Where does my audio go?

Audio is captured locally and sent directly from your machine to Groq for transcription, using your own API key. It does not pass through any Bulbul server.

Yes — Bulbul is free and open source. You only pay for your own Groq API usage, billed by Groq, not by us. There’s no subscription.

Does it work in any app?

Yes. On desktop Bulbul pastes into whatever window has focus — mail clients, editors, chat boxes, docs. On Android a floating bubble appears with your keyboard and drops text into whatever field you're typing in. If you can type there, you can dictate there.

Is Mac, Linux, or Android supported?

Yes — all of them. Bulbul runs on Windows 10/11 (x64), macOS 11+ (universal — Apple Silicon and Intel), Linux on X11 and Wayland (.AppImage, .deb, .rpm), and Android (arm64). Desktop uses a hold-to-talk hotkey; on Android you tap or hold a floating bubble.

How is this different from Windows voice typing?

Windows voice typing transcribes what you say literally. Bulbul adds a cleanup pass that removes filler and false starts, remembers your corrections, is open source, and runs on Mac, Linux, and Android too. Both send audio to a cloud model — Windows Speech to Microsoft, Bulbul to Groq using your own API key — but Bulbul keeps your history, dictionary, and snippets in a local SQLite file.

A maker at a desk on a balcony, coding above a calm river valley

Free & open sourceYour own Groq keyNo Bulbul servers

Talk. Bulbul types.

A free, open-source dictation app for Windows, Mac, Linux, and Android. Hold a hotkey — or tap the bubble on your phone — speak, release, and cleaned-up text lands where your cursor already is.

Star on GitHub

PS>irm https://bulbultypes.xyz/install.ps1 | iex

Download for other operating systems

Free on Windows · macOS · Linux · Android

What it does

Three things, done quietly well.

No commands to memorize, no dashboard to babysit. Hold, talk, and clean text appears where your cursor already is.

Hold-to-talk, anywhere

Press and hold your hotkey in any window. Bulbul captures while you speak and stops the moment you let go.

Ctrl+Win

Cleanup happens automatically

Filler words, false starts, and stray “ums” are removed. What lands is the sentence you meant to say.

um, so, like let’s ship it.

Learns your corrections

Fix a word once and Bulbul remembers. Names, jargon, and product spellings get more accurate over time.

definatelydefinitely

Messy spoken filler words drifting across a desk and resolving into a clean checklist — your words, tidied on the way out.

Cleanup

From the way you talk to the way you write.

Speech is messy — restarts, hedges, “ums.” Bulbul rewrites it into the sentence you intended, then pastes that.

As spoken

So, um, I basically wanted to, you know, follow up about the, uh, the proposal — and honestly I think we should just ship it.

As pasted

I wanted to follow up about the proposal — I think we should ship it.

How it works

Hold, speak, release.

The whole interaction is three moves. Everything technical happens in the half-second between letting go and seeing your words.

01
Hold the hotkey
Press and hold from inside any app — no window switching.
Ctrl+Win
02
Speak naturally
Talk the way you think. Bulbul listens until you release.
03
Release — it’s pasted
Cleaned text drops straight into the field you were in.
Pasted

A live waveform and rough speech becoming a clean sentence on screen — press · speak · release

Under the hood

01Captured locally
Audio is captured on your device via cpal, then streamed to Groq for transcription.
02Transcribed
Groq Whisper, using your API key.
03Cleaned up
An 8B-class model polishes the text.
04Pasted in place
Inserted where your cursor already is.

Works in any app

Same hold-to-talk. Any window.

Bulbul types into whatever already has focus — your mail client, your editor, your chat box. Pick a destination and watch.

Email, chat, notes, and a code editor open around one laptop — one shortcut, every window.

New message

Tosupport@bulbultypes.xyz

SubjectRe: the Q3 proposal

Ctrl+Win

Listening

Speed

Fast enough to feel like nothing.

Bulbul targets sub-second end-to-end — from the moment you stop speaking to text appearing in the app you’re focused on.

A laptop showing a live waveform, with a transcribe → cleanup → insert timing diagram beside it — release → text, in about 400ms.

Typical timings

Raw

250–400ms

Transcription only — no cleanup

Clean· default

400–600ms

Transcription + fillers & punctuation fix

Polished

500–700ms

Transcription + rewrite for clarity

Most dictation lands between 300 and 500ms — fast enough that you don’t notice the round-trip through the cloud.

Where the milliseconds go

feels-instant · 500ms

0≈ 380ms total · Clean700ms

01~200ms
Whisper Large v3 Turbo
Transcribes the audio
02~150ms
Llama 3.1 8B
Cleans it up — skipped in Raw mode
03~30ms
Win32 input injection
Drops text into your focused app

The “feels instant” threshold for keyboard-equivalent input is roughly 500ms. Bulbul stays under it in normal use.

Key features

More than a microphone button.

Everything is built around one loop — speak, clean, paste — then sharpened with the tools that make daily dictation actually stick.

A desk of dictated notes with hand-marked corrections and checkmarks — every fix, remembered.

Correction memory

Teach it once. It remembers.

Fix a name or a piece of jargon a single time and Bulbul keeps it. Your personal dictionary builds itself from the corrections you actually make.

namesjargonspellingsacronyms

Cards labelled expand, reformat, refine and reuse around a laptop — four ways to reshape.

Transforms

Reshape text, not just dictate it.

Captured something rough? Expand a bullet into a paragraph, reformat a list, or refine a draft — without leaving the app you’re in.

expandreformatrefinereuse

Hold-to-talk dictation

Press and hold to capture, release to insert.

