OpenWhisper

A desktop app for dictation, audio-file transcription, and meeting notes on Windows, macOS, and Linux. Transcribe with local speech models, a speech server you run, or a cloud API, then optionally use AI, including models you run yourself, to clean up the text.

Website · Download · Run from source · Install · Changelog

Use a packaged release with bundled Python and dependencies, or run from source using your own Python environment.

Feedback: If you’re unhappy with any part of the app, please open a GitHub issue and I’ll do my best to make it right ASAP.

What it does

• Dictation: Record from any app with a global hotkey, see a live preview on supported engines, and paste the result into the active window.
• Walkthroughs: Talk through a task while keeping screenshots, then get a PDF, Word and Markdown document with each screenshot after the words that describe it, ready to paste into a coding agent. Review and correct saved walkthroughs, or have your own agent write one up. Windows, macOS, and Linux (X11, Hyprland); see the Walkthrough guide.
• Audio files: Transcribe one file or a queue, keep results separate or combine them, and copy the output when ready.
• Meetings: Capture microphone and system audio, follow a live dashboard, review searchable transcripts and AI insights, play recordings, and export. Available on Windows and macOS 13+; Linux system audio is a preview.
• AI cleanup: Choose Light, Medium, or High cleanup, which handles spoken corrections and lists, then apply spelling and style rules or reusable cleanup profiles, with separate text-model choices for dictation and meetings.
• Personalize: A dictionary that steers recognition and learns from your corrections, spoken snippets, a tone for each kind of app, and Command Mode to rewrite selected text by voice or with saved transforms.
• Extras: Hands-free recording by double-tapping the record key, mouse side buttons as shortcuts, ranked microphones with automatic fallback, an overlay language switcher, a floating Scratchpad, and dictation Stats.
• Settings: Opens on Home, with where your words go, the everyday controls, and what's downloaded. Everything else is in six sidebar sections (Dictation, Personalize, Automate, Meeting Mode, Models & storage, App), and the header search (Ctrl+K) finds any setting. New installs get an optional guided setup, which File → Guided setup… reopens.
• History: Search, play back, retranscribe, and export transcripts as Markdown, plain text, or JSON, and switch an entry between its original and AI-cleaned versions. See export format support for what each format includes.
• Agent history API (preview): Search saved transcripts and meeting notes through an opt-in, authenticated local API. See API setup.
• MCP: Enable Settings → Automate → API & MCP to connect an agent to history, transcripts, and meeting insights. Optionally allow retitling and choose which settings agents may change. See MCP setup.
• Actions and automation: Send what you say to a new Claude or ChatGPT chat, your coding agent, a running Claude Code session, a webhook or a program, with a shortcut or by naming it ("Claude, …"). Follow live events, run actions and change dictated text from your own tools over a local API, webhooks and MCP. See Actions, live events and hooks.

The app also includes microphone selection, a system tray where available, and dark, light, or system-matched themes.

Install

Download the package for your platform from Releases. Native packages include Python and the app dependencies.

Platform	Install
Windows	Run the .exe installer. It installs per-user without admin rights.
macOS	Open the .dmg and drag OpenWhisper into Applications. Requires Apple Silicon and macOS 14+. Unless a release says it is notarized, approve first launch in Privacy & Security → Open Anyway. Intel Macs run from source.
Linux	Install the .deb on Debian 12+ / Ubuntu 22.04+, or .pkg.tar.zst on Arch-compatible distributions. Both packages require x86_64.

On Linux, use the matching command, replacing <version> with your download's version:

sudo apt install "./OpenWhisper-<version>-linux-amd64.deb"
# or
sudo pacman -U "./OpenWhisper-<version>-linux-x86_64.pkg.tar.zst"

Launch from your app menu; Linux packages also provide ow and openwhisper commands. Opening the app again on Linux brings the existing window forward.

Downloads include SHA256SUMS.txt for verification. Compare your file's hash using Get-FileHash in PowerShell, shasum -a 256 on macOS, or sha256sum on Linux. The Windows installer is unsigned; if SmartScreen blocks it, verify the download before choosing More info → Run anyway.

macOS permissions

In System Settings → Privacy & Security, allow Microphone for recording, Screen & System Audio Recording for meeting audio, and Accessibility for auto-paste. You can copy transcripts without Accessibility.

Settings → App → General → Set up auto-paste… shows which app to allow. If permission breaks after an update, remove its old entry, add the app again, and restart.

When macOS asks whether OpenWhisper may use your login Keychain (to read saved API keys, the MCP token, or a remote engine's connection key), enter your Mac password and choose Always Allow. Allow works for that one read only, so macOS asks again at every launch. Builds that aren't notarized may ask once more after an update.

