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+Alt+V (Mac) / Ctrl+Alt+V (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.
• Push-to-Talk (hold to talk) -- hold right-Command to insert, or right-Option to insert and submit; release to deliver. A short tap never records.
• Agent-Native Delivery -- a focused terminal agent pane (e.g. a herdr-managed Claude Code session) gets the cleaned text typed straight in -- no Accessibility permission needed -- while every other app falls back to the clipboard paste. The right template is picked automatically from the focused surface (agent / terminal / editor).
• 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)

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+Alt+V / Ctrl+Alt+V, 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