Windows
Every Scribe0 for Windows release, in plain language.
Get it from Microsoft Storev0.3.3
Added- See which graphics card Scribe0 checked, and get a heads-up before the big download. The on-device drafting settings now show your graphics card by name and memory at every step, including before you download anything and after a test. And before the 2.4 GB model download, Scribe0 reads your card's memory and warns you if it looks too small to run the model, so you are not left downloading gigabytes only to be told afterwards that it will not run. The warning never blocks you: you can always download and test anyway, and integrated graphics (which share system memory) are still given the benefit of the doubt and tested for real.
v0.3.2
Added- Automatic medication safety check after enhancing. Scribe0 now compares the medication names in the drafted note against what you actually dictated, on this device, and flags any drug that looks invented or dropped. It catches the rare case where the on-device drafting model smooths a garbled word into a real, plausible, but wrong medication name that would otherwise read as correct. The check runs with no added delay and never leaves your PC. It is on by default; you can turn it off under Settings, Connection (Quality check) if you need to.
Changed- Simplified the dictation toolbar. Live transcription is now always on, so its separate toggle button is gone. The markdown view toggle, the unfinished Send button, and the Save button were also removed to reduce clutter. Saving a transcript to a file still works with Ctrl+S, and notes still display as formatted text (double-click a note to edit the raw text).
Fixed- The Enhance button no longer shows a blank box while it works. During a note generation the button now correctly shows a stop icon you can click to cancel, instead of an empty square.
v0.3.1
Added- Context cart: give the AI reference text alongside your dictation. A new basket button on the toolbar holds up to 5 pasted snippets (8,000 characters) of reference material, such as prior results or a referral question, and the enhance step reads them next to your transcript. The cart empties itself when you start the next recording and is never saved to disk, so one patient's context cannot leak into the next. Works the same way as on the Mac and the web app.
- Prompt gallery. Settings, Prompt presets now has a Browse gallery button that loads community prompts from scribe0.com, searchable by title, specialty, tags, author and content, with a preview pane. When a gallery prompt you imported is revised, an Update button appears; updating keeps your own star, ordering and backend settings. Imported prompts always run on the backend you chose: the gallery can never change where your transcript is sent. The last downloaded gallery still opens when you are offline.
- Pin a prompt to a backend. Each prompt can now name the backend it runs on (Bedrock, OpenRouter, or the on-device model), overriding the global Connection setting for just that prompt, with an optional model choice within that backend. Pinned prompts show their backend in the prompt picker. If a prompt is pinned to the on-device model on a machine that cannot run it, enhance says so plainly instead of quietly using the cloud.
- Enhance now tells you how long it will take. After a few runs with the same prompt and backend, the status bar shows a real progress bar with a countdown instead of just a character count, learned from your own history on your own machine and stored only there. Before it has enough data it shows an honest indeterminate bar rather than a made-up one.
- The status line shows what enhance is actually doing. A small animated icon next to the status text distinguishes sending, receiving, and on-device processing at a glance, matching the Mac.
- Stop the speech model download from the main window. The first run downloads about 630 MB, and until now the only way to stop it was a button buried in Settings. There is now a progress bar and a Stop button next to the status line, where the download is actually happening. Stopping it no longer leaves the half downloaded file on your disk.
- Stop the graphics card test. The test that decides whether your card can run the on-device model starts on its own when you pick the on-device option, and can sit there for up to three minutes on a card that struggles. It can now be stopped.
Changed- Downloaded speech and drafting models moved to a new folder. Model files now live in a Scribe0 folder inside your user profile instead of hidden app data. Your already downloaded models are moved across automatically on first launch, nothing re-downloads. This also means uninstalling the app no longer risks taking gigabytes of model files with it, and it prepares the app for the upcoming Microsoft Store edition, where the old location would have been wiped on uninstall.
- The on-device drafting model now downloads from scribe0.com first. The 2.4 GB Qwen3 weights come from our own mirror, which is faster and more reliable than HuggingFace; HuggingFace remains as an automatic fallback if the mirror is unreachable.
- A per-prompt model choice now requires a backend choice. Previously a prompt could name a model without naming a backend, which could silently send the id to the wrong service. Existing prompts that pinned a known model have been assigned the matching backend automatically; a custom model id we could not attribute keeps working on the global backend.
- New toolbar icons, matching the Mac. Scribe0 for Windows was drawing its icons from the Windows system set, which meant the two apps did not look like the same product side by side. The toolbar now uses the same icons, in the same style, and the ones that mean the same thing on both platforms now look the same: a microphone that becomes a stop square, a sparkle for enhance, a clipboard for copy, a paper plane for send.
Fixed- The Stop button during enhance was greyed out just when you needed it. Enhance turns into a Stop button, but it went dim the moment the note started being drafted and only came back when the first words arrived. On a local model that is the whole model loading time, which is exactly the stretch you would want to stop, and if the model never produced anything you could never stop it at all.
