Audiomuse Ai
Audiomuse Ai: AudioMuse-AI is a self-hosted, Dockerized music playlist generator using sonic analysis to discover related tracks for Jellyfin, Navidrome, LMS, Lyrion, and Emby: no metadata or external APIs needed.
Last verified:
What is Audiomuse Ai?
Audiomuse Ai is AudioMuse-AI is an open-source, self-hosted music intelligence platform that performs deep sonic analysis on personal music libraries to generate automatic playlists and enable advanced music discovery. Built as a Dockerized microservices application, it uses machine learning models (Librosa, TensorFlow, ONNX) to extract acoustic features, mood classifications, and 200-dimensional audio embeddings locally without relying on external APIs. The system integrates with popular media servers including Jellyfin, Navidrome, Lyrion, Emby, and LMS through their native APIs.
Key features include automatic playlist generation through sonic clustering using multiple algorithms (K-Means, DBSCAN, GMM, Spectral), instant playlist creation from natural language queries via AI chat, similar song discovery using Voyager (Spotify's ANN library), sonic fingerprint playlists based on listening history, song path generation that creates sonic journeys between two tracks, and AI-powered playlist naming using Ollama, Gemini, or Mistral. The system also supports lyrics semantic search with 5 semantic axes for mood-based discovery, text search for lyrical meaning, and song similarity search weighted by lyrical and audio similarity.
AudioMuse-AI is designed for self-hosted music enthusiasts, home server users, and anyone with a personal music library who wants AI-powered music discovery without streaming service subscriptions. It works on both AMD64 and ARM64 architectures (including Raspberry Pi 5), requires 4-core CPU with 8GB RAM and SSD as suggested specifications, and optionally supports NVIDIA GPU acceleration for faster analysis. The project is in beta status and intended for testing purposes rather than production environments.
Audiomuse Ai pricing
Pricing model: Freemium
Completely free and open-source with no paid tiers or subscription plans. The project is developed as a community-driven effort for testing purposes. All features are available without cost including sonic analysis, playlist generation, AI chat, similarity search, sonic fingerprint, song path, and lyrics semantic search. Optional AI model providers (Ollama self-hosted, Gemini free tier, Mistral free tier) may have their own usage limits but AudioMuse-AI itself has no pricing. Docker containers are freely available from GitHub Container Registry with automatic nightly builds.
Audiomuse Ai pros
- Open-source and completely free with no subscription fees
- Self-hosted - all audio analysis happens locally without external APIs
- No reliance on streaming services or third-party music databases
- Integrates with multiple media servers (Jellyfin, Navidrome, Lyrion, Emby, LMS)
- Supports both AMD64 and ARM64 architectures including Raspberry Pi 5
- Multiple clustering algorithms available (K-Means, DBSCAN, GMM, Spectral)
- Instant playlist generation from natural language queries via AI chat
- Sonic fingerprint feature creates personalized playlists from listening history
- Song path feature generates sonic journeys between any two tracks
- Voyager index provides 99% recall for similarity search with less memory
- Optional NVIDIA GPU support significantly speeds up audio analysis
- Lyrics semantic search with 5 defined semantic axes for mood-based discovery
- AI playlist naming using Ollama, Gemini, or Mistral models
- Resumable analysis - interrupted analysis only re-analyzes missing songs
- Helm chart available for easy Kubernetes/K3S deployment
- Docker Compose files provided for local deployment
- Web UI accessible at port 8000 with setup wizard for configuration
- PostgreSQL database persists all analysis data for fast queries
- Redis Queue manages background job processing efficiently
- Can support multiple music libraries within single media server instance
Audiomuse Ai cons
- Beta status - intended for testing purposes not production environments
- Analysis can take 1 week+ for large collections (100k+ songs) on older hardware
- Requires minimum 4-core CPU and 8GB RAM which may be too much for older systems
- Intel I7 first gen or older CPUs don't work due to TensorFlow AVX requirement
- Some very old processors may have issues due to unsupported instructions
- GPU analysis still experimental despite NVIDIA support
- Collection Sync feature requires OAuth GitHub login and accepting privacy policy
- Centralized cloud database server currently offline limiting sync functionality
- Initial setup requires configuring environment variables or using setup wizard
- Clustering algorithm defaults tuned for 50,000-100,000 song collections
- Admin password reset requires direct PostgreSQL database access
- CLAP model analysis can be slow and may need to be disabled for speed
- Multiple worker containers needed for parallel analysis on large libraries
- Frontend parameters require understanding for optimal clustering results
- No official connection to audiomuse.ai website despite similar name
Frequently asked questions about Audiomuse Ai
What are the hardware requirements for AudioMuse-AI?