Run from source

You'll need Git and Python 3.11 or 3.12. Source installs support Windows, macOS (including Intel Macs), and Linux distributions without native packages.

One-time setup

Clone the repository and enter its folder:

git clone https://github.com/Knuckles92/OpenWhisper
cd OpenWhisper

Then create the virtual environment and install dependencies for your platform.

Windows (PowerShell):

python -m venv venv
. .\venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt

macOS / Linux:

python3 -m venv venv
source venv/bin/activate
python -m pip install -r requirements.txt

On Linux, install Python's venv and development packages and a C compiler if needed. The launcher checks audio and Qt system libraries and prints the package-manager command for any missing dependencies. Meeting capture also needs PulseAudio or PipeWire-Pulse.

Launch the app

Run these commands from your OpenWhisper checkout folder:

Windows (PowerShell):

. .\venv\Scripts\Activate.ps1
python main.py

macOS / Linux:

./scripts/openwhisper

The macOS/Linux launcher selects the virtual environment and handles macOS app-bundle identification for Accessibility. Then follow Get started.

Optional: launch from any terminal

To register ow and openwhisper, run .\install.cmd in PowerShell on Windows or ./install.sh on macOS/Linux once from your checkout folder. Open a new terminal, then use ow or openwhisper from any folder without activating the virtual environment. See source launchers for details.

Optional: develop with uv

Contributors can use uv sync and uv run python main.py. See development with uv for setup, tests, and GPU support.

GPU acceleration

For Local Whisper on Windows or Linux x86_64, install GPU Acceleration from Settings → Models & storage → Downloads → Components, or choose Use this GPU when prompted. Requires an NVIDIA driver providing CUDA 12 (525+); the CUDA Toolkit is not needed. macOS uses CPU.

Source installs can use python -m pip install -r requirements-gpu.txt in the activated virtual environment. Auto selects a compatible compute type; Downloads shows GPU memory estimates.

Other speech engines use their own runtimes from Downloads. On Apple Silicon Macs, Auto runs Parakeet MLX on the Apple GPU through MLX, Parakeet and Nemotron through the NVIDIA Speech GPU (Metal) runtime, and Qwen3-ASR through PyTorch MPS. The Pi and OpenCode SDK meeting agents are also separate components, available on Windows, Linux and macOS.

Updates

OpenWhisper asks GitHub for the latest release when it starts and every 4 hours while it runs, sending only the app version. A newer release shows as an Update to x.y.z button in the title bar and the tray menu until you update. To check yourself, use OpenWhisper → Check for Updates on macOS or Help → Check for Updates on Windows and Linux. Turn automatic checks off in Settings → App → General → Updates. Git checkouts don't check on their own.

• Windows: Apply the in-app update or follow the setup prompt.
• macOS: The verified update installs and restarts the app, restoring the previous version if startup fails. If the app cannot replace itself, it opens the verified DMG for replacement in Applications.
• Linux: Install the new package with the same command used for installation.
• From source: Run git pull --ff-only and reinstall requirements if dependencies changed.

When a package upgrade or pull replaces the app while it runs, OpenWhisper offers Restart now once nothing is recording or transcribing.

Get started

1. Open Settings → Dictation → Voice model and choose a speech backend. New Windows x64 and Linux x86_64 installs default to Parakeet, and Apple Silicon Macs on macOS 14 or later to Parakeet MLX; other platforms default to Local Whisper. Existing installs keep the engine they use.
2. Download the model and runtime through Settings → Models & storage → Downloads, or add an OpenAI key in Settings → App → API keys for cloud transcription.
3. Choose your microphone in Settings → Dictation → Recording. On macOS, grant the required permissions.
4. Use Quick Record or the recording hotkey. Stop recording to transcribe; dictation follows your clipboard and auto-paste settings. Upload File results stay in the app and have Copy buttons.

For meetings, open Meeting Mode. Add an optional brief to guide AI notes and the live copilot; you can edit it during the meeting. After transcription, Continue in the background lets you start another meeting while reports finish. Find results in Past Meetings.

Optional insight review helps clarify uncertain decisions and commitments. Fast judgments add citation checks, meaning-based history search, and live meeting cues; enable them under Meeting Mode → Fast judgments, with a TypeSafe key, the OpenRouter key you already use, or your own text model. With a model on this computer they send nothing, and live cues and history search work even with AI insights off.

Hotkeys

Change shortcuts and choose Toggle or Push and hold in Settings → App → Hotkeys. Toggle starts and stops recording with successive presses; push-and-hold records until you release the shortcut.

