feature tab 01
No actuation delay
Start talking first. Trigger while the thought is already moving.
turns speech into polished text in every app
Dictation: Stop waiting. Just talk typing less, unlost idea.
start speaking while your hand moves to the hotkey. sequential → parallel
feature tab 02
Left Shift lives on home row
The shortcut is close enough that capture feels parallel, not ceremonial.
voice to text basics
Accuracy without waiting
Keep the start of the sentence, then let local or cloud STT do the work.
voice to text basics
Polish after capture
Raw speech can become usable text without breaking the flow of speaking.
My responsibility to you.
Your laptop holds your work and private conversations. Asking you to put Roma on it means taking care of more than dictation. It means caring about what the app can access, whether you can depend on it, and what happens when something goes wrong.
I'm Felix. I build Roma and handle support myself. I want to earn your confidence through the choices I make and the work you can check. These are my commitments as I develop Roma, with the current release's limitations below.
What I will do to earn your trust
Today, you can read the source and release notes and contact me directly. The commitments below describe the standard for future releases. They are not all delivered in v1.95.1.
Make your privacy choices work as expected
Off should mean Off. I will make the context switches stop collection itself, explain when audio leaves your Mac, and make local transcription and cloud text processing clear as separate choices. I will explain what history Roma saves and how to remove it.
Take care of the access you give Roma
My priorities for future public releases include an Apple-notarized download, provider API keys stored in macOS Keychain, and clear permission explanations before you grant access. I will say which protections each release has, so you can distinguish a plan from something you can use.
Test the app you will actually download
Before I recommend a release for everyday use, I will test its published download on a clean Mac, including installation, permissions, dictation, text insertion, and quitting. I will also check what happens when context is off, which data leaves the Mac, where keys are stored, and whether history deletion works. I will publish the version, Mac and macOS tested, results, and remaining problems. Testing is my responsibility.
Keep you informed when something goes wrong
You can email me directly. I will investigate, tell you what I can reproduce, and explain what happens next. If I do not have a fix yet, I will say so. When a fix ships, I will name it in the release notes. Helping with debugging or retesting is your choice.
Current v1.95.1 recommendation
v1.95.1 is an early beta. It has no Apple Developer ID signature and is not Apple-notarized. Read the release notes for the checks completed on this download and the cases still untested. I do not recommend it for your main Mac yet.
In v1.95.1, recording includes buffered audio from before Shift. Context switches do not stop clipboard or window collection, and API keys are stored outside macOS Keychain. These are gaps the commitments above are meant to address.
Read v1.95.1 privacy, permissions, setup, and test details
Trying the current download
Apple has not notarized this download, so macOS may block it when you first open it.
If you choose to try the Mac app
Use a secondary Mac without sensitive data. Choose a local transcription model to keep audio on your Mac, and turn AI Enhancement off to avoid sending your transcript and other text for polishing. Cloud transcription can include audio captured before Shift from the rolling buffer; choose a local model if you do not want that audio sent to a provider.
Leave Screen Recording permission off. Granting it lets v1.95.1 capture a visible window when recording starts, even with Screen Context off. Roma also reads your clipboard when recording starts, even with Clipboard Context off. Do not grant Automation unless you need it. Quitting Roma stops microphone capture.
Do not disable Gatekeeper, the macOS security protection. If macOS blocks the app and you are not comfortable choosing Open Anyway, wait.
What the available checks show
Build, unit-test, package, and hands-on runtime checks cover different things. The release notes identify which checks passed on the candidate and which remain untested. A successful build does not establish that Roma works reliably on your laptop.
This download is an ad hoc signed Release build, without Apple notarization. Its checksum lets you check that the file matches the one published; it does not establish safety or stability.
You can explore the dictation interaction in the browser demo without installing the Mac app. The demo starts speech recognition when opened and uses your microphone if permission is available. Your browser may send audio to its speech provider. Roma servers receive or store no demo audio or text. The demo does not establish the Mac app's safety or stability.
What Roma reads and sends
Local model setup suggestion
During first setup, Roma asks its website, hosted by Vercel, for the approximate country of your IP address to suggest Chinese + English or English-only transcription. This request sends no audio. Your explicit choice takes precedence, and an existing model selection is kept. A local model needs an initial download; after that, its transcription runs on your Mac without uploading audio. Cloud transcription and cloud text enhancement are separate choices.
Microphone and cloud audio
With microphone permission, v1.95.1 listens while its rolling buffer is active. Its default buffer holds the latest three seconds of audio and replaces older audio as you speak.
Buffer Preload has been removed in v1.95.1. Recognition starts when you start recording. With a cloud model, the rolling buffer and audio that follows may reach that provider, including speech captured before you pressed Shift.
Local transcription and cloud text processing
A local transcription model keeps audio on your Mac. AI Enhancement can still send text to its provider, even with local transcription. Turn AI Enhancement off to avoid those text transfers. Cloud transcription sends audio directly to the provider you choose.
AI Enhancement sends your transcript, polishing instructions, custom vocabulary, and selected text Roma can read. With Clipboard Context on, it also sends clipboard text. With Screen Context on, it sends the app name, window title, and recognized window text. The provider's retention and training rules apply.
Clipboard and window collection
Roma reads your clipboard whenever you start a recording, even if Clipboard Context is off. With Screen Recording permission, it also captures and reads a visible window, even if Screen Context is off. It tries the frontmost window first, but v1.95.1 can fall back to another visible non-Roma window. Those switches only control what AI Enhancement receives. Revoke Screen Recording to stop window capture.
Saved history and API keys
Roma can save audio, transcripts, polished text, and the messages sent for polishing in local history on your Mac. Automatic cleanup is off by default. History cleanup can remove old records and their audio. Audio cleanup removes old audio but keeps the text record. Deleting local history cannot delete data a provider retained.
The public v1.95.1 download stores provider API keys in Roma's local settings, not macOS Keychain. Use a separate key you can revoke and set a provider spending limit when available.
Automatic connections
This website measures visits with Vercel Web Analytics. v1.95.1 checks Roma's GitHub releases for updates and beingpax.github.io for announcements. The Mac app does not automatically send usage or crash reports to me, so I may not know about a problem unless you contact me.
Why Roma asks for broad access
Here is how v1.95.1 uses each permission.
Revoke any permission in macOS System Settings. The related feature will stop working.
If Roma causes a problem
Contact me directly.
You can report a problem in your own words. Investigating it is my responsibility. Reporting, debugging, and retesting are optional for you.
Start with a short description. Do not send a raw recording, transcript, API key, or unredacted screenshot. If I need evidence, I will ask for the smallest useful piece and explain how to remove private details. I will tell you what I can reproduce. When a fix ships, I will name it in the release notes.
For a privacy or security concern, email me before opening a public issue. I will never publish your private message or personal data without permission. After the risk is addressed, I may publish a technical summary with no identifying information or quotes from your private message.
Felix C. Order, builder and support