Offline Transcription

By default, Verba sends recorded audio to the Deepgram Nova-3 API for transcription. As an alternative, the VS Code extension can transcribe entirely on-device via whisper.cpp — no audio ever leaves your machine, and there's no per-minute API cost.

Platform support

Offline transcription currently supports macOS. Linux and Windows support is planned for a future release.

When to Use It

• Privacy — sensitive dictation (credentials, proprietary code, confidential notes) never crosses the network.
• Zero cost — no Deepgram usage charges; only your one-time model download.
• Offline environments — works without an internet connection once a model is downloaded, though Claude post-processing still requires network access to the Anthropic API.

Setup

1. Install whisper.cpp:

   brew install whisper-cpp
   

2. Download a model with the command Verba: Download Whisper Model (dictation.downloadModel). Models are fetched from Hugging Face and stored in VS Code's global storage, shared across all workspaces.

3. Switch the transcription provider in settings.json:

   {
     "verba.transcription.provider": "local",
     "verba.transcription.localModel": "base"
   }
   

Available Models

Model	Size	Speed	Quality
tiny	~75 MB	Fastest	Lower accuracy
base	~148 MB	Fast	Good balance
small	~488 MB	Moderate	Better accuracy
medium	~1.5 GB	Slow	High accuracy
large-v3-turbo	~1.6 GB	Slowest	Best accuracy

Tip

Start with base. If accuracy is insufficient, upgrade to small or medium. large-v3-turbo gives the best quality but needs the most disk space and processing time.

How It Works

TranscriptionService uses a strategy pattern: setProvider('deepgram' | 'local') selects the active backend, and process() delegates to it. The cloud path goes through core/deepgramProvider.ts; the local path goes through localWhisperProvider.ts, which shells out to the whisper-cli binary and parses its output. Both implement the same TranscriptionBackend contract, so the rest of the pipeline — Claude post-processing, glossary, voice commands, course correction — is identical regardless of which provider transcribed the audio.

Related Settings

Setting	Default	Description
verba.transcription.provider	"deepgram"	"deepgram" (cloud API) or "local" (whisper.cpp)
verba.transcription.localModel	"base"	tiny, base, small, medium, or large-v3-turbo

See Configuration for the full settings reference.