes or for discover new music in my library (nearly 84200 songs)

[deluan]

Thanks @NeptuneHub and @danieldids for working on this.

From my point of view I would really like to ask to Navidrome developer to directly integrate AudioMuse-AI in their code base.

As @daniele-athome mentioned, the ideal would be to create a plugin to connect Navidrome to external playlist sources. There's an example created by @kgarner7 here.

Let me know if the current plugin API is enough for creating such integration.

[Th3K1ngP1n]

@danieldids That sounds really nice. I am however more interested in an "on the fly" "instant mix" method of using Audiomuse and Navidrome together, through smart queuing of tracks using the data from the sonic analysis, as it really expands the listening experience in a unique way, discovering tracks similarities etc.

This is also working with Symfonium and Jellyfin, but, I still see Navidrome being the superior music server, for my needs at least. Smart playlist generation with Navidrome is already possible, and, to have it automised in a scheduled manner daily / weekly / monthly would indeed be pretty cool.

I would still be very interested in trying Jujubaz when you are ready.

[danieldids]

What do you mean by "on the fly" "instant mix" ? Is it like the playlist being created as the user listens to music?

[Th3K1ngP1n]

What do you mean by "on the fly" "instant mix" ? Is it like the playlist being created as the user listens to music?

Not a playlist as such, but, a queue of relevant tracks in accordance with the settings.

Here it is in symfonium
https://support.symfonium.app/t/smart-queue-and-smart-flow/7042

Example: I play a Song A and hit echo match, it will now continue playing songs sonically related to song A.

[isZYKerman]

I believe limiting audiomuse-ai to generating simple playlists is a waste of its potential. This AI should be integrated to provide a more intelligent and dynamic music experience within Navidrome, offering features that are standard in modern streaming services. These features are the "soul" of a music player, and implementing them would bring Navidrome to life :) .

1. Intelligent "Random Sorting" (Contextual Shuffle)
   This feature would allow users to exert some control over the "random" playback of a playlist, making the shuffle smarter and more attuned to the user's current mood or preference. When playing a playlist on shuffle, the AI would analyze the upcoming tracks and offer intelligent sortings or filters for the next few songs.
   Click here for the referrence screenshot; this picture is too long so I think not displaying it directly is better.
   For example, I am playing a shuffled playlist but currently wants to hear only Fast-Paced songs or tracks from a specific Genre or Mood for the next part of the session. In Netease Music I can open a menu, like in the screenshot above, to select this context, ensuring the "random play" is less random and more engaging.

2. Personalized Daily Mixes and Listening Summaries
   The AI should proactively curate content based on the user's history and library.
   Daily Mix: Automatically generated, highly personalized playlists (e.g., "Mix for You") based on listening habits, similar to popular streaming services.
   Weekly/Monthly Conclusions: Offering data-driven summaries of the user's listening habits (e.g., top artists, genres, total listening time).

3. Track-Based Discovery and Suggestions
   Move beyond album-centric organization to focus on individual track discovery, which is the preferred method for users migrating from streaming services.

To implement the functions above, maybe a dedicated new page should be added to the UI to host all AI-driven recommendations, including:

• Tracks based on your "Favourites" (tracks marked as favorites).
• Tracks based on your recent listening history.
• Songs similar to the track you are currently listening to.
  These are quite similar to current music apps eg Spotify.

ps For many users, especially those migrating from streaming services like Netease Music or Spotify eg me, the listening focus is on individual tracks rather than whole albums (as buying a full album/cassette is no longer necessary). Currently, Navidrome's UI heavily features album pages, which are less relevant to us. A major UI overhaul, including the introduction of a "Suggestions" home page with daily mixes and track recommendations, will make Navidrome feel modern and dynamic. This is the current standard for most music software.
Click here for the referrence screenshot; this picture is too long so I think not displaying it directly is better.
Implementing these AI-driven features would be a massive leap forward for Navidrome, transforming it from a simple music server into a truly intelligent and personalized music player.

[danieldids]

I'm gonna steal those ideas, ok? :-) I can´t say they all will be on the first Jujubaz release, but I like them.

[isZYKerman]

Since we are already scrobbling songs to Last.fm and ListenBrainz, is it possible for Navidrome to integrate further with these services (Last.fm or ListenBrainz) to directly receive suggested tracks?

This approach offers several benefits:

• It requires less computing power compared to an AI-based solution.
• Their recommendation algorithms are more mature (e.g., Daily Mixes, Weekly listening summaries).
• It is more versatile, as recommendations could be limited to tracks within the user's library or extended to external songs by providing suggestions or links (e.g., Last.fm links) for discovery,.

