OpenWhispr is a free, open-source voice-to-text dictation app for macOS, Windows and Linux that types what you say straight into whatever window has focus. It is built for people who write all day — email, notes, chat, documents, code comments — and especially for those who would rather not have a recording of their voice sitting on a vendor's server. The project describes itself as an open alternative to WisprFlow and Granola, and it covers both sides of that comparison: press-to-talk dictation, and live transcription of meetings.
What it does
You bind a hotkey, press it, speak, and the transcribed text appears at the cursor. That works in any app, because the text is inserted where you are already typing rather than in a separate window you have to copy out of.
Around that core the app adds a few things:
- Local transcription with Whisper or NVIDIA Parakeet, where the audio never leaves the machine.
- Optional cloud transcription with your own API key, when you want more speed than your hardware gives you.
- A built-in assistant that can route what you dictate to Claude, GPT-5 or Gemini, so a spoken instruction becomes an answer rather than just text.
- Live meeting transcription with speaker labels, running on-device.
- Notes, so dictated material has somewhere to land.
The project states that there is no data collection and no telemetry, and that nothing is recorded, logged or sent anywhere without your permission.
How it works
Two processing paths sit behind the same hotkey. In local mode a speech-to-text engine — Whisper or NVIDIA Parakeet — runs on your own machine, which is what makes the privacy claim mean anything: there is no upload step to trust. In cloud mode you supply your own key and the app talks to that provider directly; the repository's topics name OpenAI, Anthropic, Gemini and Groq. This is bring-your-own-key throughout, so there is no subscription to OpenWhispr itself — you pay your model provider, or you pay nobody.
The same split applies to the assistant features. Meeting transcription with speaker labels is described as fully on-device, while the chat-style assistant runs against whichever model you have configured.
The codebase is mostly JavaScript, released under the MIT license, and shipped as a desktop application with prebuilt releases for all three platforms. The project also publishes an API overview in its documentation, so the app is not the only entry point.
Getting started
The simplest route is the packaged build: the repository links a download page to its latest GitHub release, alongside a project site at openwhispr.com and documentation at docs.openwhispr.com. Installation is the ordinary desktop routine rather than a build from source, though the source is there under MIT if you want it.
The first-run decisions are the ones that matter: which engine you want — a local model, or a cloud provider you paste a key for — and which hotkey you are willing to give up. A changelog is kept in the repository, worth skimming, since the project has been pushing changes steadily since it was created in mid-2025.
When to use it / when not
Reach for OpenWhispr when the privacy constraint is real — client calls, medical or legal notes, anything under an agreement that forbids sending audio to a third party — or when you simply resent paying a monthly fee to dictate. It is also the obvious pick if you work offline or on restricted networks, since the local path needs no connection at all.
It is a weaker fit in a few cases. Local models run on your own CPU or GPU, so transcription speed depends on your hardware; the project's own answer to that is the cloud path, which trades the privacy guarantee back away. There is no mobile version — the supported platforms are macOS, Windows and Linux desktops. And as a community project it comes with community support: issues and documentation rather than an account manager.
Alternatives
The project positions itself directly against WisprFlow for dictation and Granola for meeting notes, both commercial and hosted. Against those, the trade is the usual one: you give up a polished managed service and get source you can read, a license you can keep, and audio that stays put.
The other alternative is assembling this yourself, since Whisper and Parakeet are both available as models you can run directly. That is a fine path if what you want is a transcription pipeline in a script, but it is not a hotkey that puts text at your cursor in any application, which is most of what this repository is.
Anyone who dictates more than occasionally and has any reason to care where the audio goes should try OpenWhispr before paying for anything else. It is free, MIT-licensed, actively developed, and it has collected over seven thousand stars in roughly a year, which suggests the local-first framing found a real audience. If you have no privacy constraint and want the smoothest possible product, a hosted paid tool may still suit you better — but then you are choosing convenience over control knowingly, which is the right way to make that choice.