Hold a hotkey. Speak. Release. The transcript pastes itself at your cursor.

Local Whisper dictation for Windows and Linux. No account, no cloud, no monthly bill.
_{macOS builds and runs but isn't officially supported yet.}

_{Latest: v1.6.0 · all releases}

New in v1.6.0 — Meetings: long-form recording with mic + system audio, local transcription, and bring-your-own-LLM summaries.

VoiceFlow lives in your system tray. Hold a global hotkey, a small popup pops up with a live amplitude meter, you talk, you release, and the transcript is typed at the cursor. That's it.

The inference runs on your machine through faster-whisper. CUDA when you have it, CPU when you don't. The audio never touches a network socket.

Long-form recording that captures your mic plus system audio (Zoom, Meet, anything that plays through your speakers) into one stereo file, transcribes it locally with Whisper, and runs the summary through an LLM provider you choose.

• System audio + mic in one file. Stereo capture via WASAPI loopback (Windows) and PipeWire/PulseAudio (Linux).
• Pause, resume, stop from the dashboard or the tray menu — recording survives across hour-long calls.
• Re-transcribe any saved recording with a different model, device, or language without re-recording.
• Bring your own LLM. OpenAI, Groq, OpenRouter, Ollama, or any OpenAI-compatible endpoint. Keys live in your OS keychain.
• Auto-rename from a default timestamp to a real topic once the transcript lands.
• Export to Markdown, plain text, SRT, or structured JSON.
• Built-in playback via the voiceflow:// URL scheme — jump straight from any transcript line into the audio.

Recording, transcription, search, and storage stay local. The only network call is the optional summary request — skip it, point it at a local Ollama, or send it to a provider you already pay for.

• Fully local. Audio stays in RAM. No telemetry, no analytics, no phone-home.
• 16+ Whisper models. Tiny (75 MB) through Large-v3 (3 GB), plus Turbo, distilled, and .en variants. The picker shows speed, accuracy, parameter count, and disk size for each.
• CUDA when available. Auto-detects your GPU, falls back to CPU.
• Hold or Toggle modes. Configurable hotkeys including modifier-only combos like Ctrl+Win.
• Wayland and X11. Native evdev input on Linux, Hyprland window rules, wl-copy and wtype/ydotool for paste.
• 99+ languages. Whisper handles language detection automatically.
• Searchable history. SQLite log of every transcript, stored at ~/.VoiceFlow/.
• Dark mode by default. Light and system themes if you want them.

Grab the latest binary from Releases — currently v1.6.0:

• Windows 10/11: .exe installer (Inno Setup)
• Linux: .AppImage or .tar.gz

64-bit only. First launch walks you through a seven-step setup: microphone, compute device, Whisper model download, hotkey. If you delete the model later, a recovery dialog lets you re-download or pick a different one.

git clone https://github.com/infiniV/VoiceFlow.git
cd VoiceFlow
pnpm run setup        # installs Node and Python deps
pnpm run dev          # Vite frontend + Pyloid backend

Platform installers (run on the matching OS):

pnpm run build:installer          # Windows (.exe via Inno Setup)
pnpm run build:installer:linux    # Linux (.AppImage and .tar.gz)
pnpm run build:installer:macos    # macOS (.dmg, unsupported)

MIT. See LICENSE.

Releases · Issues · Website