Ctrl+Win

Speak raw, paste polished

Hold the hotkey and just talk — Bulbul pastes the cleaned-up, polished version, never the raw transcript.

Shift+Alt+P

Per-app context

Tone adapts — terse in chat, fuller in mail and docs.

mail · chat · code

Snippets

Reusable blocks — signatures, boilerplate — dropped in by voice.

/snippet

Scratchpad

A local holding space for dictations before they land.

local · private

Insights

Words dictated, time saved, and your most-made corrections.

on-device stats

Privacy & trust

A trust ledger, not a privacy promise.

The honest version of “your data is safe”: here is exactly where everything goes, line by line.

Bring your own Groq key

BYOK

You supply the API key. Bulbul never ships or proxies one for you.

No Bulbul servers

direct

Audio goes from your machine straight to Groq. There is nothing in the middle to trust.

Local storage

SQLite

History, snippets, and settings live in a SQLite file on your PC — not a cloud account. Audio itself goes to Groq via your own API key for transcription; nothing else leaves your machine.

Signed releases

minisign

Every build is signed with minisign so you can verify exactly what you’re running.

Anonymous usage stats are on

on by default

Bulbul is solo-built — counts and error categories help me know what to fix. Never your transcripts, audio, dictionary, or which app you’re typing into. Turn it off anytime in Settings → Privacy.

Source available

open

The code is on GitHub. Read the data path and audit it for yourself.

Comparison

It types what you actually said.

Windows already includes voice typing — but it transcribes literally, so misheard words, shaky punctuation, and filler land in your text as-is. Here’s the line-by-line picture; Bulbul’s column is what ships today, the others reflect well-known traits.

Accurate transcription

Bulbul
Windows Voice Typing
Other providers

Punctuation & casing

Bulbul
Windows Voice Typing
Other providers

Automatic cleanup

Bulbul
Windows Voice Typing
Other providers

Fixes filler & false starts

Bulbul
Windows Voice Typing
Other providers

Learns your corrections

Bulbul
Windows Voice Typing
Other providers

Free to use

Bulbul
Windows Voice Typing
Other providers

No weekly word cap

Bulbul
Windows Voice Typing
Other providers: ~2k/wk

No subscription

Bulbul
Windows Voice Typing
Other providers

Open source

Bulbul
Windows Voice Typing
Other providers

Works in any app

Bulbul
Windows Voice Typing
Other providers

Bring your own key

Bulbul
Windows Voice Typing: —
Other providers

No account required

Bulbul
Windows Voice Typing
Other providers

Local storage

Bulbul
Windows Voice Typing
Other providers

Signed releases

Bulbul
Windows Voice Typing: —
Other providers: —

Platform

Bulbul: Win · Mac · Linux · Android
Windows Voice Typing: Windows
Other providers: mostly Mac

Capability	Bulbul	Windows Voice Typing	Other providers
Accurate transcription
Punctuation & casing
Automatic cleanup
Fixes filler & false starts
Learns your corrections
Free to use
No weekly word cap			~2k/wk
No subscription
Open source
Works in any app
Bring your own key		—
No account required
Local storage
Signed releases		—	—
Platform	Win · Mac · Linux · Android	Windows	mostly Mac

varies by tool · Comparison is informational, not a benchmark.

How it’s built

Honest engineering, all the way down.

If you’re going to inspect it, here’s the stack. Native where it matters, open where it counts.

A workspace of open-source notes, architecture sketches, and a code editor — build with care, share freely.

Tauri 2

A native webview shell — not Electron. Small binary, low memory.

Rust core

Audio capture, global hotkeys, and the local store run in Rust.

React 18 + Vite

The interface is React 18, bundled and served by Vite.

Win32 APIs

Global hotkeys and paste-at-cursor go through native Win32.

cpal

Microphone capture via the cross-platform cpal audio crate.

Groq Whisper + 8B

Transcription on Whisper, cleanup on an 8B-class model.

SQLite

Every local record lives in a single SQLite database file.

minisign + updater

Signed builds with a signature-verifying auto-updater.

Install

Up and running in a minute.

Download the signed installer for your OS, or paste one line into your terminal. Windows · macOS · Linux · Android.

GitHub Releases

PS>irm https://bulbultypes.xyz/install.ps1 | iex

Download for other operating systems

On first run, Bulbul asks for your name and your Groq API key. That’s the whole setup — no account, no sign-in.

Bulbul's onboarding — a “you’re all set” screen beside a terminal reading all good — you’re all set.

FAQ

Questions, answered plainly.

No. There’s no sign-up and no login. On first run it asks for your name and your Groq API key, and that’s it.

Talk. Bulbul types.

Three things, done quietly well.

Hold-to-talk, anywhere

Cleanup happens automatically

Learns your corrections

From the way you talk to the way you write.

Hold, speak, release.

Hold the hotkey

Speak naturally

Release — it’s pasted

Same hold-to-talk. Any window.

Fast enough to feel like nothing.

More than a microphone button.

Teach it once. It remembers.

Reshape text, not just dictate it.

Hold-to-talk dictation

Speak raw, paste polished

Per-app context

Snippets

Scratchpad

Insights

A trust ledger, not a privacy promise.

Bring your own Groq key

No Bulbul servers

Local storage

Signed releases

Anonymous usage stats are on

Source available

It types what you actually said.

Honest engineering, all the way down.

Up and running in a minute.

Questions, answered plainly.

Does Bulbul require an account?

Where does my audio go?

Is Bulbul free?

Does it work in any app?

Is Mac, Linux, or Android supported?

How is this different from Windows voice typing?