Voice Transcriber

A VS Code extension that records your voice and transcribes it using OpenAI Whisper, OpenRouter, or a local Whisper-compatible API. Can optionally clean up the text with an LLM.

Features

• Record audio directly in VS Code with real-time visualization
• Upload audio or video files — audio track is extracted automatically (MP4, MKV, MOV, AVI, WebM)
• Transcribe long recordings (1hr+) — split into 10-min chunks behind the scenes
• Transcribe via OpenAI, OpenRouter (Whisper, Deepgram Nova-3, GPT-4o Transcribe, Voxtral), or your own local server
• Separate API key per provider — switching providers doesn't wipe the other key
• Copy the transcript with per-phrase timestamps as SRT, WebVTT, or plain [hh:mm:ss] lines
• Clean up filler words and fix punctuation with LLM (optional)
• Keep your last 10 transcriptions
• Auto-copy results to clipboard
• Recover recordings if VS Code crashes

ffmpeg Installation (required for native recording and long files)

The extension uses ffmpeg for native recording, splitting long files, and extracting audio from video uploads. Without ffmpeg, recording falls back to the browser (works, but lower quality and no long-file support).

macOS:

brew install ffmpeg

Linux:

sudo apt install ffmpeg

Windows:

winget install ffmpeg
# or
choco install ffmpeg

Usage

1. Click the microphone icon in the top-right of your editor
2. Set up your provider (OpenAI, OpenRouter, or local)
3. Hit "Start Recording" and speak
4. Hit "Stop" — text is automatically copied to clipboard

Configuration

OpenAI

Get an API key from platform.openai.com/api-keys, select "OpenAI" as provider, paste your key, and save.

OpenRouter

Get an API key from openrouter.ai/keys, select "OpenRouter" as provider, paste your key, and save. One key covers both transcription and LLM cleanup, and you can pick the speech-to-text model under Advanced Settings (Whisper 1, Whisper Large v3 Turbo, GPT-4o Transcribe, Deepgram Nova-3, Voxtral Mini).

Local server

Any Whisper-compatible API works:

• faster-whisper-server
• whisper.cpp server
• Anything with a /v1/audio/transcriptions endpoint

Just enter the URL, e.g. http://localhost:8000/v1/audio/transcriptions.

LLM text cleanup

With OpenAI or OpenRouter you can enable "Clean up text with LLM" to remove filler words, fix punctuation, and add paragraph breaks. Not available for local servers, which have no chat endpoint.

OpenAI models: gpt-4.1-nano (default, cheapest), gpt-4.1-mini, gpt-4.1. OpenRouter models: gpt-4.1-nano (default), gpt-4.1-mini, gemini-2.5-flash-lite, claude-haiku-4.5, llama-3.3-70b.

Timestamps and subtitles

After a transcription finishes, a SRT button appears next to the copy button in the Transcription header. It offers three formats:

• SubRip (SRT) — the subtitle standard, read by YouTube, Premiere, DaVinci, VLC
• WebVTT — the web subtitle standard, for <track> elements and HTML5 players
• Plain text with timestamps — one [hh:mm:ss] phrase line, for notes and chat

The plain copy button is unchanged and still copies just the text, and transcriptions still auto-copy as plain text.

Timings are stitched across chunks, so an hour-long recording gives one continuous timeline. Not every model can produce them — the model picker marks those that cannot:

With LLM cleanup enabled the timestamps come from the raw transcript, because cleanup rewrites and merges sentences and the alignment no longer holds.

Languages

Auto-detect or pick manually: English, Russian, Ukrainian, Spanish, French, German, Italian, Portuguese, Polish, Japanese, Korean, Chinese, and more.

Troubleshooting

Microphone access denied

macOS: System Settings → Privacy & Security → Microphone → enable VS Code → restart VS Code

Windows: Settings → Privacy → Microphone → allow app access

Linux: Check PulseAudio/PipeWire settings with pavucontrol, make sure no other app is blocking the mic

How to check logs

Command Palette (Ctrl+Shift+P / Cmd+Shift+P) → "Developer: Open Webview Developer Tools" → pick Voice Transcriber → Console tab

Transcription fails

• Check your API key
• For local API — make sure the server is running and URL is correct
• Check your internet connection

Large files and video uploads

Recordings over 24 MB are transcoded to 128 kbps MP3 and split into chunks (10 minutes for OpenAI and local, 5 minutes for OpenRouter, whose routing overhead makes long chunks much slower), each transcribed separately and concatenated. Video uploads (MP4, MKV, MOV, AVI, WebM) have their audio track extracted automatically. Both features require ffmpeg.

Privacy

• API keys are stored in VS Code's secure storage (system keychain)
• Audio goes directly to the provider you picked (OpenAI, OpenRouter) or your local API
• Nothing is saved to disk

For Developers

Setup

npm install
npm run compile

Press F5 to launch the Extension Development Host.

Commands

npm run compile   # build once
npm run watch     # rebuild on changes

Verifying a provider end-to-end

Hits the live API with real audio through the compiled services:

npm run compile && OPENROUTER_API_KEY=sk-or-v1-... node scripts/check-provider.js openrouter

Swap in OPENAI_API_KEY=sk-... node scripts/check-provider.js openai for OpenAI. Pass an audio file path as a second argument to skip the generated sample (sample generation needs macOS say plus ffmpeg).

Publishing to VS Code Marketplace

Prerequisites

1. Microsoft account — account.microsoft.com
2. Azure DevOps org — dev.azure.com
3. Publisher ID — marketplace.visualstudio.com/manage

Get a Personal Access Token (PAT)

1. Go to dev.azure.com → profile → Personal access tokens → New Token
2. Organization: All accessible organizations
3. Scopes: Custom defined → Marketplace → Manage
4. Copy the token (shown only once)

Update package.json

{
  "publisher": "your-publisher-id",
  "icon": "resources/icon.png"
}

Icon must be a 128×128 PNG.

Publish

npm install -g @vscode/vsce
vsce login your-publisher-id
vsce publish

Update version

vsce publish patch  # 0.1.0 → 0.1.1
vsce publish minor  # 0.1.0 → 0.2.0
vsce publish major  # 0.1.0 → 1.0.0

Other useful commands

vsce package                      # create .vsix without publishing
vsce show publisher.extension     # show extension info
vsce unpublish publisher.ext      # remove from marketplace

License

MIT