Voice Typing — VS Code / Cursor Extension

Speak into your microphone, get a clean coding prompt inserted into your editor. Everything runs locally — no cloud services required.

Flow: Mic → whisper.cpp (local STT) → Ollama (local LLM rewrite) → Insert into editor

Platforms: macOS · Linux · Windows — works in both VS Code and Cursor

Install

From Marketplace: Voice Typing on VS Code Marketplace

Or from VSIX: Download the latest .vsix from GitHub Releases, then in VS Code/Cursor: Cmd+Shift+P → "Install from VSIX"

---

Quick Start (macOS)

brew install whisper-cpp sox       # STT engine + audio recorder
brew install ollama                # optional: LLM rewrite
ollama pull llama3.2:3b            # optional: pull rewrite model

Install the extension from the VS Code Marketplace, or build from source:

git clone https://github.com/bread22/voice-typing.git
cd voice-typing && npm install && npm run package
# Then: Cmd+Shift+P → "Install from VSIX" → select voice-prompt-*.vsix

Press Option+V to start recording. The whisper model (~75 MB) downloads automatically on first use.

---

Cross-Platform Installation

Step 1: Install whisper-cpp (speech-to-text engine)

macOS

brew install whisper-cpp

Installs the whisper-cli binary to your PATH. Verify with:

whisper-cli --help

Linux (Ubuntu / Debian)

Option A — Build from source (recommended):

sudo apt install build-essential libsdl2-dev
git clone https://github.com/ggerganov/whisper.cpp.git
cd whisper.cpp
cmake -B build
cmake --build build --config Release
sudo cp build/bin/whisper-cli /usr/local/bin/

Option B — Download pre-built binary:

Download from whisper.cpp releases, extract, and place whisper-cli somewhere in your PATH.

Verify with:

whisper-cli --help

Windows

Option A — Build with CMake + Visual Studio:

git clone https://github.com/ggerganov/whisper.cpp.git
cd whisper.cpp
cmake -B build
cmake --build build --config Release

The binary will be at build\bin\Release\whisper-cli.exe. Add it to your PATH or set voicePrompt.stt.whisperCppPath in settings.

Option B — Download pre-built binary:

Download from whisper.cpp releases. Place whisper-cli.exe in a directory on your PATH.

Verify with:

whisper-cli.exe --help

Step 2: Install an audio recorder

The extension auto-detects the best available recorder for your platform:

Platform	Recommended	Install
macOS	SoX (auto-detected first)	brew install sox
Linux	arecord (usually built-in)	sudo apt install alsa-utils (if not present)
Windows	FFmpeg	winget install ffmpeg or ffmpeg.org

Detection priority:

Platform	Tries in order
macOS	SoX → FFmpeg
Linux	arecord → SoX → FFmpeg
Windows	FFmpeg → SoX

If no recorder is found, the extension shows the appropriate install command.

Step 3: Install Ollama (optional — LLM rewrite)

Ollama rewrites your raw voice transcript into a clean, structured prompt. Without it, the raw transcript is inserted directly (still useful!).

macOS

brew install ollama
ollama pull llama3.2:3b

Ollama starts automatically via Homebrew services. Verify: curl http://127.0.0.1:11434

Linux

curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3.2:3b

Start the server: ollama serve (or configure as a systemd service).

Windows

Download from ollama.com and install. Then:

ollama pull llama3.2:3b

Step 4: Install the extension

git clone https://github.com/bread22/voice-typing.git
cd voice-typing
npm install
npm run package

This creates a voice-prompt-*.vsix file. Install it:

• VS Code: Ctrl+Shift+P → "Extensions: Install from VSIX..." → select the file
• Cursor: Cmd+Shift+P (Mac) / Ctrl+Shift+P (Win/Linux) → "Extensions: Install from VSIX..." → select the file

Reload the editor when prompted.

Step 5: First run

1. Press Alt+V (or Option+V on Mac)
2. On first use, the whisper model (ggml-tiny.en.bin, ~75 MB) downloads automatically
3. Grant microphone access if prompted by your OS
4. Speak naturally — the extension auto-stops when you pause

---

Usage

Start recording

Method	Shortcut / Action
Keyboard	Alt+V (Windows/Linux) · Option+V (Mac)
Command palette	Ctrl+Shift+P → Voice Prompt: Start Recording
Status bar	Click the $(mic) Voice Prompt button (bottom bar)

Recording flow

1. Press shortcut → status bar shows "Recording..."
2. Speak naturally — e.g. "add a function that validates email addresses using regex"
3. Pause speaking → silence detection (VAD) auto-stops recording
4. Status bar progress: Listening → Transcribing → Rewriting → Done
5. Result inserted at your cursor position (cursor moves to end, space appended)

Example

You say:

"um so I need a function that uh validates email addresses and it should use regex and return true or false"

Gets inserted as:

Add a function that validates email addresses using regex. It should return true for valid emails and false otherwise.

Without Ollama (no rewrite)

If Ollama isn't running, the raw transcript is inserted directly. You can also disable rewrite explicitly by setting voicePrompt.rewrite.provider to none.

---

Settings Reference

Open settings (Ctrl+,) and search for voicePrompt.

