Voice typing that shows its working

Hold a key, speak, let go. The text lands in whatever window has focus. The difference is that you watch the words appear while you are still talking, so you catch a misheard name before it reaches your document.

Download for Windows

Version 1.0.0 · 32 MB · Windows 10/11 64-bit · built 1 September 2026

SHA-256 e3822e9f8606f2ac697c7bd50ce99b56b785ec6efa1222200899e0b9b9c65bf7

A pill showing the live transcript with a waveform beside it, and a second pill reading Transcribing with a progress bar — What you see while speaking, and in the moment after letting go.

Read this before installing. SenFlow transcribes with Google's Gemini API, so while you hold the key your audio is sent to Google. It needs a free API key and an internet connection, and there is no offline mode. On Gemini's free tier, Google's terms allow them to use what you submit to improve their products, which includes everything you dictate. On the paid tier they do not. SenFlow asks which you are on during setup and explains the difference there rather than burying it.

Why it feels different

It transcribes as you speak

Audio streams out while you talk rather than after you stop, so letting go is a wait for the last few words. Measured at about 450 ms, and a sixty-second passage costs roughly the same wait as a five-second one.

You see it listening

The live transcript appears in the pill as you speak. A wrong language, a microphone pointed at the television, a mangled name: you notice before a single character is typed.

It hears the start of words

A rolling 400 ms buffer means speech beginning a fraction before the key goes down is still captured, and it is sent ahead of the live audio so the model receives one ordered stream.

Letting go early is survivable

Release mid-word and capture continues to the next natural pause instead of cutting you off. Press again during that moment and it carries on as one sentence.

Nothing is lost to a bad network

If a transcription fails the recording is kept and retried, not discarded. You are never asked to say it all again because a socket blinked.

No GPU, no model download

A 32 MB installer and nothing to fetch on first run. It works on any machine with a microphone, including laptops with no discrete graphics.

Watch it think

Gemini streams a running hypothesis while you are still talking. SenFlow shows it and types nothing until the text is final, because interim guesses are revised often enough that committing them would rewrite your sentence mid-flow.

Three pills showing a transcript growing from empty, to a few words, to a full sentence — The pill widens once and then holds still, keeping the most recent words.

Controls

Do this	And
Hold `Right Ctrl`	Speak, then release. The text is typed where your cursor is.
Tap `Right Ctrl`	Recording latches hands-free. Tap again to stop, for paragraphs you would rather not hold a key through.
`Ctrl`+`Alt`+`Space`	The same latch on a separate key, if you prefer it.
`Ctrl`+`Alt`+`Backspace`	Delete exactly what SenFlow last typed. It never eats into your own text.
"new line", "new paragraph"	Spoken alone, these insert a break rather than being typed. Said inside a sentence they are typed as normal.
"scratch that"	Removes the last insertion.

Five looks

The on-screen pill and the tray icon share a palette, switchable from the tray with no restart. Flow, Iris, Mono, Frost and Classic, in that order.

The same pill rendered in five colour themes: teal, plum, graphite, light glass and charcoal

The tray icon is the whole status display: connecting, ready, recording, finishing, error.

Installing

The installer is per-user and needs no administrator prompt. That is also functionally right: Windows will not let injected keystrokes reach an elevated window unless the sender is elevated too, so a per-user install keeps SenFlow at the same level as the apps you dictate into.

On first launch it opens a setup window, asks for a Gemini API key, and checks it against Google before saving anything, so a typo is caught there rather than at your first dictation. Get a free key at aistudio.google.com/apikey.

To start it with Windows, tick the box during setup. Settings live in a commented file you can open from the tray.

%APPDATA%\SenFlow\config.toml

What it needs

Windows 10 or 11, 64-bit.
A microphone. Pick one from the tray; if it is unplugged later, SenFlow falls back to the Windows default and switches back when it returns.
A Gemini API key, free from Google AI Studio.
An internet connection, every time you dictate. There is no local mode.
No GPU, and nothing downloaded after the installer.

Good to know

Transcripts are never written to disk. The tray keeps the last twenty in memory so you can recover one that went into the wrong window, and they are gone when you quit.
Usage is metered. Every utterance is a billable API call on the paid tier. The tray shows tokens and an estimated cost, and senflow --usage prints a per-day breakdown.
Sessions last ten minutes. Google's limit, handled by recycling the connection between utterances so you never meet it.
Elevated windows are off limits. Text cannot be typed into an application running as administrator unless SenFlow is too. That is a Windows security boundary, not a limitation of the app.
85+ languages, detected automatically if you want, including switching mid-sentence.
Names and jargon can be listed in the config, which biases the recogniser itself rather than patching the output afterwards.