Thoughtborne — Hotkey dictation for Windows

Thoughtborne in the terminal: the ready screen with the four-engine model carousel and the key overview on top, below it a recording from the REC notice to the inserted-text confirmation.

It is operated by keyboard shortcut. The terminal is not a control window but a display: it shows which engine is currently active, which keys are bound, and what happened last. Nothing here is clicked — the one exception is a small window for the settings.

What’s inside

Thoughtborne has been in daily use since 2023. Over that time it has gained the things you actually need in continuous operation:

Transcription & engines

Four engines, switchable at runtime — Soniox Live, Soniox, Groq Whisper Large & Turbo, cycled with Ctrl+Alt+L, no restart.
Real-time live transcription — Soniox Live streams while you speak; the text is ready ~0.5 s after you let go.
German hallucination filters — model-specific, against known speech artifacts, tuned deliberately for German.
Filler-word filter — “ähm”/“äh” removed safely at word boundaries (spares “ähnlich”), meaning-bearing “naja”/“mhm” stays.
Recognition vocabulary — your own names and technical terms go to the engine as context and come back spelled right.
Automatic engine fallback — if one Soniox variant fails, the REST variant steps in — transparently.

Nothing gets lost

Four-layer safety net — exit salvage, crash sidecar, startup recovery, device loss — no dictation is ever lost, even on a crash.
Retry hotkey — Ctrl+Alt+R re-transcribes the last failed recording.
Parallel dictation — start the next recording while the previous one is still processing — the order is preserved.
Local archive — audio and transcript, timestamp-paired, on your disk only, no cloud storage.

Everyday operation

Four stop modes — insert by typing or clipboard, insert + Enter to send, or just transcribe and insert later.
Push-to-talk (optional) — hold a key, talk, release; robust against QWERTZ false triggers and Modern Standby.
Self-test by hotkey — Ctrl+Alt+T checks the whole pipeline end to end.
Archive-folder hotkey — Ctrl+Alt+6 opens the archive directly in Explorer.

Setup & customization

uv setup with no pre-installed Python — uv run pulls Python and dependencies itself, and re-syncs after a git pull too.
Free entry path — starts with just one key, skips missing APIs cleanly instead of crashing.
Adaptable by AI agent — readable code instead of a settings maze; AGENTS.md ships with it so your coding agent can get going right away.

Why Thoughtborne exists

Working with LLMs is working with text, and the keyboard is the bottleneck. Commercial dictation tools are plentiful; none of them hit everything at once — German, quality, price, data control. That is exactly what this is about:

Built for German — not merely “also supports German”.

Almost all competitors are English-first. Here German is the starting point of every decision.

The engine choice follows real German dictation use — curated provider-independently and kept continuously up to date: the aim is the best available cloud model for German at any given time; if something better appears, it gets adopted.
Hallucination filters from the maintainer’s own German transcripts, annotated since 2023 — model-specific.
Filler-word filtering and vocabulary handling for German terms and proper nouns.

No subscription. A free path, or ~$3/month for heavy use.

You bring your own API keys and pay only for what you use — a fraction of the $12–15/month of the subscription tools.

Groq’s free tier as a genuine everyday path, not a trial: up to ~8 hours of audio a day.
Soniox: $0.12 per hour of audio, prepaid.
~$3/month, measured, not estimated — the maintainer at ~25 h of audio a month.

Your data: verifiably under your control.

No need to trust a privacy policy — you can read for yourself where your audio goes.

Audio goes only to the API you choose, nowhere else. The archive sits locally on your disk.
No training on your data is a hard inclusion criterion — APIs without an opt-out are dropped.
Open source, readable code: read it yourself instead of believing.

Independent: four engines, no lock-in.

If a provider gets worse or pricier, it drops out of the lineup — you are never stuck with one.

Switchable at runtime with Ctrl+Alt+L, curated provider-neutrally.
Two styles as a side effect of the breadth: verbatim-instant or smoothed.

Built for talking to AI.

Not dictation in general, but the way into an ongoing AI conversation — without touching the keyboard.

A dedicated hotkey inserts and presses Enter — whole AI conversations by voice.
The quality bar behind every model choice: “good enough to send unread”.

Adapts to you — just tell it, if need be.

The tool bends to your workflow, not the other way around. No settings jungle, just readable code.

Readable Python, config.py with plain-text constants and the recognition vocabulary.
“Tell your coding agent what it should be able to do” — AGENTS.md ships with it.

The loop

Ctrl+Alt+W Start recording and talk.
Ctrl+Alt+A Stop: the text lands at the cursor position.
Ctrl+Alt+D Or insert and press Enter — whole AI conversations without a keyboard.

Prefer to hold a key? Optional push-to-talk: hold, talk, release.

Reliable, recording after recording

Instantly thereThe live engine transcribes while you speak, ~0.5 s after stop.
Never lose a recordingRetry hotkey plus a safety net for crash, exit, device change.
Dictate in parallelStart the next recording while the previous one is still processing — the order holds.
Local archiveEvery recording + transcript as a pair on your disk.
Transcribe first, insert laterWith Ctrl+Alt+Y, when the cursor is still somewhere else.
Two insert pathsSimulated typing or clipboard paste.
Push-to-talk (optional)Hold, talk, release — off by default.
Self-test by hotkeyChecks the whole pipeline end to end.

What it costs

Nowhere in this stack a subscription. You bring your own API keys and pay the providers’ usage prices — a fraction of the $12–15/month that subscription dictation tools charge.

$0 the Groq path: A genuine free tier, not a trial. Both Groq engines run on Groq’s generous free tier — up to ~8 hours of audio a day, enough for daily use.
$0.12 per hour of audio: Soniox is the paid path. No free tier: top up a small prepaid balance and pay only for what you actually transcribe.
~$3 per month, heavy use: Measured, not estimated. The maintainer dictates ~25 hours of audio a month, averaged over half a year. Lighter use lands at around a dollar.

An honest note: the quality bar above — German good enough to send unread — is measured against the paid Soniox engine. The Groq engines are strong; the bar hangs on Soniox.

Who it’s built for

A fit if you …

want a tool you control yourself — open source, no subscription.
want to dictate for free — even with regular use.
want something reliable, in daily use since 2023.
want transcripts you can send unread — in German too.
spend a lot of time with chatbots and coding agents — ChatGPT, Claude Code, Codex and the like.
type long texts and would rather just say them.

Get started

You need Windows, a microphone, and one API key — free with Groq. One terminal command installs everything, without admin rights and inside your user profile — you can read the script first. The setup wizard then asks for your key, and Ctrl+Alt+W starts your first dictation.

Install with one command Download ZIP Setup guide (README)

Rather not paste a command? Unzip the ZIP and double-click setup.bat — the same setup. And if you want full control, follow the README.