Beltr for Unraid

Unraid Community Applications templates for Beltr, a self-hosted karaoke server. It takes a song from your own music library, separates the vocals out with a local AI model (UVR MDX-Net on ONNX Runtime), finds synced lyrics online or drafts them locally, force-aligns them to the vocal word by word, and puts the words on a TV while everyone's phone acts as a remote. Guests scan the QR code on screen; there is no app for them to install.

Everything runs on your server. Nothing about your library is uploaded.

beltr.app · Full Unraid guide · Reverse proxy and HTTPS · Support

Which template

Install one of the three. They share the same default paths.

Both GPU images fall back to the CPU if passthrough is misconfigured, so a broken setup looks slow rather than broken. Confirm the GPU was found in Settings after install.

Install

Community Applications: Apps, search Beltr, pick one of the three, set the four paths, Apply. Then open the WebUI, go to Settings, and add /media as a music folder. Beltr never scans your library until you point it at one.

Plain Docker:

docker run -d --name beltr \
  -p 8477:8477 \
  -e PUID=99 -e PGID=100 -e TZ=America/New_York \
  -v /mnt/user/appdata/beltr:/config \
  -v /mnt/user/beltr:/library \
  -v /mnt/user/appdata/beltr-cache:/cache \
  -v /mnt/user/Music:/media:ro \
  --restart unless-stopped \
  ghcr.io/casavargas/beltr:latest

Compose: docker-compose.yml with .env.example as a starting point. Pick one profile:

cp .env.example .env        # edit PUID/PGID and the paths
docker compose --profile cpu up -d      # or --profile gpu (NVIDIA), --profile openvino (Intel)

The four paths

Two of these matter more than they look. /library must not live in appdata, because every processed song leaves a pair of lossless stems behind and a few hundred songs will fill a cache pool sized for config files. And /cache must stay mapped, or the ~2 GB of model weights are re-downloaded every time you update the image.

Worth knowing before you install

• First run downloads about 2 GB of model weights into /cache. Once. If a first song seems to sit still, that is what it is doing: docker logs -f beltr.
• Microphones need HTTPS. Browsers only expose the microphone on a secure page, so scoring and phone-as-mic need Beltr behind a reverse proxy with a certificate. Queueing, playback, lyrics and stem mixing all work fine over plain http. Details.
• Beltr assumes everyone who can reach it is trusted. Room joins are open by design so a guest can scan a QR code and sing. AUTH_PASSWORD puts a password on the TV and dashboard screens, but do not port-forward this container.
• AMD GPUs are not supported in the container. ONNX Runtime's AMD path needs a full ROCm userspace in the image, several gigabytes, so it is not built; on the Windows desktop app AMD and Intel graphics can use the optional DirectML GPU pack instead. Intel iGPUs and Arc A-series cards are the Beltr-Intel image above; Arc B-series and Panther Lake are newer than its driver and are not supported yet. Everything also runs on the CPU alone.

Licensing

Beltr is a one-time purchase ($19.99, lifetime updates, 14-day money-back guarantee), not a subscription. The container runs free for five songs so you can try it against your own music, then asks for a key from beltr.app. Activation happens once and then works offline; updating or recreating the container does not re-activate, as long as you keep /config. A key covers two installs.

The MIT license in this repository covers these templates and docs only. Beltr itself is proprietary software distributed under its own EULA.

Keeping up

Templates here are copied from the Beltr application repository, which is where they are edited. File issues and support requests at CasaVargas/beltr-releases.