Hold a key and speak — your words land wherever the cursor is. Say how a selection should change, and I rewrite it in place. Ask, and I'll read your screen aloud. All of it on your machine. No cloud, no account, nothing to sign into.
Hold Ctrl + Win, talk, let go. Your words appear at the cursor — Slack, Outlook, VS Code, Notion, that one internal tool nobody else supports. Terminals, vim and remote desktops hate paste, so I type the words in for them instead, key by key. Either way: you talk, text lands.
Select text anywhere in Windows, hold Alt + Win, and say how it should change — "make this sound more confident", "turn this into bullet points", "translate to French". Let go, and the rewrite streams into place, live, right where the selection was.
The report is late again. Send the corrected version today.
Hi! Just checking in on the report — could you send the corrected version over today? Thanks so much!
Alt and Win are pure modifiers — the chord physically can't type a character, so it can never type over your selection
Select something — or just copy it — and press the hotkey. I read it aloud through your speakers in a voice that doesn't sound robotic. I synthesize sentence by sentence, so I start speaking almost before you finish pressing. Press Esc and I stop mid-word.
A little local AI sweeps up after you — the "um"s go, the stutters untangle, the punctuation lands. That's everything it's allowed to do. Every cleanup has to pass a gate before it can touch your text: drop a phrase, rewrite your phrasing, soften how hard you said something — rejected, and your raw words ship instead. I keep your words exactly as you said them. Including the spicy ones.
every transcript in History keeps both layers, word-level diff included — flip between them any time
Teach each app its own voice — buttoned-up in email, loose in chat, tidy markdown in your editor. Give an app a custom style and the AI gets real creative freedom. Anywhere you haven't set one, the default holds: your words, exactly as you said them, just tidied.
None of these need a settings safari. They're just there, quietly making every dictation a little more yours.
Names, jargon, that one client nobody can spell. Add a word once and I bias recognition toward it on every engine, forever. No retraining, no cloud roundtrip.
Say a trigger phrase, get a block of text — your address, your signature, that prompt you paste eleven times a day. Spoken shorthand for anything you retype.
Every dictation is kept — audio and text, raw and tidied — on your own disk. Search it, flip the diff, copy yesterday's paragraph again. Nothing auto-deletes.
Pull the network cable mid-sentence; I don't notice. Planes, tunnels, cabins, hostile hotel Wi-Fi — every feature keeps working, because nothing needs a server.
The hold-to-talk chords are pure modifiers — they physically can't type a character, so they can never destroy a selection. Esc stops everything, globally. Bindings survive lock and unlock.
Releases are signed, downloads are SHA256-verified, and the updater refuses anything that doesn't check out. One click to install, and you're back to talking.
Cleanup, Edit Mode, per-app styles — they all need a language model, so I ship one. It downloads itself once and then lives entirely on your machine. No account to create. No API key to paste. No cloud to trust.
The built-in brain listens only on 127.0.0.1 — your own machine talking to itself. It's contained so it can never outlive the app: close tellmebaby and the brain is gone with it. Your text goes in, better text comes out, and the network never hears about either.
No cloud accounts. No telemetry. No analytics beacons. The only network traffic is downloading models the first time — open-weight files you can delete and re-fetch any time — plus a small signed check for updates. Your voice never travels.
Recognition, cleanup and speech all run on your CPU or GPU from locally-stored model files. No microphone audio is ever sent over the network. We say "100% local" and we mean it — open Wireshark and watch nothing happen.
Recordings are ordinary .wav files. Transcripts live in a SQLite database any client can open. Settings are JSON. Everything sits in one folder you can browse, back up, or delete — deleting it is a complete reset. Nothing hidden, nothing "synced for your convenience."
"100% local" is easy to claim. Here's the technical reality so you can verify rather than take our word for it.
The recognizer is sherpa-onnx, an open-source ONNX Runtime wrapper that runs INT8 model weights on your CPU. No code we write touches a network socket during transcription. Open Wireshark; watch nothing happen.
Every release is signed with an ed25519 key (separate from Authenticode) and the public half is embedded in the app. The auto-updater refuses to install anything that doesn't verify against it. You can't be tricked into installing a malicious "update."
SHA256 of every release is published right next to the download button. certutil -hashfile setup.exe SHA256 confirms what you got matches what we shipped — no MITM, no tampered binaries. Model downloads are pinned and SHA256-checked the same way.
The default models — Parakeet TDT and Whisper Large v3 Turbo — ship under permissive open licenses. You can fork tellmebaby, swap the models, redistribute. No vendor can pull the rug.
The built-in language model binds to 127.0.0.1 only and runs inside a Windows Job Object, so it can never outlive the app or accept a connection from outside your machine. Your drafts never become anyone's training data.
No paid tier planned. No telemetry to monetize. If maintenance ever shifts to paid, current installs keep working at the price they were installed at. The version you have now is the version you have forever.
tellmebaby ships speech engines that pick themselves automatically based on the languages you speak. The defaults rank near the top of the public Hugging Face Open ASR leaderboard. All of them run in INT8 on a normal laptop CPU at multiples of real-time.
| Model | WER ↓ | Speed (RTF) ↑ | Languages | Size |
|---|---|---|---|---|
| Parakeet TDT 0.6B v2 | ~6.05% | ~30× | English | 660 MB |
| Whisper Large v3 Turbo | ~7.4% | ~4× | 99 (multilingual) | 1.0 GB |
| Parakeet TDT 0.6B v3 | ~6.3% | ~30× | 25 EU (incl. English) | 660 MB |
| Cohere Transcribe 2B | ~5.4% | ~12× | 14 (EU + zh/ja/ko/vi/ar) | 1.7 GB |
| OmniLingual ASR 300M | varies | ~6× | 1600+ (low-res. fallback) | 700 MB |
Long recordings decode as silence-bounded windows, so a ten-minute ramble comes out whole. If a decode comes back blank, I retry down a ladder — drop the hotwords, split on silence, switch to a different engine architecture. And if a stretch still couldn't be decoded, the transcript says so. I'd rather admit a gap than invent words.
Picked so they don't fight your browser, your editor, or Windows itself. Ctrl+Shift+Alt + a letter is essentially never claimed by anything else — so the chords stay yours.
all five are remappable in Settings → Activation
Esc is a true global stop — cancel a capture or silence a read-aloud from anywhere. Bindings survive lock and unlock.
%USERPROFILE%\.tellmebaby where you can delete and re-fetch them any time. The brain runs only on 127.0.0.1 and shuts down with the app.
%USERPROFILE%\.tellmebaby — recordings as .wav, transcripts in a SQLite database, settings as JSON. Nothing's compressed, nothing's encrypted, nothing's hidden. You can browse it all directly in File Explorer. Deleting that folder is a complete reset.
Win + letter at the shell level (so we never see Win+R, etc.), and browsers eat Ctrl+Shift+letter shortcuts (Ctrl+Shift+R reloads, Ctrl+Shift+V pastes plain text). Ctrl+Shift+Alt + letter is essentially never claimed by anything else, so the chords stay yours no matter which app is in front. The two hold-to-talk actions go further: dictation (Ctrl + Win) and Edit Mode (Alt + Win) use modifier-only chords — keys that physically can't type — so nothing can ever land in your document, or over your selection, by accident. Re-bind to whatever you prefer in Settings → Activation.
Run the installer, pick the languages you speak, and start talking in under five minutes. Updates are automatic and signed.
Download v0.3.6 for Windows