Verba

Verba — developer-grade voice dictation, everywhere you type.

Speak instead of type. Verba records your voice, transcribes it with Deepgram Nova-3, and refines it with Claude — system-wide across macOS (Beta) or deep inside VS Code. Bring your own keys, keep your data.

---

Why Verba

• Bring Your Own Key -- your own Deepgram + Anthropic keys; no subscription.
• Privacy & data control -- keys in the OS keystore; optional fully offline transcription (whisper.cpp); your audio and text are never routed through us.
• Developer- & code-aware -- code-aware templates, Claude Code prompt generation, commit messages, JavaDoc, and deep VS Code integration.
• Everywhere -- the same dictation intelligence in your editor and across your whole Mac.

Two ways to use Verba

VS Code Extension

Dictate into your editor and terminal, with code-aware AI templates.

The shipped flagship, available today on the VS Code Marketplace:

ext install talent-factory.verba

Or search for "Verba" in the VS Code Extensions sidebar.

Full guide: VS Code Installation

macOS App (Beta)

System-wide dictation into any app — press a hotkey, speak, and Verba pastes clean text wherever your cursor is.

Beta -- build from source. There is no packaged download yet; run it via just macos-dev (or npm run tauri dev from apps/macos).

Full guide: macOS Installation

Features

Core dictation intelligence (both surfaces)

• Streaming Post-Processing -- Claude cleans up your transcript in real time as it streams back.
• Course Correction -- self-corrections in speech are detected and removed automatically. Say "let's meet tomorrow, no wait, on Friday at ten" and only "let's meet on Friday at ten" is kept.
• Voice Commands -- spoken formatting commands like "new paragraph", "comma", "bullet point" are converted to actual formatting, in any language.
• Glossary/Dictionary -- define terms that must be preserved exactly during transcription and cleanup (e.g. "Visual Studio Code", "Kubernetes").
• Text Expansions -- user-defined abbreviations are expanded during post-processing.
• Prompt Templates -- template-driven post-processing (Free Text, Commit Message, JavaDoc, Markdown, E-Mail, and more), freely configurable with your own prompts.

VS Code

• Editor & Terminal Insertion -- Cmd+Shift+D (Mac) / Ctrl+Shift+D (Windows/Linux) starts and stops recording; text is inserted contextually in the editor or the integrated terminal.
• Multi-Cursor / Selection-Aware Dictation -- dictated text replaces a selection, fills every cursor position, or is passed to Claude as context for the transformation.
• Dictation History & Full-Text Search -- every dictation is saved locally and searchable across raw transcript and cleaned text; re-insert or copy past results.
• Continuous Dictation -- longer sessions via Deepgram Nova-3 WebSocket streaming with automatic pause segmentation (Cmd+Shift+Alt+D).
• Offline Transcription -- transcribe fully locally with whisper.cpp: zero API cost, audio never leaves your machine.
• File-Type-Aware Templates -- the right template (e.g. JavaDoc for Java/Kotlin) is selected automatically from the active editor's language.

macOS (Beta)

• System-wide Global Hotkey -- Ctrl+Alt+D toggles microphone capture from any app.
• Paste into the Frontmost App -- transcribed, cleaned text is pasted wherever your cursor is; the previous clipboard content is restored afterwards.
• Menu-Bar Configuration & Template Picker -- switch transcription provider, cleanup language, and active template straight from the tray menu.
• HUD Working Visualization -- a non-activating, click-through pill shows idle/recording/transcribing/processing state.
• Keychain-Backed Keys -- Deepgram and Anthropic keys are stored in the macOS Keychain.

Prerequisites

• ffmpeg -- required for the VS Code extension's microphone recording (the macOS app captures audio natively via cpal and does not need ffmpeg)
• Deepgram API Key (Nova-3 transcription), shared by both surfaces -- or whisper-cpp for offline transcription (VS Code only, for now)
• Anthropic API Key (Claude post-processing), shared by both surfaces

Installing ffmpeg

macOS:

brew install ffmpeg

Linux (Debian/Ubuntu):

sudo apt install ffmpeg

Linux (Fedora):

sudo dnf install ffmpeg

Windows:

Download from ffmpeg.org and add to PATH, or via Chocolatey:

choco install ffmpeg

Platform-Specific Notes (VS Code)

Platform	Audio Backend	Microphone Selection
macOS	AVFoundation	Configurable via verba.audioDevice or Quick Pick
Linux	PulseAudio	Configurable via verba.audioDevice or Quick Pick
Windows	DirectShow	Configurable via verba.audioDevice or Quick Pick

On all platforms, you can select the microphone anytime with the command Verba: Select Audio Device or by setting verba.audioDevice in Settings. Without configuration, the system default microphone is used.

Linux: PulseAudio must be running (default on Ubuntu, Fedora, and most desktop distributions).

Windows: On first use, a Quick Pick dialog lets you select the microphone. Verba detects devices via ffmpeg (v7 and v8+ formats) with a PowerShell fallback.

Get started

• VS Code: press Cmd+Shift+D / Ctrl+Shift+D, pick a template on first use, speak, then press the shortcut again -- the result appears at your cursor. Full walkthrough: VS Code Quick Start.
• macOS (Beta): build from source with just macos-dev, then press Ctrl+Alt+D anywhere to start dictating -- Verba pastes the cleaned text into the frontmost app. Full walkthrough: macOS Installation.

Configuration

VS Code reads templates, glossary, and expansions from settings.json, with project-specific overrides in .verba-glossary.json and .verba-expansions.json -- see the VS Code Configuration guide. macOS (Beta) uses a single config file at ~/.config/verba/config.json (XDG), editable via the tray menu or directly -- see the macOS Configuration guide.

Architecture

                          @verba/core
        (pipeline · cleanup · Deepgram provider · config schema)
                       /                        \
          VS Code Extension                macOS App (Beta)
          (Electron / Node.js)               (Tauri / Rust)

Both hosts wrap the same dictation pipeline (record → transcribe → Claude post-process → insert/paste) around platform-specific adapters. Full breakdown: Architecture.

Documentation

Full documentation site: talent-factory.github.io/verba -- or browse the docs/ directory directly.

Contributing

Found a bug or have a feature request? Open an issue. For local development setup (monorepo layout, build commands), see Contributing.

License

MIT