Speech-to-Text

Setting	Default	Description
voicePrompt.stt.provider	whisper-cpp	whisper-cpp (local CLI) or http (custom server)
voicePrompt.stt.model	tiny.en	Model size: tiny.en, base.en, small.en, etc.
voicePrompt.stt.whisperCppPath	(auto-detect)	Custom path to whisper-cli binary
voicePrompt.stt.modelPath	(auto-download)	Custom path to GGML model file
voicePrompt.stt.httpEndpoint	http://127.0.0.1:8765/transcribe	HTTP endpoint (when provider is http)
voicePrompt.stt.timeoutMs	30000	STT timeout in milliseconds
voicePrompt.stt.language	en	Language code for speech recognition

Model sizes (auto-downloaded):

Model	Size	Speed	Accuracy
tiny.en	~75 MB	Fastest	Good for commands and short prompts
base.en	~142 MB	Fast	Better accuracy
small.en	~466 MB	Moderate	Best accuracy for English

Rewrite (LLM)

Setting	Default	Description
voicePrompt.rewrite.provider	ollama	ollama, cloud, or none
voicePrompt.rewrite.model	llama3.2:3b	LLM model for rewriting
voicePrompt.rewrite.ollamaBaseUrl	http://127.0.0.1:11434	Ollama server URL
voicePrompt.rewrite.cloudBaseUrl	https://api.openai.com/v1/chat/completions	Cloud API endpoint (OpenAI-compatible)
voicePrompt.rewrite.timeoutMs	20000	Rewrite timeout in milliseconds
voicePrompt.rewrite.style	engineering	concise, detailed, engineering, debugging

Behavior

Setting	Default	Description
voicePrompt.previewBeforeInsert	false	Show editable preview before inserting
voicePrompt.autoFallbackToCloud	false	Auto-fallback to cloud if local rewrite fails
voicePrompt.noRewriteBehavior	stt_passthrough	stt_passthrough = insert raw text · disable_plugin = block insertion
voicePrompt.showStatusBarButton	true	Show mic button in status bar
voicePrompt.insertTrailingSpace	true	Append space after each insertion for easier consecutive inputs

Voice Activity Detection (VAD)

Setting	Default	Range	Description
voicePrompt.vad.enabled	true	—	Auto-stop recording after silence
voicePrompt.vad.silenceMs	1500	600–3000 ms	Silence window before auto-stop
voicePrompt.vad.minSpeechMs	300	100–1000 ms	Minimum speech duration to accept

VAD tuning tips:

• Getting cut off mid-sentence? Increase silenceMs by 100–200 ms
• Feels slow to respond? Decrease silenceMs by 100 ms
• Ambient noise triggering? Increase minSpeechMs to 400–500 ms

---

Cloud Rewrite (Optional)

Use OpenAI or another cloud LLM instead of (or as fallback to) local Ollama:

1. Set API key: Ctrl+Shift+P → Voice Prompt: Set Cloud API Key (stored securely in VS Code SecretStorage)
2. Switch provider: Set voicePrompt.rewrite.provider to cloud
3. Or use as fallback: Keep ollama as provider, enable voicePrompt.autoFallbackToCloud

---

Troubleshooting

Problem	Solution
"whisper-cpp not found"	Install whisper-cpp (see platform instructions above)
"No audio recorder found"	Install SoX, arecord, or FFmpeg (see Step 2 above)
"Transcription failed"	Check that whisper-cli works: whisper-cli -m <model> -f test.wav
Model download fails	Set voicePrompt.stt.modelPath to a manually downloaded .bin file
Ollama not rewriting	Check ollama ps — model should be loaded. Try ollama run llama3.2:3b
Cursor position jumps on insert	Ensure voicePrompt.insertTrailingSpace is true (default)
VAD cuts off too early	Increase voicePrompt.vad.silenceMs (try 2000–2500)
Extension not responding to hotkey	Reload window: Ctrl+Shift+P → "Developer: Reload Window"
Microphone permission denied (macOS)	System Preferences → Privacy & Security → Microphone → allow VS Code/Cursor

---

Architecture

┌─────────────┐    ┌──────────┐    ┌─────────────┐    ┌───────────┐    ┌─────────┐
│  Microphone  │ →  │ Platform │ →  │ whisper-cpp │ →  │  Ollama   │ →  │ Editor  │
│  (Alt+V)     │    │ Recorder │    │ (local STT) │    │ (rewrite) │    │ Insert  │
└─────────────┘    │ WAV file │    │ Raw text    │    │ Clean text│    │ @ cursor│
                   └──────────┘    └─────────────┘    └───────────┘    └─────────┘

All processing happens locally on your machine. No audio or text is sent to any cloud service unless you explicitly configure a cloud rewrite provider.

Project Structure

src/
  extension.ts           — command registration and activation
  audio/
    audioCaptureService  — cross-platform mic capture (SoX / arecord / FFmpeg) + VAD
  stt/
    whisperCppSttProvider — shells out to whisper-cli binary
    httpSttProvider        — HTTP STT for custom servers (opt-in)
    modelManager           — auto-downloads GGML models from HuggingFace
  rewrite/
    ollamaRewriteProvider  — local Ollama LLM rewrite
    cloudRewriteProvider   — cloud LLM rewrite (OpenAI-compatible)
  inject/
    cursorInputInjector    — inserts text at cursor, manages position
  config/
    settings               — reads VS Code configuration
    secrets                — manages API keys via SecretStorage
  types/
    contracts              — shared TypeScript interfaces
  orchestration/
    voicePromptOrchestrator — pipeline: record → transcribe → rewrite → inject

---

Development

npm install          # install dependencies
npm run build        # compile TypeScript
npm run bundle:dev   # esbuild development bundle (with source maps)
npm run bundle       # esbuild production bundle (minified)
npm run package      # create .vsix for distribution
npm run dev          # build + package dev vsix
npm run watch        # watch mode for development

Press F5 in VS Code/Cursor to launch the Extension Development Host with the debugger attached.

---

License

MIT