The open-source Wispr Flow alternative

Wispr Flow’s free plan caps you at roughly 2,000 words a week on desktop. To lift that cap you move to Flow Pro at $15 a month — $144 a year, every year, for as long as you keep using it.

Bulbul has no subscription and no word cap. Instead of paying us, you paste in your own free Groq API key and pay Groq directly for what you actually use. For almost everyone that comes to nothing: Groq’s free tier is generous enough for close to a thousand dictations a day — far more than any person actually does.

And if you somehow outgrew the free tier, Groq charges about $0.04 for a full hour of transcribed audio. You would have to dictate for roughly 3,600 hours in a year — about ten hours every single day — before your Groq bill matched one year of Wispr Flow Pro. For the vast majority of people, Bulbul costs exactly nothing, forever.

Bulbul listens to your microphone and types into every app you use. That is exactly the kind of software you should be able to inspect. Every line of Bulbul is on GitHub under the GPL-3.0 license — you, or anyone you trust, can read precisely what it records, where it sends your audio, and what it stores. You don’t have to take a privacy policy’s word for it.

Wispr Flow is closed source. It may well be trustworthy, but you can’t verify it; you’re trusting a binary and a legal document.

Open source also means Bulbul can’t be taken away from you. A subscription product can raise its price, change its terms, get acquired, or shut down — and your workflow goes with it. Bulbul’s code is yours under the GPL: fork it, build it yourself, keep running the version you like. There’s no account to lock you out, and no company that can decide you no longer have access to the thing you type with every day.

This one is simple. Wispr Flow has no Linux app — Linux is a “vote for it” item on their roadmap. If you work on Linux, Wispr Flow isn’t an option today.

Bulbul runs on Linux now, on both X11 and Wayland, with .deb, .rpm, and AppImage builds. Getting dictation to work on Wayland is genuinely hard — the compositor deliberately stops apps from typing into other apps — so Bulbul types through a kernel-level virtual keyboard and reads your hotkey straight from the input device, the one path even GNOME Wayland can’t block. Hold-to-talk-anywhere works on Linux exactly like it does on Windows and macOS.

For a lot of developers, that single line — “works on Linux” — is the entire reason they’re reading this page.

Let’s be honest up front: neither app is fully offline. Both send your audio to the cloud to transcribe it — that’s how the quality is this good. The difference is the route your words take.

With Bulbul, your audio goes straight from your machine to Groq, authenticated with your own API key. There is no Bulbul-owned server in the middle — we never see your audio, because it never touches us. Your dictation history, custom dictionary, snippets, and settings all live in a plain SQLite file on your own disk that you can read, back up, or delete. The only thing that ever leaves is an anonymous usage count, and that’s a single toggle to switch off.

And Groq doesn’t keep it or learn from it. Under GroqCloud’s data policy, Groq is not permitted to use API inputs or outputs to train or fine-tune any model, and by default it doesn’t retain the data from inference requests at all. Your audio is transcribed and discarded — it never becomes training data, and no Bulbul server ever saw it in the first place. You don’t have to take our word for it; it’s spelled out in GroqCloud’s data policy and its services agreement.

With Wispr Flow, you sign into an account and your dictation runs through their service. That may be perfectly fine — but “you can read the code and you hold the key” is a stronger privacy position than “trust our servers,” and it’s one only an open, bring-your-own-key tool can offer.

Free and open doesn’t mean basic. Bulbul runs a two-stage pipeline: Whisper transcribes, then a fast cleanup model rewrites the raw text — removing filler and false starts, fixing punctuation and casing, handling self-corrections, and turning a spoken list into a real bullet list. You choose how far it goes: Raw, Clean, or Polished.

On top of that: a custom dictionary for names it keeps mishearing, snippets that expand a trigger into boilerplate, per-app tone so it writes more formally in your mail client than in chat, and transforms bound to hotkeys that rewrite selected text on the fly — translate, make formal, bulletize, or your own prompts. There’s a history-and-insights view and a standalone scratchpad. It’s a serious daily driver, not a demo.