SparkFlow: hold a key, talk, it types
Posted by Chris
SparkFlow: hold a key, talk, it types SparkFlow is private voice-to-text for your computer: hold a key, speak, release — and clean text appears in whatever app you're typing into. Transcription runs entirely on your machine, so your voice never touches a cloud, and it keeps working with the Wi-Fi off. It's $5/month or $49 lifetime at tomsparkflow.com, lifetime is included in the SparkBox Ultimate bundle, and every install starts with a 7-day full trial — no card, no account. This guide covers install, first-run setup, the hotkeys, and activating your license. 1. Install — Windows 10 or 11 SparkFlow is Windows-first right now (Mac and Linux are coming; it's an active beta that updates itself). Download the installer from tomsparkflow.com and run it. You'll likely see a blue "Windows protected your PC" screen. That's SmartScreen reacting to the beta installer not being code-signed yet — not a virus warning. Click More info → Run anyway. (Signing is on the roadmap; the honest FAQ on the site says the same.) 2. First run: microphone + pick a model SparkFlow walks you through two things on first launch: 1. Microphone permission. Windows will ask, or SparkFlow will point you to the right privacy setting if it's blocked. 2. Pick a transcription model. This is the one download the app needs — the "brain" that turns your voice into text, stored locally. The recommended pick is around half a gigabyte and is the sweet spot of speed and accuracy for most machines. Smaller models (~50 MB) run on anything; bigger ones are more accurate if your machine has the muscle. A graphics card helps speed but isn't required — without one, the CPU does the work. 3. Using it — the hotkeys SparkFlow lives in your system tray and works everywhere. The defaults: Hold Ctrl+Space and talk, release to stop — your words appear in whatever app has focus: editor, browser, Slack, a terminal. Escape — cancel a recording you didn't mean to start. Ctrl+Shift+Space — dictate with AI cleanup on top (optional — see section 6). Ctrl+Alt+R — Command Mode: select text anywhere and tell SparkFlow how to rewrite it (also part of the optional AI layer). All of these are changeable in Settings → General → SparkFlow Shortcuts — click a binding and press your preferred keys. Filler-word removal ("um", "uh") and repeated-word cleanup happen automatically, on your machine, no AI required. 4. Activate your license Your license key arrives by email from licenses@tomsparkflow.com after purchase — including if SparkFlow came in your Ultimate bundle. The key starts with SPARKFLOW-. 1. Open SparkFlow from the tray. 2. Go to Settings → About — the License card is there. 3. Paste the key from your email and click Activate. The status line flips to "Licensed to your email" and that's it. Activation is checked on your machine — SparkFlow doesn't phone home. Until you activate, the 7-day trial is the full product: nothing is feature-gated, so you can put off the key until the trial nags you. Ultimate buyers: your SparkFlow email is separate from your SparkBox license email. If it hasn't arrived within a few minutes of the others (spam checked), message @tomspark on Demox with your checkout email — he can verify the purchase and issue your key directly. 5. What "private" means here, concretely The transcription model runs on your computer. Your audio is processed locally and nothing about your dictation is uploaded — you can turn off Wi-Fi mid-sentence and keep dictating. Downloads, update checks and optional integrations can use the network; ordinary local transcription does not upload your audio. The one exception is opt-in: the AI-cleanup layer (next section) talks to whatever AI you point it at — which can also be fully local. 6. Good-to-know settings AI cleanup (optional). The Ctrl+Shift+Space layer and Command Mode need an AI model to rewrite text: bring your own cloud API key, or point it at a local Ollama / llama.cpp server to keep even this piece on your machine. Out of the box, plain dictation with the deterministic cleanup is the whole experience — you don't need to configure any of this. Custom vocabulary. Add names, jargon, or product words SparkFlow should always get right. History. A local, searchable list of your past dictations — yours to keep or clear, with a retention setting. Languages. Dozens, depending on the model you loaded — plus a "translate to English" toggle. Paste behavior. By default SparkFlow pastes text in and restores your clipboard after. A few locked-down apps block that; SparkFlow falls back to simulated typing automatically, and you can force a method in Settings → Advanced. One honest limit: apps running as administrator won't accept injected text from a normal-privileges SparkFlow — you'll get the text on your clipboard instead. 