Action	Windows / Linux	macOS
Start/stop recording	Numpad *	Control+Option+R
Cancel	Numpad -	Control+Option+Escape
Enable/disable program	Ctrl+Alt+Numpad *	Control+Option+Shift+R
Minimize to tray	Ctrl+Alt+M	Control+Option+M

In push-and-hold mode, double-tap the record shortcut to keep recording hands-free, then press it again to stop. Command Mode, the Scratchpad, switching dictation language, and pasting the original of your last dictation have optional shortcuts with no default. On Windows and X11, mouse side buttons (Mouse 4 and Mouse 5) can be shortcuts too.

On X11 Linux, hotkeys also reach the focused app. Omarchy supports desktop shortcuts and auto-paste; other native Wayland desktops use focused-window shortcuts and manual paste. macOS auto-paste requires Accessibility permission; global hotkeys do not.

Omarchy / Hyprland

OpenWhisper adapts to Omarchy 3 and 4 with a compact interface and desktop colors. Hyprland manages window placement and sizing; Omarchy mode supports desktop shortcuts and auto-paste on Hyprland 0.55+.

Choose Settings → App → Appearance → Theme → Omarchy desktop to follow your desktop palette. An optional Omarchy 4 bar widget provides recording controls, status, and live text.

Speech models

Backend	Models	Platform / device	Workflows
Local Whisper	Standard Whisper sizes, turbo, Distil-Whisper	All platforms, CPU; NVIDIA CUDA on Windows/Linux	Dictation with preview, uploads, meetings
Parakeet	TDT 0.6B v3; Orukeet TDT 0.6B community adaptation	Windows x64 and Linux x86_64 CPU / NVIDIA GPU; Apple Silicon CPU / Apple GPU (Metal); Intel Mac CPU	Dictation with preview, uploads, meeting chunks
Parakeet MLX	TDT 0.6B v3, MLX Community weights	Apple Silicon, macOS 14+: Apple GPU through Metal or CPU	Dictation with preview, uploads, meeting chunks
Parakeet Unified	Unified EN 0.6B, English	Windows x64 and Linux x86_64 CPU / GPU through Vulkan (NVIDIA, AMD, Intel)	Dictation with native preview, uploads
Qwen3-ASR	0.6B, 1.7B	Windows x64 CPU / NVIDIA GPU; Apple Silicon CPU / Apple GPU (MPS)	Dictation, uploads
Nemotron Streaming	3.5 ASR Streaming 0.6B	Windows x64 and Linux x86_64 CPU / NVIDIA GPU; Apple Silicon CPU / Apple GPU (Metal); Intel Mac CPU	Dictation with preview, uploads, meetings with native preview
Moonshine	Streaming Small / Medium, English	Windows x64 CPU; Apple Silicon CPU, macOS 15+	Dictation, uploads, meetings with native preview
API	OpenAI: GPT-Transcribe; GPT-4o Transcribe, GPT-4o Mini Transcribe, and Whisper until OpenAI retires them on February 26, 2027. Groq: Whisper Large v3 / Turbo. Your own server: whatever it serves	OpenAI or Groq in the cloud (API key required), or an OpenAI-compatible speech server you run: Speaches, LocalAI, vLLM, whisper.cpp, or custom	Dictation, uploads
Remote computer	Whichever engine the paired computer has selected	Another computer running OpenWhisper on your network	Dictation with that engine's preview, uploads, meetings with a supported host model

• Downloads: Install both the model and its required runtime in Settings → Models & storage → Downloads. Each entry shows its size, publisher, license, and download checks.
• Languages: Choose a supported language or Auto for detection. Coverage varies by engine; Moonshine and Parakeet Unified are English-only. Language selection controls transcription, not translation.
• Apple Silicon: Choose Auto to use the Apple GPU where supported. See GPU acceleration for engine-specific support.

See speech model details for language coverage, platform requirements, download verification, and Orukeet's license and validation status.

Custom Whisper models

Choose Add custom models… in Settings → Models & storage → Downloads, Dictation → Voice model, or Meeting Mode → Voice & speakers.

• Local files: Choose a model folder or a parent folder containing several models. OpenWhisper lists complete model folders for you to select.
• Hugging Face: Enter an owner/model repository and, optionally, a subfolder such as ct2_int8_float16. Find on Hugging Face lists compatible model folders without downloading weights. You can also discover models already in the local Hugging Face cache.

Click Add selected, then choose a model in Voice model or Voice & speakers to load it.

• Format: CTranslate2 Whisper folders with model.bin, a valid config.json, and tokenizer.json. Convert original PyTorch weights before use.
• Storage: Downloads follow your Hugging Face policy; cached models load offline. Removing a custom entry keeps its source files.

