10 November 2026
Why voice runs on your PC
Windows dictation was almost always wrong. Transcription moved to Whisper on the machine that is already recording.
The wake word is “hey clip”. It has to hear a quiet desk mic, grab the last moments, and settle a title without sending the audio to this website.
What failed
Windows dictation was almost always the wrong sentence. A cloud speech route existed for a short time. It is gone. The site does not receive your voice.
Quiet microphones also missed the wake word. That became a mic picker and a gain control, with a level meter so you can see the signal.
What runs now
Whisper small.en runs on NVIDIA through CUDA. base.en runs on CPU when there is no NVIDIA GPU. The model stays loaded. Endpointing is short, so the clip does not wait on a long silence. Windows TTS talks back. You can barge in to settle the title.
That path is the free, on-device voice. Connecting an advanced model, for smarter titles and a conversational assistant, plus auto upload, is the supporter plan. It needs an account, and it is not for sale yet.
The unsigned installer you can download today is the free beta. The open-source fork is still coming. The signed daily build from main waits on code signing.