OpenWhispr is a free, MIT-licensed, open-source voice-to-text desktop application for macOS, Windows, and Linux that turns speech into text, notes, and actions at the cursor, and it is built for anyone who wants dictation and meeting transcription without sending audio to a third party.

What it is

OpenWhispr is a cross-platform desktop application written in JavaScript that provides voice dictation, meeting transcription, and notes with AI agents. A user presses a hotkey, speaks, and the transcribed words appear at the cursor in whatever application has focus. Transcription can run fully offline against local speech-to-text engines such as Whisper and NVIDIA Parakeet, in which case audio never leaves the device, or it can be routed to cloud providers when speed matters more than locality. The project ships no data collection and no telemetry.

The concrete problem it solves is the paid, cloud-only dictation service. OpenWhispr positions itself directly as the open-source and free alternative to WisprFlow and Granola, so it lives in the same Productivity & Utilities / Input & Dictation space while replacing the subscription model and the requirement that speech be uploaded for processing. The choice between local engines and hosted models is exposed to the user rather than fixed by the vendor, and local inference is GPU-accelerated across Metal, CUDA, and Vulkan, which covers AMD and Intel graphics as well as Apple silicon and NVIDIA hardware.

Key capabilities

• Global hotkey voice dictation that automatically pastes the transcript into the focused application, plus a separate dictation translation hotkey that dictates in one language and pastes the text in another.
• AI agent access to GPT-5, Claude, Gemini, Groq, Tinfoil, OpenRouter, or local models behind a named voice assistant.
• A dedicated Voice Assistant hotkey that sends speech directly as a command with no wake word and no cleanup pass, edits highlighted text in place, and can opt in to sending a screenshot of the current screen as context.
• Meeting transcription that auto-detects Zoom, Teams, and FaceTime calls, with live speaker diarization, voice fingerprinting, and Google, Microsoft, or Apple Calendar integration; diarization and fingerprint recognition also run on-device without cloud.
• Notes with folders, semantic search, cloud sync, and AI actions, plus team spaces and sharing through link, domain, or invite-only visibility with roles, invitations, and server-enforced membership.
• Audio import that transcribes existing audio and video via drag-in files, batch upload, or a pasted YouTube or audio URL, with optional speaker detection.
• Enterprise controls for organization policy enforcement, company SSO and SCIM, and centrally managed Amazon Bedrock or Azure OpenAI access without distributing cloud keys, alongside a Public API and an MCP server for programmatic note and transcription management.

Who uses it and how

• Individual knowledge workers who want dictation available inside every application rather than in a single dedicated editor, using one global hotkey and automatic pasting.
• People working with sensitive speech, who select local Whisper or NVIDIA Parakeet engines so that audio never leaves the machine.
• Meeting-heavy users and teams who need automatic detection of Zoom, Teams, and FaceTime calls, speaker labels, and calendar-linked transcripts, with notes searchable semantically across meetings.
• Teams that share notes through link, domain, or invite-only visibility and collaborate in team spaces with roles and invitations.
• Regulated or large organizations that enforce policy centrally and route AI calls through Amazon Bedrock or Azure OpenAI without handing out cloud keys.

Getting started

Install by cloning the repository and running npm install followed by npm run dev, which requires Node.js 24 or newer, or download a prebuilt package for the relevant platform: .dmg for macOS, .exe for Windows, and .AppImage, .deb, .rpm, or .tar.gz for Linux.

How it compares

WisprFlow and Granola are paid products, while OpenWhispr is MIT-licensed and free, so the cost model shifts from subscription to whatever local hardware or cloud API keys the user already has. Those products process speech in their own cloud, whereas OpenWhispr can keep transcription, diarization, semantic search, and AI reasoning entirely on the local machine, with no telemetry or data collection. Self-hosting and provider choice are therefore under the user's control, including the ability to point AI requests at Bedrock, Azure OpenAI, or local models instead of a vendor-fixed endpoint.

When to use it — and when not to

A self-hoster must supply the compute: local transcription and semantic search depend on running models, and GPU acceleration is the practical path to usable speed, so machines without that capacity will fall back to cloud providers or slower processing. On Intel Macs, live speaker identification and voice fingerprinting are unavailable because ONNX Runtime stopped shipping macOS x86_64 binaries in 1.24, and notes search falls back to keyword matching instead of semantic search, though meetings still record and transcribe normally. Anyone unwilling to configure engines, manage API keys, or tolerate a project carrying 341 open issues should consider a fully managed paid service instead.

git clone https://github.com/OpenWhispr/openwhispr.git
cd openwhispr
npm install
npm run dev