Remote engine

Use another computer's OpenWhisper engine for dictation, uploads, or meetings:

1. On the computer with the engine, open Settings → Dictation → Remote engine, choose Share this computer, and turn on Share this computer's engine. Allow OpenWhisper through the firewall if Windows asks.
2. On the other computer, open the same page, choose Use another computer, find the host on your network or tailnet, and click Connect. Compare the six-digit number on both screens, then click Allow on the host only if they match. If the host isn't listed, use Enter an address on the client and Show a pairing code on the host.
3. Click Use for dictation, or choose Remote computer as the recording engine.

View → Host Mode turns a computer that mostly serves others into a host dashboard: sharing status and address, the engine served, who's connected and what they asked for.

The host's selected engine handles transcription. Traffic is encrypted and limited to paired computers; remove access under Paired computers.

If macOS Keychain refuses to save the connection key, click Retry Keychain access and choose Allow or Always Allow when macOS asks. If no prompt appears, use Open Keychain Access, unlock the login keychain and check the OpenWhisper item's Access Control, then retry. The approved connection stays in memory while OpenWhisper is running, so retrying does not need another pairing code or approval from the host. The connection key is only saved in the OS credential store.

• Meetings: Select Remote computer in Settings → Meeting Mode → Voice & speakers → Speech engine. Audio goes to the host for transcription; capture, recordings, and transcripts stay on the computer running the meeting.
• Host models: Use Manage host models to choose or download a model and install speech runtimes. The host must first enable Allow paired computers to manage models.
• Shared history: Enable Allow the paired host to query this computer's history on the client so agents on the host can search it while the client is online. See MCP setup for queries and offline copies.
• Away from home: Install Tailscale on both computers. Hosts on your tailnet appear in the app; computers on the same Tailscale account can connect without a pairing code when the host allows it.

AI cleanup and meeting intelligence

Use AI cleanup to polish dictation and meeting intelligence to generate notes and insights. Each has its own text-model settings.

• Dictation: Enable Settings → Dictation → AI cleanup, choose a cleanup level, and choose who runs it: an installed Claude Code, Codex, or OpenCode with its own sign-in, or a text model below. The Thinking level lists only the levels the chosen model accepts. Add words to the Dictionary, pick a tone per kind of app in Styles, teach other behaviour in Learned rules, or use Profiles for reusable formats. See Personalize dictation.
• Providers: OpenAI, OpenRouter, Ollama, LM Studio, vLLM, llama.cpp, LocalAI, Groq, OpenCode Go/Zen, and custom OpenAI-compatible endpoints. Add keys in Settings → App → API keys; the local servers are preconfigured with their standard local addresses and need no key — just run the server. Edit changes its address or sets an optional API key variable. The picker marks local servers that are running, reads each model's context and thinking support from the server, and lists only models that generate text.
• OpenAI cleanup: The suggested model is gpt-6-luna. Saved OpenAI cleanup assignments using the old gpt-4o-mini default resolve to Luna; other models, providers, and saved meeting assignments are kept. Thinking Off explicitly requests reasoning.effort: none through the Responses API. Cleanup remains off by default, and fresh text-model assignments still use OpenRouter Free.
• Meetings: Choose an engine in Settings → Meeting Mode → Intelligence using the options below.

Meeting engine	Setup
Pi (default)	Uses your selected text endpoint and API key. Install from Models & storage → Downloads → Components where available.
Claude Code, Codex, or OpenCode	Select an installed agent's tile to use its sign-in, providers, and models. Install, update, and sign in to the agent yourself.
OpenCode SDK	Install from Models & storage → Downloads → Components where available, then select it under Agent core.

Meeting passes count toward the installed agent's account or API usage. These agents refresh live insights less often. If an engine is unavailable, select another or continue recording without AI insights.

See AI and meeting setup for custom endpoints, environment variables, agent behavior, and compatibility with older settings.

Offline use

Downloaded speech models load from the local cache without network metadata checks. Install any required runtime before going offline.

Choose whether to ask, allow, or block missing-model downloads in Settings → Models & storage → Downloads → When a model is missing from this computer. Set HF_HUB_OFFLINE=1 before launch to block model downloads entirely. Cloud transcription and remote text providers still need a network connection; speech and text servers on this computer do not. See Network and privacy for every connection the app can make. Only the update check (at launch and every 4 hours) is on by default, and there is no telemetry.

Contributing and license

See CONTRIBUTING.md for development setup, validation, pull requests, and adding a provider or engine. Release history lives in CHANGELOG.md. Report vulnerabilities privately as described in SECURITY.md. Participation follows the code of conduct.

OpenWhisper builds on OpenAI Whisper, faster-whisper, NVIDIA NeMo-Speech.cpp, Qwen3-ASR, and Moonshine, with converted Whisper weights from Systran and Mobius Labs. The app source is MIT licensed; model weights and dependencies retain their own licenses. See Third-party notices.