[metalheim]

There is a plugin to get "Daily Jams" and "Weekly Jams" Playlists from ListenBrainz:
https://github.com/kgarner7/navidrome-listenbrainz-daily-playlist

Theoretically, plugins like this (that dont exist yet) could also import other playlists, like "Weekly exploration" from Listenbrainz or "last.fm recommended tracks"

[isZYKerman]

There is a plugin to get "Daily Jams" and "Weekly Jams" Playlists from ListenBrainz:\nhttps://github.com/kgarner7/navidrome-listenbrainz-daily-playlist

I've seen that. If I'm getting it right, it needs MBID tagged in the media file, which is rather inconvenient, especially for people like me who download loads of individual tracks from music streaming apps.

[Th3K1ngP1n]

Limiting Audiomuse/Navidrome to playlists generation only, is like having a sports car only to sit in it and not drive it.

Audiomuse already expose get similar songs, artists via its API amongst a bunch of other things that open the doors for other creative implementations.

Maybe there are some hack out of the box ways to get similar results with LastFm etc. But LastFm is not free or open source. Audiomuse is in very active development and I think the DEV is pretty open to ideas and suggestions on improving the service. At least from my experience. Apart from the Analysis / Clustering jobs. It is also not very resource heavy.

There is an open discussion about getting it implemented into opensubsonic API, and, my opinion is that its the way forward, because other servers also use this API, and, it would open freedom for more than just navidrome. But it seems we need to wait for that.

[mrSaT13]

The Audiomuse/Navidrome limitation is for playlists only, it's like having a sports car just to sit in it and not drive it.

Audiomuse is already exposing a similar version to artists through its API, among many other things that open the door to other creative implementations.

Maybe there are some hack-out-of-the-box ways to get similar results with LastFM, etc. But LastFM is not free or open source. Audiomuse is in very active development, and I think DEV is pretty open to ideas and suggestions for improving the service. At least in my experience. In addition to analyzing/clustering jobs. It's also not a very heavy resource.

There is an open discussion about embedding it in an open subsonic API, and in my opinion, this is the way forward because other servers also use this API, and this will open up freedom for more than just a flood path. But it looks like we need to wait for that.

there is also a restriction on users, all of which is generated for only one user.

[rivpyl-sidfyl]

I'm not sure if this is an appropriate forum for this kind of comment, but I just want to say that I've been using AM on my navidrome library - which is relatively big and full of music that I don't really know - and it's just been an incredible way to generate playlists that allow me to discover unfamiliar music through various methods.
This combined with last.fm/listenbrainz similar music features is basically the perfect self-hosted alternative to spotify-type algorithmic discoverability. As a non technical guy, I would support in any endeavour to integrate this service deeper into navidrome. Cheers!

[PascPeli]

Thank you @NeptuneHub for creating this PR and this amazing product.

I also (similarly to others) logged here just to say that integrating AM to Navidrome should be a no brainer.

What AM is offering is much needed to bring our music experience in similar levels as paid streaming services. Without it I'm afraid Navidrome is like a library without a librarian. And don't get me wrong I love Navidrome.

So, what I suggest is:
Seamless integration, one UI, simple and powerful.

Thanks you all and please Keep us posted

[coacharnold1]

So .. I use Navidrome heavily while on the go remotely. In my house i have been a heavy user of Snakeoil Server with LMS ... ive found for my system that MPD is a much better solution ... MY POINT for this discussion ... ive made a pretty large front end for MPD that has some of the spottily recommended playlist features your looking for on Navidrome, integrating LAST FM's api heavily .. please re-use the code if it at all helps for this ... https://coacharnold1.github.io/Maestro-Server/

[NeptuneHub]

Hi all,
Finally back to share that the AudioMuse-AI plugin for Navidrome is out and downloadable here:

• https://github.com/NeptuneHub/AudioMuse-AI-NV-plugin

Actually it work only with the develop container image of Navidrome, but the stable version should be out in days. Meanwhile you can test it and share your feedback !

Thanks so much to @deluan for his effort to make this possible !

[craiglush]

I built a plugin that does exactly this! Mood Playlists — audio analysis-powered mood playlists and mood-aware Instant Mix for Navidrome.

What it does

• 6 auto-generated mood playlists — Happy, Chill, Energetic, Melancholy, Party, Aggressive — refreshed on a cron schedule
• Mood-aware Instant Mix — Replaces the default Instant Mix with Euclidean-distance mood-vector similarity matching across analyzed tracks
• Scheduled audio analysis — Periodically scans your library and analyzes tracks using an external essentia-tensorflow service with Discogs-EffNet embeddings
• Genre-aware scoring — Context boosts for 25+ genres (e.g., DnB gets danceability correction, metal gets aggression boost, ambient gets relaxed boost)
• Fully configurable — Mood thresholds, playlist sizes, schedules, analyzer URL — all from Navidrome's plugin settings UI

How it works

The plugin ships as a .ndp file (WASM plugin) + a companion Docker container running essentia-tensorflow. The analyzer extracts mood scores (happy, sad, relaxed, aggressive, party), danceability, BPM, and energy from your actual audio files, then applies genre/BPM-aware context boosts for better accuracy. Mood data is stored in the plugin's KVStore and used to generate playlists and power similarity-based Instant Mix.

