Quickstart

Install PipeVoice, pick an engine, and start talking into any app on Windows — in about a minute.

Install Engines & keys Hotkeys & modes Privacy Troubleshooting

1 · Install

Download the installer and run it. PipeVoice opens quietly to your system tray (look for the mic icon).
If Windows shows a blue "unrecognised app" SmartScreen warning, click More info → Run anyway. It appears for any app without a paid code-signing certificate — PipeVoice is open source, so you can read every line first.

2 · Pick an engine

On first run, PipeVoice asks how you want to transcribe. You can change this any time from the tray menu.

Engine	Needs	Best for
Local (offline)	Nothing — runs on your PC	Privacy & zero cost. First use downloads a small model (~150 MB).
OpenAI Whisper	`OPENAI_API_KEY`	Best accuracy. ~$0.006/min, billed by OpenAI.
Deepgram	`DEEPGRAM_API_KEY`	Lowest latency — words appear live as you speak.

For the cloud engines, paste your key into the first-run dialog — it's stored locally on your PC and never uploaded. No key? Choose the offline engine and you're running, free.

Where to get a key: OpenAI → platform.openai.com/api-keys · Deepgram → console.deepgram.com. You pay the provider directly, usually a few cents a day.

3 · Hold & talk

By default, hold Right Ctrl, speak, and release — your words type in wherever the cursor is. A corner pill shows a live mic meter while it listens.

Modes

Push-to-talk (default) — hold the key while speaking, release to type.
Toggle — tap once to start, tap again to stop. Good for longer dictation.

Change the hotkey or mode

Right-click the tray icon → Settings. You can rebind the hotkey (with a "Capture" button), switch mode, pick your microphone, choose type-vs-paste output, set a language hint, and add vocabulary for your jargon. Changes apply live.

Privacy

PipeVoice has no servers of its own and no telemetry. What leaves your computer depends only on the engine you choose:

Local engine — nothing leaves your PC. Audio is transcribed entirely on-device.
Cloud engines — audio is sent to OpenAI or Deepgram to transcribe, then discarded. Your API key stays on your machine.

It's open source — audit it on GitHub.

Troubleshooting

My words aren't appearing in an app

If the target app runs as administrator (some terminals do), Windows blocks keystrokes from non-elevated apps. Run PipeVoice as administrator too, and they'll match.

First offline transcription is slow

The local engine downloads its model (~150 MB) the first time you use it. After that it's cached and fast. Bigger models are more accurate but slower — a GPU helps a lot.

Wrong microphone

Settings → Microphone. Leave it blank to use the system default, or pick one by name/index.

Nothing happens when I hold the key

Confirm the hotkey in Settings, and that an engine is selected (the tray tooltip shows the current engine). For cloud engines, make sure a valid key is saved.

Talk to your AI agent (MCP)

PipeVoice can act as a local MCP server so agents (Claude Code, Cursor, Cline) can use your voice — no API key, everything runs on your machine.

listen — the agent asks a question, you answer by voice, and it gets the text back.
transcribe — hand it a local audio or video file and get a transcript with timestamps, or ready-made captions (format: "srt" / "vtt").

Enable Agent MCP (listen + transcribe) in the tray menu, then register it once with your client:

claude mcp add pipevoice -- python -m wisprlite --mcp

The server is loopback-only and off by default.