7. Model cheat sheet You can switch models any time in Settings — the first pick isn't a commitment. The names below are exactly what the in-app picker shows; sizes are approximate: | Model | Download | When to pick it | | --- | --- | --- | | Moonshine V2 Tiny | ~31 MB | Instant on any machine; fine for quick notes (English-focused) | | Moonshine Base | ~55 MB | A notch better, still featherweight (English-focused) | | Moonshine V2 Small | ~99 MB | Good balance for older machines (English-focused) | | Parakeet V3 (recommended) | ~456 MB | The sweet spot — fast on plain CPUs, strong accuracy | | Whisper Small | ~465 MB | The multilingual all-rounder | | Whisper Large | ~1 GB | Maximum accuracy; happiest on a GPU | | Whisper Turbo | ~1.5 GB | Top-tier accuracy tuned for speed — wants a GPU and disk space | Rule of thumb: start with the recommended pick. If dictation feels slow, step down a size; if it's fast and you want fewer corrections, step up. 8. Troubleshooting Blue "Windows protected your PC" screen at install. SmartScreen reacting to the unsigned beta installer — click More info → Run anyway. Not a virus flag. Recording fails with a microphone message. Windows is blocking mic access. Settings → Privacy & security → Microphone → allow desktop apps; SparkFlow also offers to open that page for you. Then try the hotkey again. Nothing appears in one specific app. If that app runs as administrator, Windows blocks normal apps from typing into it — your text lands on the clipboard instead; paste it with Ctrl+V, or run that app non-elevated. Text appears garbled or not at all in a stubborn app. Settings → Advanced → Paste method: try a different clipboard variant, or "Direct" (simulated typing). Dictation suddenly refuses to start. If the window that pops up says your trial ended, that's the cause — Settings → About → paste your key → Activate. The model download stalls on first run. Pick a smaller model to get going (Moonshine V2 Tiny is ~31 MB), then switch to a bigger one from Settings once it finishes downloading in the background of your day. Hotkey clashes with another app. Settings → General → SparkFlow Shortcuts → click the binding, press a new combo. One known clash in the family: TSNotes uses Ctrl+Shift+Space for quick capture — the same default as SparkFlow's AI-cleanup dictation. If you run both (Ultimate does include both), rebind one of them; whichever app started second silently loses the shortcut otherwise. 9. If something's off No license email after checkout (spam checked)? Message @tomspark on Demox with the email you used — keys can be re-issued directly. For everything else, d/tomsparkflow is the community home, and the site's refund policy is 14 days, no questions. That's SparkFlow — talk instead of type, privately. If something doesn't match what you see, post in d/tomsparkflow — every stumble in this guide gets rewritten. tomsparkflow.com → More guides → A five-minute verification routine Before dictating into something important, open an ordinary text editor and put the cursor on a blank line. Hold the normal dictation hotkey, say a harmless sentence, release, and check the result. Repeat with a name from your custom vocabulary. Then test the same sentence in the application you actually use. This separates microphone or model problems from a target app refusing pasted text. After the model download finishes, disconnect Wi-Fi and repeat ordinary dictation. Keep optional cloud cleanup off for this test. Local transcription should not require a live provider account. Turn networking back on for updates or for a cloud AI feature you deliberately configure. If you run SparkFlow alongside TSNotes, check the global shortcut collision described above before assuming either app is broken. Give AI cleanup and quick note capture different bindings. For TomSparkCal, place the cursor in its plan box and dictate an appointment, then review the calendar's date preview before saving; dictation does not itself confirm an appointment. Moving machines and protecting local history Keep your license email and record the model and microphone you prefer. On the new Windows computer, download the official installer, finish the model download, activate the included SparkFlow license and test in a text editor before configuring optional AI. A license unlock does not automatically transfer the local dictation history or custom vocabulary from the old machine. History can contain sensitive text because it records what you dictated. Review its retention controls, remove entries you no longer need and include any history you want to preserve in your normal private computer backup. Never paste an entire history file into a public support thread. Send a harmless example sentence, selected model, microphone, target application and redacted error instead. The public desktop release is Windows-first. Native mobile builds and other desktop platforms must be taken from current official release notices; source code or a testing screenshot is not proof that a public installer is available. Check d/tomsparkflow for current build instructions. --- Updated September 14, 2026. Website guide. Please keep keys and personal data out of public replies.