Fixed: letters running together when you type fast. When a sentence started right after a full stop, the automatic capital and the layout correction could land on top of each other, and a typed "I think" came out as "iIthink". The substitution now waits for the keys you have already pressed.
The Dictation tab is rebuilt. Where speech is transcribed — on this computer, at OpenAI under your own key, on your own server, or in NeuroSwitcher Cloud — is now a row of tabs, and each one shows only its own settings. Under them, one line saying where the audio goes. Before this, the graphics-card setting and the model catalogue stayed on screen whatever you picked, so the console could claim that "recognition runs on the graphics card" while every word was going to a server.
Start engine no longer demands a downloaded local model when transcription is set to the cloud.
Meeting recorder: the server now decides how long a transcription batch may be, speaker labels hold together for longer, and long batches are reserved for NeuroSwitcher Cloud.
Cloud dictation is no longer charged about a third more than it should be — a counting error on our side.