AudioMuse-AI works on both ARM and INTEL architecture. The suggested requirements are 4-core CPU and 8GB of RAM with SSD. Some very old processors could have issues due to unsupported commands. Intel I7 CPU of first gen or older don't work because TensorFlow requires AVX support. If you want to use the NVIDIA version, a GPU with 8GB VRAM is suggested. The system has been tested on HP Mini PC with Intel i5-6500 16GB RAM NVME SSD and Raspberry Pi 5 8GB RAM NVME SSD.
How do I deploy AudioMuse-AI?
The README section has explanations and multiple examples can be found in the deployment folder. From v1.0.0, only PostgreSQL, Redis, and TZ configuration must be configured via environment variables. All other configuration values are managed through the browser setup wizard and persisted in the database. The Setup Wizard is shown on clean installation as landing page and is also available later from the menu under Administration > Setup Wizard. If you're not able to reach the front-end on http://YOUR-IP:8000 or the analysis seems to finish without analyzing anything, it usually means some parameters are missing in your .env file.
How do I start using AudioMuse-AI after deployment?
After deployment, the first thing to do is access the AudioMuse-AI frontend, which is available at http://YOUR-IP:8000. From there, run the Analysis. This process collects information about your songs and stores it in your local database. Running the analysis is mandatory before you can use any other features. The analysis process is mandatory and must complete before clustering, similarity search, or any other features will work.
How long does the analysis take and what if I interrupt it midway?
The time required for the analysis depends on several factors, such as the number of songs to analyze and the hardware on which AudioMuse-AI is running. Depending on these factors, it can take anywhere from a few hours to several days. For big collections (100k+ songs) or old hardware, 1 week+ of analysis can be totally normal. The good news is that analyzed songs are stored in the database, so if the process is interrupted, you can restart it and only the missing songs will be analyzed.
Can AudioMuse-AI support multiple music libraries?
Yes, it can support multiple music libraries within a single media server instance (e.g., two separate music folders in one Jellyfin server). However, a single AudioMuse-AI instance cannot connect to multiple different media servers (e.g., one Jellyfin and one Navidrome server) at the same time. The parameters MUSIC_LIBRARIES can be used to match multiple music libraries on the same music server. It's a comma-separated list of music libraries/folders for analysis. If empty, all libraries/folders are scanned.
What if clustering returns empty playlists or with only a few songs?
The default clustering parameters are fine-tuned for music collections of around 50,000-100,000 songs. If your clusters are too small or empty, you can adjust these values in the Advanced Parameters view: Stratified Sampling Target Percentile defines the percentile of songs sampled per genre for clustering - the higher this value, the more songs will be clustered (you can set it to 100 to include more songs). You can also adjust min clusters and max clusters - by default AudioMuse-AI creates between 40 and 100 clusters (playlists). Lowering these numbers will result in fewer clusters, each containing more songs.
What if clustering returns clusters with too many songs?
In contrast to clusters with few songs, you can just raise the Stratified Sampling Target Percentile, min clusters, and max clusters values in the advanced parameter view. Increasing these values will create more clusters with fewer songs in each one, helping to distribute your music library more evenly across playlists.
How can I make clustering run faster?
Clustering algorithm by default does 5000 runs. This means that multiple runs are executed and the best is kept. You can lower this number in the front-end in the Clustering Runs field to do fewer runs. For example, with 1000 runs the result should still be good enough and take a reasonable amount of time. Reducing clustering runs significantly decreases processing time while still producing acceptable results.
How do I reset the Admin password?
From AudioMuse-AI v1.0.0 the Admin password is stored encrypted on the database. The only way to reset is to access the PostgreSQL database and delete it. You can read more in the AUTHENTICATION documentation. This requires direct database access since the password cannot be reset through the web interface for security reasons.
Can I speed up the analysis process?
The time needed for analysis really depends on your hardware and how big your music collection is. If you want faster analysis, you can disable the text search functionality by setting parameter CLAP_ENABLED to false. This will run the analysis only for the Musicnn model skipping the CLAP model. Alternatives could include running multiple worker containers in parallel - learn more by taking a look at the ARCHITECTURE page and different deployment examples in the deployment folder. GPU analysis is also supported but still experimental - take a look at the GPU DEPLOYMENT page for details.