The biggest update in a while: two large features and one long-standing user request.
Meeting recording. NeuroSwitcher now records a whole conversation — your microphone and the system audio, which is the other participants' voices — as two separate tracks.
- The transcript already knows who said what. Not by guessing from the sound of the voices, but because they were two different tracks: yours, and everyone else's.
- Transcription is a separate action. Stop the recording and nothing goes anywhere. You pick a session and start the transcription yourself — in an hour or next week.
- The files stay yours. One folder per conversation: audio, text, subtitles with timecodes. You delete them; set a retention period and old ones clear themselves.
- It starts from the keyboard. Ctrl+Shift+F8 begins a recording and ends it — no trip to the tray while the call is already running. While it records, a small badge shows the elapsed time and a level meter per track: drag it anywhere by the grip on its left, and pause and stop are buttons on the badge itself.
- One file, and a mute button. A recording is stored as a single stereo file — you on the left, everyone else on the right. On the badge, the microphone meter is also a mute button: your voice stops going into the file while the other side keeps recording, and the meter draws struck through so you always know.
- Punctuation, put back (optional). Quiet speech comes back from the recogniser unpunctuated and lowercase, and foreign terms spoken inside another language come back mangled. The assistant can repair both after transcribing — correcting only, never rewriting. Off by default: it sends the text of the conversation to your AI provider.
- Read the warning before the first recording. It captures other people's voices, and Windows shows them no indicator. In many countries recording a conversation without everyone's agreement is illegal. The app asks you about this once — and it is not legal advice.
Recording is a Pro feature, open to everyone during the beta. Transcribing on your own PC costs nothing.
GPU acceleration for transcription. If you have an NVIDIA card, the Meeting recording section can download a separate acceleration pack (about 290 MB) and switch transcription from the processor to the graphics card. On the machine we measured, the large recognition model got 5.7× faster: 163 seconds of speech in 112 seconds instead of 637. That is the difference between a model worth turning on and one nobody waits for.
- You will need NVIDIA's CUDA 12.x with cuDNN 9.x. The pack does not contain them, although the archive is named after them — that is a separate install from NVIDIA's own site, and most people with a gaming card have never done it. Until they are on the machine the app will not even offer the download: instead of a button you get a line naming exactly what is missing.
- Qwen3 stays on the processor. It is slower on the graphics card, and the app will not move it there. The acceleration is about the Whisper models, the large ones above all.
Web search in the assistant. There is now a globe next to the input box.
- It switches on for one question and does not stick: the next one answers normally.
- The answer comes with its sources — a collapsed list of links that opens on click and takes no room in the thread meanwhile.
- It works with NeuroSwitcher Cloud only. If you have your own key or a local model selected, there is no globe: that is a capability those backends do not have, not a fault.
- The text of your question also goes to a search service, not only to the AI provider — there is nothing to search otherwise. Your ordinary typing, as always, is never sent anywhere at all.
Languages you never type in stop getting in the way. Asked for by a user.
- Correction no longer offers a language you have no keyboard layout for, and no longer pulls that layout into your system by itself. There is nothing to configure — that is simply how it behaves now.
- The list is in Settings, under "Languages / Keyboard layouts". Untick one and it is off entirely.
- If a language you want is missing from the system, the app offers to add the layout for you — and says plainly that it will join your Win+Space cycle.
Also in this release:
- The assistant stopped naming the cloud model in its header. We route it server-side and the choice moves with cost and availability, so printing today's name would be a promise we never made. A self-configured backend still shows its own — there you picked the model.
- The source list no longer follows you into the next question. It stays on screen but is kept out of the model's context: last answer's links are hundreds of wasted tokens that grow with the conversation.
To update: the console will offer the update itself, or download the installer from the site. It installs over your current copy — settings, dictionaries and licence are kept.