LymeScribe Server guide

LymeScribe Server is an always-on transcription server for a whole office, running on a machine you own. This guide covers what it is, what it needs, and how to install, configure, and run it day to day.

Last updated 2026-08-31

What LymeScribe Server is

It's a small .NET server you install on one machine in your building. It wraps WhisperX — the speech model that does the actual transcription — behind a network service, and every LymeScribe desktop app on the network sends its audio there instead of transcribing locally. The audio never leaves your network.

Around that core it provides a job queue, a browser admin dashboard on the same port, server-side history with retention rules, and per-device API keys you can issue and revoke individually.

Who it's for — and when you don't need it

There are two ways to host transcription with LymeScribe, and most people should start with the cheaper one.

Pro self-hosting (in the desktop app)

  • A paid unlock inside the free Mac, Windows, or Linux app
  • Your own computer does the transcribing
  • Roughly three attached peers — a connection cap, not a speed one
  • No separate install, no ML environment to set up

Right for one person, or a couple of people sharing one good machine.

LymeScribe Server (this guide)

  • A separate product, installed on a dedicated always-on machine
  • No peer cap — the whole office connects
  • GPU-accelerated on Windows with an NVIDIA card
  • Admin dashboard, API keys, server-side history and retention

Right for an office where several people transcribe through the day and nobody's laptop should be the bottleneck.

Both tiers queue jobs. Whichever you run, jobs are processed one at a time rather than in parallel — the difference is how many people can attach and how fast each job finishes. See Choosing a tier if you're still deciding, or the pricing section.

Requirements at a glance

WhatDetails
A machine to dedicate A Windows PC with an NVIDIA GPU (the real-time option), or an Apple Silicon Mac (M1 or later, 16 GB+ RAM). Windows without a GPU works too, just slower.
ML tooling underneath WSL2 + CUDA/cuDNN + a WhisperX Python environment on Windows; Homebrew + a WhisperX Python venv on Mac. This is the part that takes real time — budget 30–60 minutes the first time.
A HuggingFace account Free, but required: the speaker-label models are gated and need a one-time license acceptance plus an access token.
Network One TCP port on your LAN — 58994 by default. Clients connect by the server's LAN address.
Trial 14 days free, no credit card. The clock starts on your first transcription, not at install, so setup time doesn't burn it.

Hardware sizing is driven almost entirely by the GPU: the transcription model needs roughly 3 GB of VRAM, and speaker labeling adds another 1–2 GB on top, so a card with 6 GB minimum and 8–12 GB recommended is the practical floor. 16 GB of system RAM and an SSD round it out.

The guide

Also worth reading

Questions before you commit? Email webinquiry@iadev.net — a real person answers.