Frequently asked questions

Not to use it. The desktop app is free forever as a client — it connects to any LymeScribe server on your network and gives you the full dictation experience, hotkeys, history and corrections, with no account and no trial clock.

You pay only when you want your own computer to do the transcribing. That's the Host unlock: a one-time purchase, currently $39 at the early-adopter price (regularly $59). Which tier do I need? walks through the choice.

The Host unlock turns the Mac, PC or Linux box you already own into the transcription server for yourself plus a few devices on your network — roughly three attached peers, with jobs running one at a time in a queue rather than in parallel.

LymeScribe Server is a separate, office-scale product: a dedicated always-on machine (Windows with an NVIDIA GPU, or Apple Silicon) with no peer cap. It's currently a Technology Preview — the pricing page says "let's talk" rather than listing a price. See choosing a tier and the Server guide.

Client mode never involves a trial — it's simply free. The first time you turn on hosting, a 7-day trial with full features starts, with no credit card. When it ends, hosting locks until you buy; client mode keeps working exactly as before. LymeScribe Server has its own 14-day trial that starts on its first transcription.

It was emailed to you after purchase and looks like LS-XXXX-XXXX-XXXX-XXXX. Can't find it, or think a key was revoked by mistake? Email us with your order number or the address you purchased with and we'll re-send it.

Your key activates on up to 10 machines you own. If you hit the limit, deactivate a machine you no longer use first — in the app, Settings ▸ License. If you genuinely need more, email us with your key or purchase address; we're happy to raise the limit, usually the same day.

No. The hosting unlock is a one-time purchase for the current major version, and it keeps working — the software runs on your hardware, so there's no server bill for us to pass along. Point updates are free; a future major version can be upgraded to at a discount.

Every purchase carries a 30-day money-back guarantee. Email webinquiry@iadev.net and we'll refund it, no interrogation. A refund revokes the associated license and hosting locks again — client mode keeps working, free. Full terms in the refund policy.

No. There's no sign-up, no profile, and nothing to log into. Purchases are handled by Paddle.com as merchant of record and give you a license key, not an account.

No. Transcription runs entirely on hardware you own — the speech model, the audio and the resulting text all stay there. When you host, that machine is your own computer. When you connect to a LymeScribe Server, the audio goes to that server over your own network and no further. Nothing is uploaded to us, and dictation works with the network switched off entirely.

The only exception is one you make deliberately: if you point an AI tool at a cloud provider, that transcript text goes to the provider you chose. See the next answer.

Those run against whichever OpenAI-compatible endpoint you configure as an engine profile. Point one at a local model server (LM Studio, for example) and the transcript never leaves your machine; point one at a cloud provider and the transcript text you run the action on is sent there under that provider's terms. Nothing is sent unless you run an action, and nothing is sent to us either way. Setup is in AI tools setup, and the actions themselves in AI tools.

• Update checks against our release manifest.
• License activation, for paid hosting, against our licensing server.
• Engine downloads, when your machine qualifies for a GPU build that isn't bundled.
• Your own server, when you're using the app as a client on your LAN.

None of them carry your audio or your transcripts.

On your own disk, in the app's history. Every destructive and management action lives in the History window — delete a single item, clear the history, export, reveal the folder, and (when hosting) delete kept audio. Audio retention is a setting, so you can have recordings discarded automatically. See History and Settings.

We never see your audio or transcripts — that's the point of the product. The app has no analytics and no telemetry. This website uses Cloudflare Web Analytics, a cookieless service reporting aggregate traffic only. When you buy a license we store what's needed to issue and support it: name, email, license key, and basic activation records. Payment details go to Paddle, not to us. Details in the privacy policy.

• Mac — macOS 14 or later, notarized by Apple, auto-updates via Sparkle.
• Windows — Windows 10/11, 64-bit, signed installer, auto-updates in-app.
• Linux — x86_64 AppImage (also published as the AUR package lymescribe-client-bin). Still labelled Beta.

All three are the same product: the same free client, the same paid hosting unlock, the same "one desk, one list" layout. Getting started covers installing on each.

Equivalent, but honestly not identical — the Linux desktop takes some capabilities away from applications on purpose, and we'd rather name them than pretend. The short version:

• Typing into other apps on Wayland needs your permission. Wayland has no way for an app to type on its own; LymeScribe asks the desktop's remote-desktop portal, and until you grant it, dictated text goes to your clipboard instead, with one notification per session explaining why. On X11 typing just works, no dialog.
• Some Wayland desktops can't offer that permission at all — sway, river and other wlroots compositors. There, clipboard delivery is the permanent behaviour, and Settings says so rather than showing a button that couldn't help.
• Push-to-talk needs GNOME 48+, KDE Plasma 5.27+ or Hyprland. On Wayland the desktop owns the shortcut, not the app, so there's no chord picker in Settings. On older desktops the fallback is toggling recording from the tray or the window.
• GNOME needs a tray extension ("AppIndicator and KStatusNotifierItem Support") for the tray icon to appear at all. KDE hosts it natively.
• The dictation overlay is best-effort on Wayland — the compositor decides where a window goes, and clicks on the overlay are absorbed rather than passed through.
• Per-app delivery overrides only apply on X11, because Wayland deliberately hides which app has focus.
• No keyring? Secrets fall back to a permission-restricted file, and the app says so instead of hiding it.

