Privacy policy
In short. Dictation (Czech, Slovak, English) runs entirely on your phone. Your voice and the recognised text are not sent anywhere. With your consent, the floating microphone uses Android's accessibility service only to put text into the field you dictate into; it stores and sends nothing of it. The app has no accounts, no ads and no analytics. It goes online only to download the speech-recognition model file, and in doing so it checks only whether the connection is metered.
1. Who is responsible
The app is published by Jan Rosecký, ID number (IČO) 74849999 (brand: Moravian AI). Privacy questions: info@moravian.ai.
Place of business
Zámezí 301/14, Jehnice, 621 00 Brno, Czech Republic
Because your voice and text never reach us, we hold no personal data about you as a user of the app.
2. What the app processes, and where
- Microphone audio. The app listens only when you start dictation. It processes the audio on the phone, in memory, and does not save it: no file, no recording. The memory that held the audio is zeroed after recognition. Working data of the recognition model that is derived from the audio (for example a spectrogram) stays in the phone's memory until the model is unloaded, which the app does about two minutes after dictation ends, or when Android stops the app. The app does not write it to disk.
- Recognised text. It is inserted only into the field you are dictating into. If it cannot be inserted, the keyboard holds it in memory and offers to insert, copy or discard it; with the floating microphone it waits for at most 5 minutes in a bubble by the button. It reaches the clipboard only if you tap “Copy”. The text is not stored, not written to logs and not sent anywhere.
- Android voice input. The app is also offered on the phone as a voice-input service. When another app asks for voice input and you choose this one, the recognised text goes to the app that asked for it. What that app does with it is governed by its own policy.
- Settings and the model. The app's settings (for example whether to listen as soon as the keyboard opens, vibration, the length of the pause that ends a sentence, or the position of the floating button) and the downloaded model files (the speech-recognition model, about 474 MB, and the language model that joins sentences, about 137 MB), or a partly downloaded part, stay on the phone. Both are in the app's private storage, are excluded from Google backup and are not transferred to a new phone. The app keeps no history of your dictation.
3. What leaves the phone
The only thing the app does over the network: it downloads two files: the speech-recognition model and then a language model that joins sentences split by a pause. Both are on the same storage, which the app's publisher rents from Hetzner Online GmbH (Object Storage in the Falkenstein data centre, Germany, EU; the address ends in fsn1.your-objectstorage.com); the second file adds no other recipient. The download of both (one step of the setup guide) starts only when you tap “Download model” in the setup guide. If a partly downloaded file is left after an interruption, the download resumes by itself when you return to the setup guide, or when you tap “Continue downloading”. On a metered connection (for example mobile data) the app asks first and downloads only with your consent; you can also choose to wait for Wi-Fi, and the download then starts by itself once the phone joins an unmetered network, as long as the setup guide is open. If you close the setup guide, a download under way may go on until Android stops the app; no other connection is made meanwhile. The downloaded part is kept and the download resumes from it when you return to the setup guide. After that, dictation works offline.
Before and during the download the app checks only whether the active connection is metered. It reads nothing else about the network and sends nothing.
With each download the storage necessarily sees your IP address and the usual request data (time, file address, the name of the speech-recognition library czechspeech-android with its version number, and connection details). It learns nothing about your speech, your text or you from this, because the app sends no such data. Hetzner runs the storage for the publisher as a processor; the publisher does not use this technical data and keeps no access logs of the file. How Hetzner handles it is set out in its policy (hetzner.com/legal/privacy-policy).
Legal basis for this transfer of technical connection data: our legitimate interest in delivering the model file you asked to download (Article 6(1)(f) GDPR); without your IP address and connection data the file cannot be delivered.
4. What the app does not do
- It has no user account and no sign-in.
- It has no payments or purchases.
- It shows no ads.
- It does not measure usage (no analytics) and sends no crash reports.
- It shares nothing, passes nothing to third parties and sells nothing.
- It does not store recordings, transcribed text or a dictation history.
- It uses the accessibility service for nothing but the floating microphone (see section 7): it does not control other apps on its own, changes no settings and does not read the screen beyond the field you type in.
5. Permissions and why the app asks for them
- Microphone. Dictation is impossible without it. The app uses it only during a dictation that you start.
- Network (internet). Only for the one-time download of the two model files, see section 3. No other code in the app communicates over the network.
- Network state. The app checks one thing: whether the active connection is metered (for example mobile data). That lets it ask before downloading the model on such a connection and wait, if you choose, for an unmetered network. It reads nothing else about the network (no network name, addresses or location) and sends nothing.
