Whisperstream

About Whisperstream

Whisperstream is push-to-talk dictation for Windows that transcribes entirely on the local machine — built on NVIDIA's Parakeet model, with a Microsoft code-signed installer. The privacy claim is structural rather than a policy promise, and the vendor spells out the exact sequence: the microphone captures audio, the local model converts it to text, the text is typed into the active window, and the audio is discarded from RAM. Transcription makes zero network calls; the only connections the app opens are license checks and updates, neither of which carries voice or text. The single exception is opt-in — a free cloud fallback for AI cleanup on older hardware — and Lockdown Mode turns off every cloud path and puts transcript history behind a password. Beyond the core loop it has fifteen features, of which the notable ones are AI cleanup that cuts filler and applies spoken corrections locally, a custom dictionary for names, jargon, and acronyms (the site's own examples include mapping "my sequel" to MySQL and "next js" to Next.js), per-app modes that apply a custom AI prompt automatically depending on which app or site you are in, searchable transcript history with replayable audio encrypted on device, and audio-file import for transcribing existing recordings. It covers 39 languages, all running locally, and works offline. Because it types into the active window it needs no add-in — Outlook, VS Code, Gmail, Slack, and Claude Code are all called out.

Key Features

Tags