The app keeps a live ledger of exactly which of these apply to your machine — check Settings and the Activity log.

It means the guarantee is weaker than on Mac or Windows, and we'd rather you know than assume. Mac updates are Ed25519-signed and notarized by Apple; Windows installers are signed by Interapp Development, Inc. There is no equivalent for the AppImage. What the Linux updater does check: the file is fetched over HTTPS and its SHA-256 is verified against the release manifest before it runs. That protects against a corrupted or tampered download, but there's no third party attesting to who published it. If you installed from the AUR, your package manager's own verification applies instead and the app tells you to update through pacman.

More than plain transcription. On Mac, Apple Silicon M2 with 16 GB is the floor, and an M-series Pro/Max/Ultra with 32 GB is what we'd recommend. On Windows, speaker labels require an NVIDIA GPU with 8 GB of VRAM — we don't offer them on CPU, where they're too slow to stand behind. On Linux the feature also needs two tools the app doesn't bundle: ffmpeg and python3 (3.11 or newer, used by the setup wizard). Plain transcripts, without labels, work on far less. Full floors are in the system-requirements section of the home page; tuning advice is in tuning speaker labels.

On Mac, Apple Silicon uses Metal; Intel Macs get a CPU-only hosting beta. On Windows and Linux, an NVIDIA GPU uses CUDA and AMD/Intel GPUs use Vulkan, with CPU as the fallback everywhere. CUDA is preferred over Vulkan when both are available, and a machine with no real GPU is never offered a Vulkan build. There is no Vulkan tier on the Mac. Details in self-hosting.

On Windows 11 with WSL2 and an NVIDIA GPU, or on Apple Silicon. Installation is in the Server installation guide; the prerequisites are documented separately for Windows/WSL/NVIDIA and macOS Apple Silicon, plus HuggingFace setup for speaker labels.

Our download host sits behind a CDN that caches the update manifest for a few hours, so a brand-new release can take a little while to reach every machine. Waiting solves it; so does downloading the installer directly from the download section. On Mac you can force a check from the menu bar ("Check for Updates…"); on Windows and Linux the check is in the app. LymeScribe Server has no auto-update — run the new installer over it.

Windows SmartScreen can warn on installers it hasn't seen many times yet, even signed ones. How to open past the security warning shows the exact clicks.

That's the GPU engine build for your machine, fetched on demand rather than bloating the installer for people who don't need it. The CUDA build is the big one (several hundred megabytes); the Vulkan build is far smaller. The transcription models themselves also download once — budget around 5 GB of free disk, 10 GB if you use speaker labels. Progress shows in the status strip and the Activity log. See self-hosting.

• NVIDIA card present but ignored: the app asks the driver via nvidia-smi. If the driver is installed but not answering, hosting falls back to CPU — a reboot usually fixes it, and the app raises a notice when it sees exactly this.
• A "GPU" that's really software: software Vulkan devices (llvmpipe, lavapipe, SwiftShader and friends) are deliberately ignored, because they're slower than the CPU build.
• Old CPU: the accelerated builds still need AVX2 on the CPU side, whichever GPU you have.
• Missing meters aren't a fault: the GPU-percentage meter only shows on NVIDIA/CUDA hosts, and the video-memory bar is unavailable on Windows with an AMD card. The work is still happening on the GPU.

Almost always the Wayland typing permission. Open Settings ▸ Dictation: if there's an "Allow LymeScribe to type" button, press it and accept your desktop's own dialog. If Settings says your desktop doesn't offer the permission (sway, river, other wlroots compositors), clipboard delivery is the honest ceiling there — an X11 session gives typing with no setup. LymeScribe deliberately does not use ydotool, which would need privileges over your whole machine's input.

Open the Activity log from the status strip — it's the app's running account of what it's doing. On Linux there's also a session-scoped diagnostics switch in Settings ▸ General ▸ Diagnostics, which writes one self-contained file (no transcripts, no audio, no license key) and is always off again on the next launch. To capture start-up itself, launch with the flag:

LymeScribeClient-x.y.z-x86_64.AppImage --diagnostics

Attach the file to an email and we'll take a look.

Start at the support hub — it indexes the app and server guides — or just email us. Include your platform, the app version, and what you were doing when it went wrong.