Beltr-NVIDIA ghcr.io/casavargas/beltr:latest-cuda https://github.com/orgs/CasaVargas/packages/container/package/beltr bridge bash false https://github.com/CasaVargas/beltr-releases/issues https://beltr.app Beltr turns any song in your music library into a karaoke track. It separates the vocals out with a local AI model, finds or transcribes synced lyrics, and displays them on a TV while everyone's phone acts as a remote — scan the QR code on screen, no app to install. Everything runs on your server. Nothing about your library is uploaded. THIS IS THE NVIDIA GPU BUILD. It needs the Unraid "Nvidia-Driver" plugin installed and your GPU passed through. Two settings below do that: NVIDIA_VISIBLE_DEVICES must hold your card's UUID (the Nvidia-Driver plugin page in Settings shows it, like GPU-1a2b3c4d-...), and "Extra Parameters" must contain --runtime=nvidia, which this template already sets. If you do not have an NVIDIA GPU, install plain "Beltr" instead. This image is several GB larger and gains you nothing without a card. It will still run — it falls back to the CPU — so a broken passthrough shows up as "slow", not "broken". Check the GPU panel in Settings after install to confirm the card was actually found. FIRST RUN: Beltr downloads about 2 GB of model weights into /cache the first time it needs them. That happens once and survives container updates as long as /cache stays mapped. Watch the container log if a first song seems to sit still — the download is what it's doing. SPEED: with a working GPU, separating a 3-4 minute song takes well under a minute. Most songs then get their lyrics instantly from an online lyrics database; the ones no provider knows fall back to local transcription, which the GPU also accelerates. LICENSE: Beltr is a one-time purchase (no subscription). Run it free for five songs, then enter your key from beltr.app in the License key field below — or drop it in a file at /config/license.key. It activates once and then works offline. Container updates and restarts do not re-activate, as long as you keep the Config path mapped; a key covers two installs. AFTER INSTALL: open the WebUI, go to Settings, and add /media as a music folder. Then open http://your-server:8477/tv on whatever screen you're singing at, and scan the QR with a phone. MICROPHONES NEED HTTPS: browsers only allow microphone access on a secure page, so singing scores and the phone-as-mic feature need Beltr behind a reverse proxy with a certificate. Everything else — queueing, playback, lyrics, stem mixing — works normally over plain http. NOTE: Beltr assumes everyone who can reach it is trusted. Set AUTH_PASSWORD to put a password on the TV and dashboard screens, but do not expose this container to the internet without a reverse proxy doing real authentication in front of it. MediaApp:Music MediaServer:Music http://[IP]:[PORT:8477]/dashboard https://raw.githubusercontent.com/CasaVargas/beltr-unraid/main/templates/beltr-cuda.xml https://raw.githubusercontent.com/CasaVargas/beltr-unraid/main/icon.png https://raw.githubusercontent.com/CasaVargas/beltr-unraid/main/README.md karaoke sing singalong lyrics vocals stems demucs whisper party microphone cdg mp3g nvidia gpu cuda Proprietary — one-time purchase, five-song free trial https://raw.githubusercontent.com/CasaVargas/beltr-unraid/main/screenshots/tv-lyrics.webp https://raw.githubusercontent.com/CasaVargas/beltr-unraid/main/screenshots/dashboard.png https://raw.githubusercontent.com/CasaVargas/beltr-unraid/main/screenshots/phone-remote.webp 2026-08-14 ### 2026-08-14 First Community Applications release. The container tracks the Beltr desktop release line — per-version notes are at https://github.com/CasaVargas/beltr-releases/releases --runtime=nvidia Unraid Nvidia-Driver plugin, and an NVIDIA GPU with driver 525 or newer (the image ships CUDA 12.8 builds of PyTorch). Internet access on first start to activate a license key, or to begin the five-song trial. 8477 /mnt/user/appdata/beltr /mnt/user/beltr /mnt/user/Music /mnt/user/appdata/beltr-cache all compute,utility 99 100 Etc/UTC large-v3 htdemucs first-run