Self-hosted playlist engine that listens to your waveforms

What it does

AudioMuse-AI is a Dockerized Python stack that bolts onto Jellyfin, Navidrome, LMS, Lyrion, or Emby to build playlists from the raw audio characteristics of your files. Using Librosa and ONNX models, it runs sonic analysis locally to cluster similar tracks, map your library in 2D space, and answer natural-language requests like “calm piano songs.” An optional OpenAI-compatible LLM provider can handle the text parsing, but the audio fingerprinting stays entirely on your hardware.

The interesting bit

The project treats your music collection as a navigable vector space. You can chart a “Song Path” to bridge two tracks through sonically similar intermediates, or use “Song Alchemy” to add and subtract reference tracks like ingredients until the playlist matches your desired vibe. It is essentially a self-hosted recommendation engine that does not need Spotify’s graph or MusicBrainz tags to surface connections.

Key highlights

• Runs on both ARM64 and AMD64, tested from Raspberry Pi 5 up to Intel i5 boxes
• “Sonic Fingerprint” builds playlists based on your recent listening habits
• Lyrics search supports 72 languages, letting you query by theme or story rather than sound
• Ships with dedicated plugins for Jellyfin and Navidrome, plus a Helm chart for Kubernetes

Caveats

• Requires a 4-core CPU with AVX2 (Intel 2015+) or ARM; QEMU virtual CPUs without AVX2 passthrough will prevent the app from starting
• PostgreSQL v15 is explicitly required; other versions are known to cause errors
• Major and minor version bumps can force a full library re-analysis when the embedding models change

Verdict

A solid fit for music hoarders running local media servers who want Spotify-style discovery without the cloud. If you do not self-host your audio files, there is nothing here for you.