Your own stem separation server

Songverse splits recordings into stems (see Separating a recording into stems) on a Songverse Sonic server: Demucs, which does the separating, and an analysis of each recording - its tempo, beats and bars, key and sections. A team or an organisation can run its own, on its own computer or server, and connect it. Its songs are then split there, by its members, without needing a role from this Songverse’s admin.

Before you start

• This Songverse allows it. Its admin turns on Teams and organisations may connect their own Songverse Sonic server in Admin > Stem separation. Until then, the Stem separation section doesn’t show on a team’s or an organisation’s page.
• A computer that stays on: Linux with Docker, or a Proxmox host. The processor is enough (a few minutes a song); an NVIDIA graphics card is much faster. Count 8 GB of memory and 10 GB of disk for the models and the jobs.
• Songverse can reach it: a public address starting with https://, usually behind a reverse proxy. A private or local address (http://192.168.1.20:8000) only works when this Songverse’s admin allows it (On private or local addresses too), which makes sense only when both servers are on the same network.

With Docker

docker run -d --name sonic --restart unless-stopped -p 8000:8000 \
  -e DEMUX_API_KEY="$(openssl rand -hex 32)" \
  -e DEMUX_PRELOAD_MODELS="htdemucs htdemucs_ft" \
  -v sonic-data:/var/lib/demucs-api \
  ghcr.io/bricedupuy/songverse-sonic:latest

• Its API key, to paste into Songverse: docker exec sonic printenv DEMUX_API_KEY.
• With an NVIDIA graphics card: the ghcr.io/bricedupuy/songverse-sonic:cuda image, with --gpus all (the NVIDIA container toolkit must be installed). Without a usable card, it says so in its log and uses the processor.
• Docker Compose: the repository has a docker-compose.yml and a .env.example to copy.
• Updating: docker pull the image and create the container again; the jobs and models stay in the sonic-data volume.

Every setting is in the Songverse Sonic repository.

On Proxmox

On the Proxmox host, as root:

bash -c "$(curl -fsSL https://raw.githubusercontent.com/bricedupuy/songverse-sonic/main/ct/demucs-api.sh)"

It creates a Debian container running Songverse Sonic, with a guided choice of processor or graphics card. Its API key: pct exec <CTID> -- cat /etc/demucs-api/api-key. Inside the container, update updates it.

Making it reachable

Put it behind a reverse proxy that gives it an https:// address. With Caddy, for example:

sonic.example.org {
  reverse_proxy localhost:8000
}

Then tell it its public address and to trust the proxy: DEMUX_PUBLIC_BASE_URL=https://sonic.example.org, DEMUX_TRUST_PROXY_HEADERS=true and DEMUX_FORWARDED_ALLOW_IPS set to the proxy’s address. Sonic tells Songverse when stems are ready; if it can’t reach Songverse, Songverse asks it every 2 minutes instead.

Connecting it

On the team’s page - or the organisation’s - its admins find Stem separation:

• Server address and API key, then Connect;
• Then a finer pass: a slower, better separation that replaces the first stems when the server has time;
• Separations per person in 30 days: empty for no limit.

Test connection checks the address and the key: Connected: songverse-sonic 1.6.0. Models: …. If its version doesn’t find beats and bars, it says to update it. The key is never shown again: leave it blank to keep it. Disconnect stops using the server: anything still being split there fails, and the stems already split stay.

Which server a song uses

• A team’s song: the team’s own server, else its organisation’s, else this Songverse’s own, if it has one.
• Your own songs: this Songverse’s own server.
• Who can separate: on a team’s or an organisation’s server, the team’s members, within its limit. On this Songverse’s own server, whoever its admin allowed.

When you separate a recording, the window names the server it goes to: On Grace Church’s Sonic server.