VS Whisper

Voice-to-text dictation for VS Code. Free, local, fast and private.

Dictate code, comments, commit messages, and prompts — right inside your editor. Powered by whisper.cpp, runs entirely on your machine. No API key required.

Think superwhisper.com, but open source and built into VS Code.

Demo

Cmd+Shift+; → speak → Cmd+Shift+; → text appears at cursor

[Recording...] → [Transcribing...] → "function that returns the sum of two numbers"

Features

• Push-to-talk — toggle recording with Cmd+Shift+; (macOS) / Ctrl+Shift+;
• 100% free & local — whisper.cpp runs on your machine, no cloud, no API key
• 3 dictation modes:
  • Dictate — raw transcription, inserted as-is
  • Code — strips filler words, converts spoken punctuation ("open paren" → (, "arrow" → =>)
  • Command — cleans and formats speech as a coding instruction
• Cloud backends — optional OpenAI Whisper API and Groq support for those who prefer cloud
• Status bar — shows recording state, current mode, transcribing spinner
• Multi-language — supports 100+ languages via Whisper
• Clipboard fallback — copies to clipboard when no editor is focused

Quick Start

1. Install prerequisites

You need an audio recording tool:

# macOS
brew install sox

# Linux
sudo apt install sox

# Windows
# Install ffmpeg: https://ffmpeg.org/download.html

2. Install the extension

# From source
git clone https://github.com/roussi/vs-whisper.git
cd vs-whisper
npm install && npm run compile
npx @vscode/vsce package
code --install-extension vs-whisper-0.1.0.vsix

3. Setup local transcription

On first launch, VS Whisper will prompt you to download whisper.cpp and a model. Or run manually:

Cmd+Shift+P → "VS Whisper: Setup Local Transcription"

This will:

1. Clone and compile whisper.cpp
2. Download a Whisper model from Hugging Face

Build requirements: git, cmake, make, C++ compiler

# macOS
xcode-select --install
brew install cmake

# Linux (Debian/Ubuntu)
sudo apt install build-essential cmake git

4. Start dictating

Press Cmd+Shift+; to start recording, speak, press again to stop. Done.

Backends

Using OpenAI or Groq

Cmd+Shift+P → "VS Whisper: Select Transcription Backend"

Then set your API key in VS Code settings:

• vsWhisper.openaiApiKey — your OpenAI key
• vsWhisper.groqApiKey — your Groq key

Dictation Modes

Switch modes via Cmd+Shift+P → "VS Whisper: Set Dictation Mode" or the status bar.

Dictate (default)

Raw transcription. What you say is what you get.

Code

Optimized for coding. Removes filler words and converts spoken punctuation:

Command

Formats your speech as a clean coding instruction — useful for AI-assisted workflows (Copilot, Cursor, Claude, etc.):

"um so basically I want to refactor the user service to uh use dependency injection"

→ Refactor the user service to use dependency injection.

Configuration

Model sizes

Commands

Architecture

src/
├── extension.ts      # Commands, status bar, text insertion
├── recorder.ts       # Cross-platform audio capture (sox/ffmpeg)
├── transcriber.ts    # OpenAI, Groq, and local whisper.cpp backends
├── modes.ts          # Post-processing: dictate, code, command
└── localSetup.ts     # Auto-download and compile whisper.cpp + models

Contributing

Contributions are welcome! Some ideas:

• [ ] Hold-to-talk mode — record while holding the shortcut, stop on release
• [ ] Custom vocabulary — user-defined word mappings for technical terms
• [ ] Audio level indicator — visual feedback in status bar while recording
• [ ] Inline suggestions — show transcription as ghost text before confirming
• [ ] Multi-cursor dictation — insert at multiple cursors
• [ ] Streaming transcription — show partial results while still recording
• [ ] Windows testing — verify ffmpeg/dshow recording on Windows

Development

git clone https://github.com/roussi/vs-whisper.git
cd vs-whisper
npm install
npm run watch    # compile on change
# Press F5 in VS Code to launch Extension Development Host

License

MIT

Acknowledgments

• whisper.cpp by Georgi Gerganov — the engine behind local transcription
• OpenAI Whisper — the model that makes this possible
• superwhisper — the inspiration for this project