VoiceBox turns Whisper dictation into formatted text

VoiceBox is an open-source speech-to-text tool built around Whisper and a cloud-hosted formatter, designed to drop polished text directly into whatever app you are using. The workflow is simple: hold Ctrl+Cmd, speak, release the keys, and the app transcribes your speech, formats it with an LLM, copies the result to the clipboard, and auto-pastes it into the originating app.

The desktop client is built with Wails, using Go and a React WebView, while the backend runs as a Cloudflare Worker with a Durable Object handling audio accumulation and the AI pipeline. According to the project page, the worker uses @cf/openai/whisper-large-v3-turbo for speech recognition and @cf/qwen/qwen3-30b-a3b-fp8 for formatting.

On macOS, VoiceBox captures the focused element’s context — including app name, bundle ID, element role, title, placeholder, and current value — and sends that along with audio settings in the WebSocket session so the formatter can tailor the output. Auto-paste requires macOS Accessibility permission.

The app currently targets macOS, with window_darwin.go handling overlay, settings, and dock behavior, while non-macOS builds are stubbed out. Setup requires:

• Go 1.24+
• Node.js + pnpm
• Wails v2 CLI
• A Cloudflare account with Workers AI access
• macOS for auto-paste support

The default audio pipeline runs at 16kHz, mono, PCM signed 16-bit LE, in roughly 4096 byte chunks or about 128ms each. Maximum recording size is listed at about 25 MiB, or roughly 13 minutes.

Recommended reading

Signal tests Android phones as linked devices

The repository also outlines a planned Phase 2 local backend using faster-whisper for transcription and Ollama for formatting. For now, the main path is cloud-based, configured through a voicebox.toml file that points the desktop app to a deployed Cloudflare worker and shared secret.