Set up VoxCast Server

Version 1.0.0 · Windows, Linux, Docker

VoxCast Server runs the speech recognition on a single machine with a strong GPU. Every VoxCast on your network sends dictation, meetings and file conversions there instead of computing locally. Everything stays on your network; no cloud is involved.

Windows: done in 3 minutes

1. Install. Download the server setup for Windows and run it on the machine with the strong GPU. The setup asks for the network port (default 8590, just keep it), registers the Windows service and opens the Windows Firewall. Both checkboxes are already ticked: just click through.

2. Find the server in the client. On any PC, open VoxCast, go to Settings, section Remote server, enable it and click "Find server". Your server shows up in the list; one click fills in the address.

3. Paste the API key. On the server machine, the Start menu has an entry "VoxCast Server API key". Copy the value (starts with vxc_) into the key field in the client, save, done.

Use "Test connection" in the client to confirm everything works: it shows the active model and whether the server computes on the GPU.

Changing the port

The port is chosen during installation (default 8590). To change it later, edit server.toml in the installation folder (C:\Program Files\VoxCast Server), then restart the "VoxCast Server" service. The client's server search finds the server whatever port you pick.

Linux and Docker

Download the server package (tar.gz) and unpack it.

Docker (recommended):

cd voxcast-server-1.0.0/compute_server/deploy
docker compose up -d
docker compose logs | grep -A2 "API key"

GPU operation needs the NVIDIA driver and the nvidia-container-toolkit on the host. The compose file uses host networking so the server search works.

Native (Debian/Ubuntu, systemd):

sudo compute_server/deploy/install.sh
journalctl -u voxcast-server | grep -A2 "API key"

Firewall

The server needs two openings on the local network: the chosen port (default TCP 8590) for transcription and UDP 8590 for the server search. The Windows installer creates a program-scoped rule for both automatically (profiles "Private" and "Domain"). On Linux with an active firewall: sudo ufw allow 8590 or the equivalent firewalld rule. Corporate firewalls between the devices must allow the port as well.

If the server is not found

Network is set to "Public". Windows blocks server services on public networks. Set the network on the server machine to "Private" (Settings, Network & internet, properties of the active network).

Different subnet or VPN. The search uses broadcasts and only finds servers on the same subnet. Across subnets, VLANs or VPNs, enter the address manually: http://SERVER-IP:8590.

Guest Wi-Fi. Many routers isolate Wi-Fi devices from each other (client isolation). Use the regular network or disable the isolation.

Is the service running? Check the Services console on the server machine for "VoxCast Server", or open the console shortcut in the Start menu: it prints the address and the API key in plain text.

HTTPS (encrypted connection)

For an encrypted LAN connection without running your own certificate authority, enable HTTPS with a self-signed certificate. Set tls_enabled = true in server.toml (or run VoxCastServer.exe --gen-cert to create the certificate up front) and restart. The server prints the certificate fingerprint on start.

On the client, use an https:// address; the first time you connect, VoxCast shows the certificate fingerprint and asks you to trust it (like connecting to a new server over SSH). Confirm it matches the fingerprint the server printed, and it is pinned from then on. For a public address, terminate TLS at a reverse proxy (e.g. nginx) with a real certificate instead.

Central AI enhancement (for companies)

Instead of every user entering their own Anthropic or OpenAI key, store one key on the server. In server.toml set llm_provider (anthropic or openai) and llm_api_key. Clients then see "Use the server" under AI enhancement and need no key of their own. The key never leaves the server.

Company AI modes (system prompts)

Define enhancement modes (system prompts) centrally so everyone shares the same company modes. Manage them with the server tool, on Windows and Linux alike:

VoxCastServer.exe --add-prompt "Formal e-mail" "Rewrite the text as a formal e-mail."
VoxCastServer.exe --list-prompts
VoxCastServer.exe --remove-prompt formal-e-mail

On Linux: python compute_server/server_main.py --add-prompt "..." "...". Clients pick these up automatically.

Rolling out to many PCs (enterprise)

To ship VoxCast pre-configured, place a policy file at %PROGRAMDATA%\VoxCast\managed.json (Windows) or /etc/voxcast/managed.json (Linux) and distribute the normal installer via GPO/Intune/SCCM:

{ "values": { "remote_enabled": true, "remote_url": "https://server:8590", "enhance_source": "server" }, "secrets": { "remote_server_api_key": "vxc_..." }, "locked": ["remote_enabled", "remote_url", "remote_server_api_key", "enhance_source"] }

Fields under locked are pre-filled and cannot be changed by the user. The file may contain the key in clear text, so restrict its permissions to administrators.

Updating the server

On Windows: VoxCastServer.exe --update checks for a newer build, downloads and verifies it, and runs the installer (the service restarts automatically). Use --update --check to only check. On Linux, pull the new server package or Docker image.

Security

VoxCast Server is built for trusted local networks. Without HTTPS the LAN connection is unencrypted; access is protected by the API key. When connecting, verify the shown IP really is your server. For access over the internet, put a TLS reverse proxy (e.g. nginx) in front; never expose the server to the internet unprotected.

Questions? info@hashfox.com