Changelog
What changed in each release, and when.
1.2.0 this is the current download
Published 6 September 2026.
- Voyce can finally learn from the way you actually speak. Until now the only place it could learn was the test box inside its own settings window. Everything you dictated into Outlook, a browser or your CRM was typed into that other program, and Voyce never found out what you changed it to. So the promise that it improves with your voice could only come true if you sat and tested it, never from doing your job. On one machine that meant twenty recordings and not a single usable one after weeks of daily use.
- There is now a "Teach Voyce your voice" section in settings. It lists what you have dictated, plays each one back, and lets you fix anything it got wrong. Every one you check is kept as your audio paired with your words, which is the only thing that can be used to train it on how you speak. Correcting a word there also fixes it everywhere immediately, exactly as correcting it in the test box always did. None of it leaves your PC.
- Groundwork for real training. The machinery that trains on those recordings is built and tested. It measures how well the current model does on recordings it has never been shown, trains, measures again on those same recordings, and refuses to install anything that is not genuinely better. It also refuses to run at all until there are about ten minutes of checked speech, because below that it would simply memorise those clips and get worse at everything else.
1.1.2 superseded by 1.2.0
Published 5 September 2026.
- The settings panel was dead, in every version so far. Opening Voyce and looking at the panel, it said "Loading" forever and nothing in it responded: the switches, the microphone list and the licence box all did nothing. One stray line break inside the panel's own code stopped the whole page from running. Dictation itself was never affected, which is why this went unnoticed for so long, and it also means anyone who tried to enter a licence key was told nothing was wrong while the button was not connected to anything.
- Your words are transcribed on a far better model, and much faster. Voyce now sends what you say to OWT's own transcription server first, which uses a larger model that also understands languages other than English. If that server cannot be reached, for any reason, it falls straight back to transcribing on your own PC exactly as before. You are never left unable to dictate, and your audio never goes to anyone outside OWT.
- The Remove button no longer hangs out of the licence box, and Voyce's character now sits behind that card with only his eyes showing, instead of his whole face floating over it.
- The push-to-talk key is more reliable on unfamiliar keyboards. Voyce used to recognise your key only by its name, which is not the same on every keyboard and every language setting. When it did not match, the key simply did nothing and the app looked broken. It now also learns the key's hardware code the first time it starts.
1.1.0 superseded by 1.1.2
Released 27 August 2026.
- AI formatting works. It never has before. The setting has been in every release, and the part that actually runs the language model was never included in the build. On a PC with nothing else installed it quietly did nothing and handed your words back unchanged. It now ships with the app. Turn it on under Settings, and it will punctuate and tidy what you dictate without changing your words.
- The indicator tells you which stage it is on. It used to say "transcribing" the whole time, including while the AI was still shaping the text, which is a different job that happens after the listening has finished. It now says thinking during that part, so a pause has an explanation.
- Voyce keeps a recording of everything you dictate, on your own PC. It used to keep only the ones you corrected, which meant anyone who dictates all day and corrects nothing gave it nothing to learn from, and that is most people most of the time. It keeps up to 500 MB and deletes the oldest first. It is never sent anywhere. You can switch it off or erase all of it under Settings, and the panel now describes this accurately, which for one release it did not.
1.0.9 superseded by 1.1.0
Released 26 August 2026.
- You can see your licence key before you remove it. Signing out cleared the key and there was no way to read it first, so anyone who had lost the original email had no way back in.
- Older versions of this site are no longer publicly reachable. They were being served normally to anyone with the address.
1.0.8 superseded by 1.0.9
Released 26 August 2026.
- Correcting a word now keeps the audio with it, on your PC, so the correction can teach Voyce how you actually say it rather than only how it is spelled.
- A rule for typing while you are still speaking, so live text commits cleanly instead of rewriting itself as you go.
- The status indicator wears proper artwork, and the site has a face.
1.0.7 superseded by 1.0.8
Released 25 August 2026.
- Voyce works the moment you install it. The speech model used to be a separate 145 MB download hidden in Settings, and if you missed it, holding the key recorded you and produced nothing. The model now comes with the app. The installer is 209 MB instead of 90 MB as a result, which is a one-off download in exchange for never being silent on a fresh machine.