Install

1. Download mood-playlists.ndp from Releases
2. Start the analyzer service (Docker image included)
3. Drop the .ndp in your plugins directory
4. Configure via Settings > Plugins

Requires Navidrome develop branch (0.61.0+ with plugin support). Full details in the README.

Would love to hear feedback from anyone who tries it out!

[craiglush]

I've built a plugin that does exactly this — audio analysis-powered mood playlists and mood-aware Instant Mix.

Mood Playlists Plugin (v0.2.0)

It uses essentia-tensorflow with Discogs-EffNet embeddings to analyze your library for mood (happy, sad, relaxed, aggressive, party), danceability, BPM, and energy. Then it automatically creates 13 mood playlists:

• 6 simple moods: Happy, Chill, Energetic, Melancholy, Party, Aggressive
• 7 composite scenarios: Study, Workout, Sleep, Road Trip, Cooking, Dining, Background (using multi-condition filtering, e.g. Study = high relaxation AND low energy AND low aggression)

It also replaces Instant Mix with mood-similarity matching — Euclidean distance across the full mood vector, so "similar songs" actually sound similar.

The analyzer runs as a companion Docker container and applies genre-aware context boosts (so DnB gets a danceability boost, metal gets aggression, etc.).

• Download: https://github.com/craiglush/navidrome-mood-plugin/releases/tag/v0.2.0
• Requires: Navidrome develop branch (0.61.0+ with plugin support)

[Th3K1ngP1n]

@craiglush I love the idea of your plugin but it requires a separate analysis of my library no? Or can it plug directly into the Audiomuse database?

Also do you know when Navidrome 61.0 is expected to be released? 😀

[NeptuneHub]

I just miss to update this issue: because the plugin system is there, I was able to do the integration with the plugin. So for me this issue is good to be closed. Thanks.

[craiglush]

Hi, Well it generates moods based on analysing the actual music, one track at a time. It definitely needs a separate model to analyse, but it doesn't need heavy firepower. You can set mood for particular genres to, or even complimentary ones. Thanks Sent from Outlook for Android<https://aka.ms/AAb9ysg> Craig Lusher Principal Solutions Architect [cid:c8signature-ico-telephone_c13147e2-9017-48dc-b4b9-a130b5d0f363.jpg] +441624694646 [cid:c8signature-ico-mobile_7b9546c5-d5c7-4559-8b6a-580ca7d73e52.jpg] +44 7921 399499 [cid:c8signature-ico-email_796bf7d7-b437-440f-93ae-39d35644a294.jpg] ***@***.******@***.***> [cid:c8signature-ico-website_575254f7-d4c7-4f50-b0e1-30f037a6722c.jpg] continent8.com<https://www.continent8.com> [cid:c8signature-facebook_3749357b-91aa-4f11-b614-3364c5997484.jpg]<https://www.facebook.com/Continent8Technologies/> [cid:c8signature-twitterlogo_44x44_2a3cd9e0-30a0-4063-a686-2bee2378b892.jpg] <https://twitter.com/continent8?lang=en> [cid:c8signature-linkedin_a2b3594e-0438-458f-bf52-221cb1f15737.jpg] <https://www.linkedin.com/company/continent-8-technologies> [cid:c8signature-mainlogo_fullcolourrgb_d59f52bf-e544-463f-9890-d0849fc497ad.png] <https://www.continent8.com/> [cid:c8signature-blueprintes_v4_f3df7828-edf0-4e4b-a311-cb1ad7a687d9.png]<https://www.continent8.com/blueprint-gaming-strengthens-security-and-accelerates-threat-response-with-continent-8s-managed-soc/>

…

________________________________ From: Th3K1ngP1n ***@***.***> Sent: Tuesday, March 24, 2026 5:17:53 pm To: navidrome/navidrome ***@***.***> Cc: Craig Lusher ***@***.***>; Mention ***@***.***> Subject: Re: [navidrome/navidrome] Song similarity and automatic playlist generation (Discussion #4332) *** This email is from a non-C8 source *** @craiglush<https://github.com/craiglush> I love the idea of your plugin but it requires a separate analysis of my library no? Or can it plug directly into the Audiomuse database? Also do you know when Navidrome 61.0 is expected to be released? 😀 — Reply to this email directly, view it on GitHub<#4332?email_source=notifications&email_token=AR425A4DOLZIIBFSW5PVAVL4SK7LXA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNRSHE3DMMZYUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-16296638>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AR425A5ZE4RIBBRUTIYDAJT4SK7LXAVCNFSM6AAAAACBKQ7PP2VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMMRZGY3DGOA>. You are receiving this because you were mentioned. This email, along with any attachments, is for the sole review and use of the intended recipient(s) and may contain confidential and/or legally privileged information belonging to the sender. Delete this email and notify us if you are not the intended recipient.