Show HN: AudioMuse-AI Sonic Analysis (github.com)

AudioMuse-AI is an open-source, Dockerized toolkit for adding AI-driven playlist generation to self-hosted media servers (Jellyfin, Navidrome and any Subsonic-compatible server). Packaged for Docker Compose, Podman and Kubernetes (tested on K3S with a Helm chart), it extracts sonic features with Librosa, builds embeddings with TensorFlow and AI models, and exposes a UI, chat-driven “instant playlist” and API for analysis, clustering and similarity tasks. The project is beta, supports amd64 and arm64, can use local models (Ollama) or cloud models (Gemini), and offers optional experimental cloud Collection Sync via OAuth/GitHub (opt-in, privacy policy required). Technically notable updates include a Voyager index that boosts recall from ~70–80% to ~99% for top-100 similar-song queries while reducing memory use, a Sonic Fingerprint that converts listening history into a fingerprint for similarity discovery, and a Song Path feature to compute a sonic path between two tracks. Clustering uses an evolutionary search over K (defaults: kmeans, NUM_CLUSTERS_MIN 40, MAX 100, CLUSTERING_RUNS 5000) with tunable scoring weights (diversity 2.0, purity 1.0) and options for GMM/DBSCAN. For operators, the repo supplies deployment examples, hardware notes, GPU/Nvidia worker support and deep dives on analysis, clustering and concurrency—making it a practical sandbox for researchers and tinkerers wanting to experiment with large-scale, self-hosted audio similarity and playlist generation.