- A slow connection was blamed on you. If drafting or a model download ran past its time limit, Scribe0 said you had cancelled it: a quiet, ordinary looking message for something that had actually failed and was worth retrying. Whether you stopped the work or it broke is now told apart properly, so a timeout reads as a timeout and offers you a retry.
- A stopped download no longer costs you disk space. Stopping a model download used to leave the partial file behind, up to 2.4 GB of it for the on-device drafting model, and nothing in the app ever mentioned it again. It was only reclaimed if you happened to start the same download a second time.
v0.3.0
Added- On-device AI enhance. A new "On-device (no network)" option in Settings > Connection runs Qwen3 4B on your own computer to draft the note. Nothing leaves the machine, there is no daily limit, and no key is needed. It downloads about 2.4 GB the first time you turn it on, and then drafts a note in a couple of seconds on a capable graphics card. It needs a graphics card that supports Vulkan; Scribe0 tests yours and tells you if it cannot run the model, in which case you stay on a cloud route.
- Stop button for enhance. The enhance button turns into a stop button while a note is being drafted. Stopping restores your transcript exactly as it was, rather than leaving a half written note behind.
- A graphics compatibility check. Before on-device enhance is switched on, Scribe0 runs a short test that loads the model on your graphics card and drafts a few words. If your card can run it, on-device enhance is enabled and Settings tells you which card is being used and how fast it is. If it cannot, the option stays off and Scribe0 says so rather than being slow or unstable. The test runs in a separate background process, so a faulty graphics driver cannot take Scribe0 down with it.
Changed- The built-in "Clinical polish" prompt now spells out what it should produce: the polished transcript first, then the Plan. It also gained worked examples for units and a rule to keep your own voice ("I reviewed", not "the patient was reviewed"). The old wording only ever described the Plan format, and left the rest to be inferred. Cloud models inferred it; smaller on-device models did not, and returned the Plan alone. Your copy is updated automatically unless you have edited it, in which case it is left alone.
- The transcript is now sent to the model on its own, instead of being wrapped in "Please enhance the following transcript. Return only the enhanced text." That wrapper contradicted any prompt asking for more than one section, so the Plan could be dropped.
Fixed- Saving the enhancement provider. The Save button in Settings > Connection was hidden whenever a provider with no API key was selected, so the on-device option could be picked but never saved and quietly reverted. Save is now always visible and confirms when it has saved.
- Fidelity rules on every enhance. Language models can quietly replace a dictated drug dose with the dose they expect to see, so a dictated aspirin 300 mg comes back as 325 mg. Every enhance now carries rules that forbid changing any dictated number, dose, unit, route or date, and forbid stating a result that was never dictated. The rules are applied by the app itself rather than by the prompt text, so they also cover prompts you write yourself. Rewriting a spoken number into digits and short units ("one thousand two hundred and forty picograms per mil" becomes "1240 pg/mL") is still allowed, because that changes the notation and not the value. This is a safeguard, not a fix for a bug you would have seen: the current cloud models pass this check on their own. It matters because the rules travel with the app rather than with any one model.
v0.1.5
Added- Free and Pro plans for the AI enhance step, matching macOS. The enhance step works with no setup on the free tier; a Pro key unlocks the larger models and removes the daily limit.
v0.1.4
Added- Microphone selector. Capture now defaults to the Windows default device, and you can pick a different one.
v0.1.3
Added- Automatic updates. Scribe0 now installs via a proper Windows installer and updates itself in the background using small delta downloads.
- Enhanced notes now render as formatted text rather than raw markdown.
- The AI enhance response streams into the transcript pane as it is generated, so you see it arrive instead of waiting for the whole thing.
- Session diagnostics panel, with realtime factor, CPU, memory and latency, saved alongside each transcript.
- The speech model now adapts to your machine: it scales threads to the hardware and offers a faster, smaller model tier on modest PCs.
- Transcript version chips can now be closed, matching macOS.
Fixed- Long enhancements no longer fail with a dropped connection. The response is now streamed.
- Importing a long audio file no longer crashes. Audio is now split at natural silence boundaries before decoding.
v0.1.2
First public release for Windows. Local speech to text on your own PC, with no cloud
transcription and no account.
Added- Local speech to text using Parakeet, running entirely on your PC.
- Live transcription as you speak, with an option for near realtime output.
- Import an audio file and transcribe it, with speaker labels for multi speaker recordings.
- AI Enhance via the sovereign AWS Bedrock (Australia) backend by default, with OpenRouter as a fallback.
- Editable prompts, each able to pin its own model.
- Snippets and text replacements.
- A live microphone level meter, keyboard shortcuts for the core actions, and a Logs view in Settings.
Fixed- A bug that repeated the same sentence during live transcription.
- A crash when diarizing recordings longer than about 16 seconds.
- A crash caused by voice activity detection overrunning its buffer.
- Settings panels now scroll, so controls no longer clip on small windows.