Wake Word

Voice Activation for Code Editors

Meta
Registries

Say a wake phrase and the right AI assistant opens -- no clicking required.

Say "Hey Claude" and Claude opens. Say "Hey Copilot" and Copilot opens. Say "Hey Computer" and the terminal focuses. The extension handles the routing, pauses its own mic so the assistant can use it, then resumes listening when the voice session ends.

Zero config. No API keys. No accounts. No system dependencies. Install and go.

All audio processing happens locally on your machine. Nothing is recorded or transmitted.

https://github.com/user-attachments/assets/fb007095-5c4c-4927-aaa6-fa76550d7cb2

How It Works

1. Install the extension. On Windows, that's all. On macOS/Linux, a local speech model (~17MB) is downloaded on first use and cached.
2. When your editor opens, the extension starts listening on your microphone
3. Audio is processed locally -- through voice activity detection and keyword spotting on macOS/Linux, or the built-in recogniser on Windows -- and matched against your configured wake phrases
4. If a phrase is detected with sufficient confidence, the extension releases the mic and fires the mapped command
5. The target assistant (Claude, Copilot, etc.) takes over the microphone with no contention
6. Wake word listening resumes after a configurable cooldown, or, for routes set to manual handoff, when you click the status bar

Installation

VS Code

Install directly from the VS Code Marketplace.

Or from the command line (VS Code only):

code --install-extension analytics-in-motion.wake-word

Cursor, Windsurf, and other VS Code forks

Option A: Open VSX Registry -- Search for "Wake Word" in the extensions panel.

Option B: Manual .vsix install -- Download the latest .vsix from GitHub Releases. Open Extensions, click the three-dot menu, select "Install from VSIX".

Prerequisites

On macOS and Linux a local speech model (~17MB) is downloaded on first use and cached. If Node.js is installed via nvm or fnm and not on VS Code's PATH, set wakeWord.nodePath to the full path of your node executable.

First Run Consent

The first time the extension tries to listen, a modal dialog explains exactly what happens: continuous microphone use, fully local processing. You must click "Allow Microphone Listening" to proceed.

If you decline, listening does not start. You can enable it any time via the status bar or command palette, which will re-prompt for consent. Reset the consent prompt with Wake Word: Reset Microphone Consent.

Wake Phrase Routing

The core feature. Each spoken phrase maps to a VS Code command.

Default routes

These work out of the box with no configuration:

The Claude route uses manual handoff (see Handoff mode): after it fires, listening stays paused until you click Wake: Paused in the status bar. The other two resume after the cooldown.

Custom routes

Add your own phrases in settings.json. Any spoken English phrase works:

{
  "wakeWord.routes": [
    {
      "label": "Copilot",
      "phrase": "hey copilot",
      "command": "workbench.action.chat.open"
    },
    {
      "label": "Claude",
      "phrase": "hey claude",
      "command": "claude-vscode.focus"
    },
    {
      "label": "Search",
      "phrase": "search files",
      "command": "workbench.action.quickOpen"
    },
    {
      "label": "Commands",
      "phrase": "open commands",
      "command": "workbench.action.showCommands"
    }
  ]
}

The speech engine uses a constrained grammar built from your configured phrases for accurate matching.

Phrase aliases

The phrase field accepts a string or an array of strings. Use arrays to map multiple trigger phrases to the same command:

{
  "label": "Claude",
  "phrase": ["hey claude", "open claude"],
  "command": "claude-vscode.focus"
}

Per-route cooldown

Override the global cooldown for individual routes with cooldownSeconds:

{
  "label": "Terminal",
  "phrase": "hey computer",
  "command": "workbench.action.terminal.focus",
  "cooldownSeconds": 10
}

Handoff mode

handoff chooses how listening comes back after a route fires:

• "timer" (default): listening resumes after cooldownSeconds.
• "manual": listening stays paused until you click the status bar or run Wake Word: Enable Listening. Use this for assistants whose voice sessions run longer than the cooldown, so Wake Word does not restart under them and compete for the microphone.

{
  "label": "Claude",
  "phrase": "hey claude",
  "command": "claude-vscode.focus",
  "handoff": "manual"
}

The status bar shows Wake: Paused while a manual route waits. The default Claude route uses manual handoff.

