In daily use · Windows, with Android shipped

Talk. It types. Nothing leaves the device.

Galah is push-to-talk dictation that runs entirely on your own machine. Hold a key, speak, and the words appear where your cursor is — transcribed locally and tidied by a small AI pass. A Wispr Flow replacement you actually own.

Hold a key Speak Transcribed on-device AI tidies the text Pasted at your cursor
What it is

Dictation that treats your voice as private.

Most dictation tools stream your microphone to someone else's server. That is fine until you are dictating something you would rather not hand over — client notes, half-formed ideas, anything sensitive. Galah was built to remove that trade-off.

Speech recognition runs locally, on your own hardware, using Whisper. The only thing that ever touches the network is the cleaned-up text, sent to a small model purely to fix punctuation, casing, and the stumbles everyone makes when speaking. The audio stays put.

It started as a replacement for a paid dictation app and a way to learn the speech stack end to end. It is now the daily driver on Windows, with the same product shipping to Android.

Your audio never leaves the machine. Only the finished text does, and only to be tidied.
  • Local transcription. Whisper runs on-device. No audio upload, no account, no telemetry.
  • Clean output. A small AI pass turns spoken stumbles into text you can actually paste.
  • Everywhere you type. Whatever has focus — editor, browser, chat — the text lands there.
  • Yours to keep. No subscription, no lock-in. It runs because you run it.
How it works

From key-press to pasted text

Four steps, all but one of them on your own machine.

Step 1
Capture

Hold the push-to-talk key and speak. Modifier-suppress guards stop the hotkey from leaking stray keystrokes into the app you are typing in, and optional audio cues confirm when recording starts, stops, and bails.

Step 2
Transcribe

Whisper runs locally to turn the audio into raw text. A custom vocabulary bias nudges it toward the names, jargon, and acronyms you actually use, and it handles multiple languages.

Step 3
Clean up

The raw transcript is passed to a small, fast model (Claude Haiku) that fixes punctuation and casing and removes false starts — without rewriting your meaning. The transcript is wrapped and fenced so the model tidies text and never tries to answer it as a question.

Step 4
Paste

The cleaned text is pasted straight where your cursor sits. Galah preserves whatever was already on your clipboard and gives you a hotkey to re-paste the last dictation when you need it again.

What it does

The details that make it a daily driver

On-device STT

Whisper runs locally. No audio ever leaves your machine.

AI cleanup

A small model fixes punctuation, casing, and false starts without changing your meaning.

Vocabulary bias

Teach it the names and jargon you use so they transcribe correctly first time.

Multilingual

Dictate in more than one language; the target language is configurable.

Clipboard-safe

Keeps what was on your clipboard and lets you re-paste the last dictation on a hotkey.

Audio cues

Optional sounds confirm recording on, off, and bail — eyes-free.

Privacy

The commitments are load-bearing

These are not settings to switch on. They are how the app is built.

  • Audio never leaves the device. Ever. Transcription is entirely local.
  • No accounts, no billing, no telemetry. Nothing phones home.
  • The API key is used for text only. Cleanup sees the transcript, never the recording.
Where it runs & what is next

One product, every platform

Galah is the dictation product across Windows, Mac, Android and iOS — Currawong is its meetings sibling.

Honest about what this is: Galah is a personal tool we built and run ourselves, primarily as a way to learn the speech stack properly. It is not a commercial product — there is nothing to buy. We are showing it because building and using our own tools is how we keep our judgement about AI quality sharp.