The app does not request the accessibility service of the floating microphone as a permission: you turn it on yourself in Android's accessibility settings, see section 7.
The app requests no permissions other than these three. (The permission list on Google Play also shows an internal signature permission of the AndroidX library, which serves only communication within the app and which the user does not grant.)
6. The keyboard and Android's warning
When you turn the keyboard on in settings, Android warns that a keyboard may be able to read everything you type, including passwords and card numbers. The warning is generic and applies to every keyboard. For this app it means the following:
- The keyboard is for dictation only: it has a microphone, a space key, a delete-last-sentence key and an action key. It has no letter keys, so it does not see what you tap.
- Like every keyboard, however, it receives details about the text field it is in. It reads the text before the cursor (at most 1,000 characters), one character after the cursor and any selected text, plus the field type (for example password, number), its label and the name of the app the field is in. It needs this for correct capitalisation and spacing, to delete the last dictated sentence, and to know where to insert text.
- If you turn on “Průběžný náhled” (live preview) or “Tečky na místě, kam přijde text” (dots where the text will go) in the settings (both are off by default), the keyboard on Android 13 and later also reads a snapshot of the field around the cursor: the text before and after the cursor, as much as the app with the field provides, and the position of the cursor and of the selection. It uses it to check where the temporary grey text stands in the field, so that it can remove that text safely and delete nothing else.
- What it reads from the field stays briefly in the phone's memory. It is not stored, not written to logs and not sent anywhere.
- In password and PIN fields, and in fields whose label mentions a payment card, the keyboard refuses to dictate. Recognising a field by its label is a best guess and may not catch every such field.
7. The floating microphone and the accessibility service
The floating microphone is a button that appears by the field you are typing in, above your own keyboard (Gboard, for example). You tap it, speak, and the text goes into the field. To make that work with another maker's keyboard, the app uses Android's accessibility service (AccessibilityService). Only you can turn it on, in Android's accessibility settings. The app sends you there only after the setup guide has shown you what the service reads and why, and you have confirmed with the “Souhlasím” (I agree) button. You can turn it off there at any time. The app is not a tool for people with disabilities and uses the accessibility service only for the floating microphone.
What the service reads, and why:
- the list of windows on the screen, to recognise that a keyboard is open and where it is;
- for the field you are typing in: which app it belongs to, where it is on the screen (so that the button does not cover it), whether it is editable, its input type and label, and whether it is a password field (so that it never dictates into passwords or payment-card fields);
- when it inserts dictated text, the field's text and the cursor's position, for spacing, capitalisation and the right place to insert; when the field empties, it recognises that the message was sent and ends the dictation;
- only if you turn on “Přestat poslouchat a odeslat” (stop listening and send; off by default): the labels of the buttons right next to the field, to find the Send button among them, and it presses that button when you tap;
- on Android 13 and later the system attaches it to every field you type in, much like a keyboard, and gives it the field's details including the text around the cursor (except passwords). The app uses this only to insert dictated text into the field you are dictating into; it does not read other fields' details.
Nothing the service reads is stored, written to logs, sent anywhere or analysed. It remembers only the position of the button on the screen. Text that cannot be inserted waits in a bubble by the button for at most 5 minutes, in memory only. Apart from inserting text, and pressing Send when you tap (with the option above), the service does nothing in other apps: it changes no settings, controls nothing on its own and reads nothing else from the screen. The button does not appear in password and card-number fields, on the lock screen, or in this app.
8. Children
The app is not directed at children. It collects no personal data of children, because it collects no personal data at all.
9. How to delete everything
All of the app's data is on your phone only. Uninstall the app, or open Android settings, Apps, Přepis řeči CzechSpeech, Storage, Clear data. Your settings and the downloaded models disappear with it. Turn the floating microphone off in Android settings, under Accessibility. We hold no data of yours that would need deleting.
10. Your rights
Because we keep no personal data about you, there is nothing to access, correct or erase. You can still write to info@moravian.ai. You also have the right to lodge a complaint with the Czech Office for Personal Data Protection (uoou.gov.cz) or the authority of your own country.
11. Changes to this policy
This policy describes the first version of the app: the floating microphone, the keyboard and the system voice-input service. A later version with user accounts will get its own updated policy, published here before that version is released. We will list every change here with its date; the policy at this address is always the one in force.
Version 1.0, 20 November 2026: first release.