The handoff

When a wake phrase is detected:

1. The extension asks the speech engine to close the microphone and waits for it to confirm, then kills the process (forcing it after 500 ms if no confirmation arrives)
2. The target VS Code command fires (opening the assistant)
3. The assistant's voice mode takes over the microphone with no contention
4. After wakeWord.cooldownSeconds (default: 30), wake word listening resumes. A route with handoff: "manual" waits for you instead
5. Status bar shows a live countdown (Wake: 30s → Wake: 29s → ...) during handoff, then returns to "Wake: Listening". A manual route shows "Wake: Paused" instead, with no countdown

This ensures only one thing uses the mic at a time.

Multiple windows

If you have more than one editor window open, only one listens at a time. The first window to start listening takes a lock; the others show Wake: Other window in the status bar and stand by. When the listening window closes, crashes, or has listening disabled, a standing-by window takes over within about ten seconds. The lock is held through the cooldown after a detection, so a second window never opens the microphone while an assistant has it.

The lock lives in the extension's global storage, which windows of the same editor share. Windows of different editor products do not see each other's lock.

Settings

Choosing a microphone

wakeWord.audioDevice selects the input device for the sherpa engine. Use any case-insensitive substring of the device name as it appears in your system sound settings, or the device's index number:

{
  "wakeWord.audioDevice": "Blue Yeti"
}

Changing it restarts the engine. If the value matches no device, or more than one, the error notification says so and names the value. The Windows engine (System.Speech) always uses the system default input device, so set wakeWord.engine to sherpa to choose a device on Windows.

Calibrating

Wake Word: Calibrate listens for 15 seconds and logs every wake phrase it hears, with the time and, on the Windows engine, the confidence score, without firing any route. Say each of your phrases a few times at your normal distance, then open the Wake Word output channel: the summary groups detections by phrase with the count, the average confidence, and the minimum. No detections means the microphone, the room, or the threshold needs attention; a minimum close to wakeWord.confidenceThreshold means that phrase is on the edge. The sherpa engine reports no score, so its summary has counts only. Whatever the extension was doing before, listening, a cooldown, or a manual pause, is put back afterwards.

Commands

• Wake Word: Enable Listening -- start the detector
• Wake Word: Disable Listening -- stop the detector
• Wake Word: Toggle Listening -- toggle on/off (also via status bar click)
• Wake Word: Reset Microphone Consent -- clear consent and re-prompt
• Wake Word: Open Settings -- open the Settings editor filtered to Wake Word (also linked from the status bar tooltip)
• Wake Word: Calibrate -- listen for 15 seconds and log what is heard without firing any route (see Calibrating)

Common command IDs

Useful values for the command field in your routes. Command IDs listed are for VS Code. Cursor and other editors may use different IDs for the same features.

How It Works (Technical)

The extension selects a speech engine based on platform (or the wakeWord.engine setting) and spawns it as a background child process. Both engines communicate via stdout using the same protocol: READY, DETECTED:<phrase>|<confidence>, ERROR:<message>, DEBUG:<info>. The sherpa engine also sends RELEASED once it has closed the microphone.

Windows engine (default on Windows)

Spawns a PowerShell process using System.Speech.Recognition.SpeechRecognitionEngine, the same engine built into Windows. A constrained grammar is built from your configured phrases and passed via encoded command. Zero model downloads; the speech engine ships with Windows.

Sherpa engine (default on macOS/Linux, optional on Windows)

Spawns audio-engine.js under system Node.js (not Electron). The child process uses decibri for mic capture and sherpa-onnx for keyword spotting. Running under system Node.js is required because Electron's Node.js runtime cannot load native audio addons. A local speech model (~17MB) is downloaded to VS Code's global storage on first use and cached.

Captured audio passes through voice activity detection (Silero VAD) before it reaches the keyword spotter, so the spotter only runs while someone is speaking and an idle editor does not decode silence. decibri also conditions the signal on the way through: DC offset removal, an 80 Hz high-pass to drop rumble below the voice band, and automatic gain control targeting -18 dBFS so the confidence threshold sees a consistent level.

Shared flow

1. Extension builds phrase list and spawns the engine process
2. Engine writes READY when the mic is open
3. Each detection above the confidence threshold is written to stdout as DETECTED:<phrase>|<confidence>
4. The extension reads stdout and fires the corresponding command
5. On handoff, the engine process is killed to release the microphone
6. After the cooldown countdown, a new engine process starts

Zero runtime npm dependencies in the extension host. All native dependencies are isolated in the engine/ child process.

Troubleshooting

Privacy

All speech recognition runs locally on your machine. No audio data ever leaves your device. On Windows, speech is processed by the built-in System.Speech.Recognition engine in memory. On macOS and Linux, a local sherpa-onnx model processes audio in the engine child process. Nothing is recorded, stored, or transmitted.

Platform Support

Compatibility

License

Apache 2.0