Familiar
AI-powered local music player that understands sound, not just metadata
Familiar is a self-hosted music player that uses AI to understand what music actually sounds like, not just what it’s tagged as. It builds a semantic map of every track in a library using CLAP (Contrastive Language-Audio Pretraining) embeddings, enabling natural language search — asking for “something that sounds like driving at night” returns results based on sonic similarity, not genre metadata.
The player includes three spatial visualizations: a Mood Grid that plots tracks on energy and valence axes, a Music Map that projects CLAP embeddings into 2D clusters using UMAP, and a 3D Explorer for navigating the library in WebGL. A chat interface exposes 33 AI-powered tools for playlist generation, music discovery, and library analysis.
The backend is Python with FastAPI and SQLite, using librosa for audio feature extraction and the CLAP model for embeddings. The frontend is a React + TypeScript progressive web app. Spotify playlists can be imported, Last.fm scrobbling is built in, and a Subsonic-compatible API enables playback from mobile music apps and CarPlay. Everything runs in Docker.