What's changed

15 October 2026 · 0.3.4

A new address - vstt.alexvamp.app

New

• The site, the downloads and the update feed have moved to vstt.alexvamp.app. The installer and the speech model are now served through Cloudflare's network rather than from a single server - faster to download, and still there if that one server is not. Nothing is needed from you: copies already installed find this version at the old address as usual, and check the new one directly from then on.

10 September 2026 · 0.3.3

Dictating through AirPods Max

Fixed

• Dictation was silent while AirPods Max were connected. The headphones' microphone was listed and selected, the capsule showed it recording - and no text arrived at all, as though nobody had spoken. The sample rate is why: AirPods Max record at 24 kHz, where the built-in microphone and almost any USB one record at 48. The way VSTT took audio in could not survive that difference and was handed not one sample from the headphones. Recording now goes through a different system mechanism, one that handles any rate equally well - so AirPods Max, like any other headset, are heard as well as the built-in microphone. Ducking works properly along with it: the music in your ears is taken down while the key is held and comes back at exactly the level it was playing at, including for the couple of seconds it takes the headphones to leave headset mode. Dictation also starts noticeably sooner: audio flows about 0.2 seconds after the key goes down, rather than 0.85.

31 August 2026 · 0.3.2

The dictation key no longer takes the app down

Fixed

• On macOS 26.7 VSTT quit on the first press of any modifier key. Dictation could not be started at all: the app closed the moment a key went down - and not only its own key, but any of Shift, Ctrl, Option, Command. The cause was a housekeeping check VSTT ran on every keypress, which the new system version turns into a crash. That check is gone from the key's path, and holding to dictate works as it did before.

27 August 2026 · 0.3.1

Keeping nothing at all

New

• Transcriptions can be kept out of storage entirely. Until now everything dictated went into a file on this Mac and stayed there until you deleted it. Settings now has a Privacy section with a switch: turn it off and no new transcription is written down anywhere - no file, no list. What is already saved is left alone; deleting that stays your decision, with Delete All under Transcriptions.
• Launch at login switches itself on for a fresh install. VSTT lives in the menu bar, and the dictation key only works while it is running - yet after installing, it did not start when you logged in, and you had to work that out for yourself. It now turns itself on the first time it runs, exactly once. An existing install where you deliberately turned it off is left as it is.
• Launch at login can be toggled from the menu bar. It used to mean opening Settings.

26 August 2026 · 0.3.0

Music stays out of your dictation

Fixed

• The Dock icon disappeared while an app window was still open. Switching to another program was enough: VSTT's window ended up covered, and the icon went with it - taking Cmd+Tab along and leaving the mouse as the only way back. The icon now stays until the last window is closed.

New

• You can choose which microphone to use. With more than one to hand - built-in, USB, a headset - VSTT no longer has to take whichever the system assigns. The choice lives in Settings, under Microphone. It is remembered by the device itself rather than by its name, so renaming one, or owning two of the same, confuses nothing. If the chosen one isn't connected, dictation goes through the system's and the capsule says so - its dot turns amber for the length of the hold. Plug the microphone back in and it takes over again, with nothing to re-pick.
• Other audio is silenced while you dictate. Sound from the speakers reaches the microphone, and the engine transcribes it along with your speech - words you never said end up in the text. Holding the dictation key now takes the system volume to silence, and letting go fades it back to where it was. It works with anything: Music, Spotify, a video in a browser tab, a game. If nothing is playing at that moment, the volume is left untouched. There is a switch in Settings for anyone dictating in headphones, where there is nothing to duck.

26 August 2026 · 0.2.9

A home of its own - vstt.app

Added

• VSTT has moved to vstt.app. The site, the downloads and the update feed now live on their own domain instead of vstt.alexvamp.ru. Nothing is needed from you: the old address redirects to the new one, so copies already installed find this update by themselves, and check vstt.app directly from then on.

25 August 2026 · 0.2.8

The first run finishes what it starts

Added

• The dictation HUD shows two lines. A dictated phrase outgrows one line almost immediately, and only its tail used to be visible. Two lines now show the last two rows, so what you said can be read and checked rather than guessed at, and the capsule no longer changes height mid-dictation. A setting switches it back to one line for anyone who wants the smaller capsule.