- It uses your graphics card when it finds one it can genuinely use, and falls back to the processor otherwise.
1.0.6 superseded by 1.0.7
Released 25 August 2026.
- Punctuation and capital letters are back. The shared vocabulary added in 1.0.5 was written in lower case with no full stops, and that taught the recogniser to write everything else the same way. It stripped the capitals and full stops out of ordinary dictation. If you installed 1.0.5, update.
- The Settings toggle that the team and docs pages already described now exists.
- The version embedded in the file properties was stuck on 1.0.4.
1.0.5 superseded by 1.0.6
Released 25 August 2026.
- Voyce learns as a team. 317 destination, hotel and transfer names are built in, taken from our own supplier records, and a shared pool means one person teaching Voyce a word gives it to everyone. Only words are shared, never audio and never anything you dictate, and a word is only shared once several machines agree on it.
- One mark everywhere, and a window that is black all the way to its edge.
1.0.4 superseded by 1.0.5
Released 24 August 2026.
- Stop automatically now works. It never had. The speech detection model was not being included in the build, so the feature sat in Settings doing nothing in every release up to this one. The error it reported blamed something else, which is why nobody caught it.
- Your settings survive a corrupted config file. If the file ever picked up a byte order mark, from Notepad or a script, Voyce treated it as broken and started fresh, which cost you your licence and every preference.
1.0.3 superseded by 1.0.4
Released 24 August 2026.
- OWT agents can activate with a company code instead of a licence key. Open the panel, go to Licence, type the code your IT team gave you, and press Activate. That is the only step that needs the internet.
- Voyce tells you when a newer version exists. The check sends nothing about you or your machine: it reads a small file on this site and compares it with the version on your own PC.
1.0.2 superseded by 1.0.3
Released 24 August 2026.
- Uninstalling no longer deletes your licence key. It used to remove everything: your settings, your key, and every word you had taught Voyce to spell correctly. Anyone reinstalling to fix a problem lost all of it, and the learned words were not recoverable from anywhere. Now an uninstall removes the application and the downloaded speech model, which you can fetch again in a click, and keeps the rest. We found this by actually uninstalling it, which nobody had done before.
- Uninstalling cleans up properly. It was leaving about 110 files behind while reporting success.
1.0.1 superseded by 1.0.2
Released 24 August 2026.
- Wake word is gone from the panel. It had been visible but permanently disabled, labelled "coming in a future update". The usable pre-trained wake models are licensed for non-commercial use only, so a paid product cannot ship one, and that is a licence term rather than something a future release fixes. A switch nobody can ever use should not be in the settings, so we removed it instead of leaving it greyed out with a promise attached.
- Optional error reports. Off unless you turn them on. When something breaks on your machine we currently learn nothing about it unless you write to us. If you switch this on, a failure sends the app version, your Windows version, which feature failed and the error message. It cannot send audio or transcripts: there is no field for them, and the server refuses anything that is not on the list.
- The publisher name is corrected to MindScope Solutions FZC. It is what Windows shows you in the file properties and in Add or remove programs, and it was previously spelled two different wrong ways.
1.0.0 superseded by 1.0.1
Released 23 August 2026.
- Automatic stop on silence now works, and without the 531 MB of machine learning framework it used to need. The installer is smaller for it.
- AI formatting can obtain its model, and loads on machines where Windows blocks the bundled component, by using a signed local runtime instead.
- Speak to edit. Select text anywhere, hold a key and say what to do to it.
- The correction box teaches. Fixing a word in the panel adds it to the glossary that biases the next transcription, so names and jargon stop coming out wrong.
- Fixed: the panel was unresponsive for the first few seconds after opening. Every control looked ready and did nothing until loading finished.
- Fixed: starting to set a new push-to-talk key and then changing your mind disabled dictation until you restarted the app.
- Fixed: a Pro switch could appear on while doing nothing. Controls the machine cannot actually run are now disabled and say why.
- The real Voyce mark ships at every icon size, and the app identifies its publisher.
1.0.0 is the first public release, so there is nothing before it. Earlier work is in the repository history rather than here, because a changelog for versions nobody could install is a list of things that never happened to anyone.