Whispering

Whispering is built around one gesture: hold a key, speak, release. The browser build handles in-app shortcuts, and the desktop app adds system-global shortcuts that fire no matter which window is focused. That is the core loop, and it's why the app stays out of your way. You don't open a separate tool, record a file and then hunt for it later. The text arrives at the cursor you were already using.

The app records speech and hands the audio to whichever provider you configure. You can connect a cloud service, the hosted Epicenter gateway or a self-hosted endpoint, and you supply the credentials. That design keeps Whispering free and keeps your options open, since you're not stuck with one vendor's accuracy or pricing. Change your mind and switch the provider in settings, and the app keeps working the same way.

The Epicenter desktop build can transcribe with a local GGUF model instead of sending audio anywhere. That option only exists on desktop, because the native runtime that loads the model lives there rather than in the browser. For anyone working with confidential recordings, it means the audio never leaves the device and no internet connection is required. It's the feature that turns Whispering from a cloud dictation tool into a private one.

After the raw transcript comes back, Whispering can run an optional pass that cleans it up. The polish step is a separate layer, so you decide whether you want the plain output or a tidier version. It's useful when you're dictating in a rush and the first pass is rough. You can leave it off entirely if you'd rather see exactly what was spoken.

Getting text out of a dictation app is often the annoying part. Whispering targets the active cursor directly. In the desktop app it uses native delivery when the operating system allows it, and falls back to the clipboard when it doesn't. In the browser it uses the clipboard. Either way, the transcript ends up in the field you were already in, so there's no window-switching step in the middle of your workflow.

Whispering keeps settings and recording metadata on your device first. Audio only leaves the machine when the provider you selected requires an upload, and if you're using a local model, it never leaves at all. Recordings are stored in the browser's IndexedDB or in Epicenter app-data files on desktop. That local-first approach is what separates it from dictation tools that route every clip through a company server by default.

The same app runs as a hosted web page and as a window inside Epicenter. The browser version is a real target, not a stripped-down fallback, and it handles browser recording, local blobs and web-safe shortcuts. The desktop version swaps in native recording, system permissions, local model transcription and floating overlays. You pick the host that matches how you work, and the core dictation flow stays the same.