Fixed

• Dictation stopped working after the model downloaded, until the app was restarted. Loading the weights into memory for the first time takes several seconds - noticeably longer on Intel processors. Letting go of the key during that wait broke recognition for good: the session finished opening after everyone had stopped waiting for it, leaving nobody to close it. Every later hold then failed on that abandoned session, and only a full restart cleared it.
• Long error messages were unreadable. The HUD trimmed them from the front, leaving a meaningless tail like "…& Security ▸ Microphone" that never said what had gone wrong. Messages now wrap onto a second line and trim from the end when they still don't fit, so the start of the sentence is always there.
• Without microphone access the app acted as though it were recording. Where the permission had never been asked for, the HUD said "Listening…" at exactly the moment nothing could be heard: the app was waiting on the system prompt, and releasing the key just took the HUD away without a word. Access is now checked before anything claims to be recording, and the HUD says what is missing - with different wording for "never asked" and "refused", because the way out of each is different.
• The first dictation right after the download sat silent for several seconds. The model was only loaded into memory the first time something asked it to transcribe, and the HUD showed nothing while it did. It now loads the moment the file lands on disk, so the download finishing and the app becoming usable are the same moment.
• "Grant Microphone…" in the menu did nothing. It opened the System Settings pane that a freshly installed VSTT isn't listed in - macOS lists an app there only once it has requested access. The item now brings up the system prompt, and only sends you to Settings once access has been refused and the prompt won't appear again.

25 August 2026 · 0.2.7

The dictionary file actually exists

Fixed

• "Reveal Dictionary File" opened nothing. dictionary.txt was only written once you had edited the dictionary inside the app - before that there was no file on disk, so Finder had nothing to show. The same gap broke editing the dictionary in a text editor: the app can't watch a file that isn't there. The file is now present from the start, with the format documented inside it.

24 August 2026 · 0.2.6

Settings in plain sight

Added

• A settings button right in the window. Settings used to open only via ⌘, - if you didn't know that shortcut, there was no way to find them. There's now a gear icon next to the record button.

Fixed

• The push-to-talk key could go dead after sleep or a locked screen. If the Mac slept or the screen locked mid-hold, the app could stop responding to the key entirely - only a restart fixed it. That interruption is now recognized and the key works again right away.
• An error message sometimes showed up in English. A rare speech-engine failure could show a technical English message in the HUD instead of a translated one. Those now show translated text.

24 August 2026 · 0.2.5

The tray behaves like a tray

Fixed

• The window still opened on a plain launch. The tray icon showed up correctly, but launching the app by hand - rather than through login - opened the main window on screen anyway. It no longer opens on any launch path, only via "Open VSTT" in the tray.
• The Dock icon stayed after closing the window. Open the window from the tray, close it, and the icon never went away. The check for "is any window still open" was counting the tray icon itself as one, so the condition could never be met. It now correctly disappears on close.

24 August 2026 · 0.2.4

Apple's engine knows its own limits

Fixed

• Apple's engine is no longer offered for languages it doesn't know. Apple transcribes 30 languages and Russian isn't one of them - yet the engine stayed selectable, and holding the key then silently did nothing. It is now shown as unavailable, with the reason, and the setting moves to Nemotron on its own. It also removes the case where Russian speech was quietly run through an English model.
• A dictation error disappeared before you could read it. The HUD hid itself the same instant the error message appeared, so a failed hold looked like nothing had happened at all. The message now stays up for the few seconds it was always meant to.

24 August 2026 · 0.2.3

An actually universal build

Fixed

• On Apple Silicon the app was built as an Intel binary and had been running under Rosetta this whole time - that's what triggered the "will stop working" warning and left Apple's on-device engine silently refusing to run, since it needs a native process. The build is universal now: real arm64 on Apple Silicon, x86_64 on Intel.

24 August 2026 · 0.2.2

The model, back on HuggingFace

Fixed

• The model downloads from HuggingFace again first - their CDN is faster for most people and doesn't load our server. If that doesn't work, the download automatically continues from our site.

24 August 2026 · 0.2.1

Launch at login, and a quiet start

Added

• Launch at login. A switch in Settings, under Startup, mirrors System Settings → General → Login Items, so both places always agree.
• Starts tray-only. No Dock icon appears until you open the window yourself with "Open VSTT" in the menu bar, and it disappears again once you close it.

24 August 2026 · 0.2.0

The model, updates, and a repaired interface

New

• The speech model downloads itself on first launch - before, you had to find a menu item nobody knew to look for. Progress is visible, and trying to dictate mid-download tells you how far along it is.
• Downloads resume where they stopped. A closed laptop or a dropped connection used to mean fetching all 716 MB again.
• The model releases its memory when unused - after 15 minutes by default, configurable in Settings, or never. Worth about 1.2 GB.
• Automatic updates. The app checks for new versions and offers to install them.

Fixed

• The app put four identical icons in the menu bar instead of one.
• Settings let you pick an engine that doesn't run on this Mac. It is now unselectable, with a note saying why.
• The level meter was rebuilt: bars respond to actual sound and read as a waveform rather than a battery gauge.

23 August 2026 · 0.1.0

First release

• Push-to-talk dictation: text lands wherever the cursor is.
• Recognition entirely on your Mac - Nemotron 3.5 via transcribe.cpp. Runs on Intel as well as Apple Silicon.
• A dictionary for names and terms the engine keeps getting wrong.
• Cleanup: filler words removed, spacing and punctuation fixed.
• Searchable history of past transcriptions.
• Interface in Russian and English.