Nilo Privacy Policy
Nilo is dictation that runs on your phone or your computer. There is no Nilo account, cloud transcription or advertising. Nilo does not upload your speech or text. Model-download providers, Google Play and (for desktop update checks) GitHub may receive technical request, identifier and diagnostic data as described below. Text you insert into another app or explicitly export is handled by that destination.
The short version
| What | Where it stays |
|---|---|
| Microphone audio and shared voice messages | In your phone's memory while you dictate, use Listen to speaker audio or transcribe a voice message you shared to Nilo, then discarded. Never uploaded, never written to storage. |
| Transcribed text | On your device only. Optionally kept in a local history you control. |
| Dictionary, snippets, settings | On your device only. |
| Account, email, contacts, GPS location | Nilo does not request these. Download providers can infer approximate location from your IP address. |
| Download requests and technical identifiers | Hugging Face and its download hosts receive request metadata, including IP addresses. |
| Billing diagnostics and session identifiers | Sent to Google by the Play Billing library, including when purchase UI is disabled. |
| Purchases (if you buy Nilo Pro) | Handled by Google Play. We never see your payment details. |
| Feedback you send | Only if you choose to email it. Your email app sends it; you see everything in it first. |
Your speech and transcripts stay on your device. We do not sell personal data. Third-party technical data handling is described in section 6.
1. Who we are
Nilo is developed and published by KindaSoren ("we", "us"). This policy explains how Nilo's Android and desktop applications (the "apps") handle your information. The sections below describe the Android app; section 16 covers the desktop app. For any privacy question or request, email support@sorens.dev.
2. How dictation works, and what happens to your voice
Nilo transcribes speech on your device using the Whisper model and the whisper.cpp engine. The engine is bundled with the app; speech models are downloaded or imported separately. The microphone is open only while you are dictating or while a Listen to speaker audio session you started is running. Audio is held in memory for the current session and then discarded.
- Audio is never transmitted to us or to any third party.
- Audio is never written to storage by the app.
- Android shows an ongoing notification while the microphone is active, and you can stop dictation at any time.
Listen to speaker audio
“Listen to speaker audio” transcribes a message that another app, such as WhatsApp, plays through your phone's speaker. You start it from Nilo's Home screen or from Nilo's notification. Nilo records what the microphone hears, so it captures the voice of whoever is speaking in that message as well as any nearby sound. It does not read other apps' audio internally, and it does not work with headphones.
- A session ends when you tap Done or cancel it. It transcribes at most two minutes of the message, and the microphone stays on for at most three minutes in total, including the time it waits for the message to start playing.
- The audio is handled exactly like dictation: transcribed on your device, kept only in memory, never written to storage and never uploaded.
- The transcript is not typed into any app or copied to the clipboard until you choose Copy or Paste. If history is on, it is saved to your local history like a dictation.
- Only transcribe audio you have the right to transcribe. The person speaking in the message is not notified by Nilo.
Voice messages you share to Nilo
You can also share a voice message or audio file to Nilo from another app, for example by long-pressing a voice note in WhatsApp and choosing Share, then Nilo. Nilo reads that one file through the permission Android grants for the share, decodes it in memory, transcribes it on your device and then discards the audio. Nilo does not copy or save the file, does not upload it, and cannot read other files. The transcript is handled as described above: it waits for you to copy, share or paste it, and is saved to your local history only if history is on.
3. The Accessibility service
Nilo offers an optional Android Accessibility service that types transcribed text into the field you are already using, so you do not have to switch keyboards. If you enable it, the service:
- Reads the currently focused editable field and basic window state, including which app is in front, only while dictation is enabled;
- Uses that information only to insert the text you dictated and to show or place the dictation bubble (beside the keyboard, or only in the apps you chose for it);
- Keeps which app is in front in memory only; it is never stored or sent. The list of apps you choose for the bubble is a setting stored on your device;
- Ignores password fields and never reads or stores their contents;
- Never logs, saves to disk or uploads existing field contents. Previous field text may be held temporarily in memory to support Undo;
- Never clicks buttons or changes settings; it only inserts or restores text when you request it.
You can turn the service off at any time in Android Settings → Accessibility. If you decline it, Nilo's other insertion methods (the Nilo keyboard and the clipboard) remain available.
4. The Nilo keyboard (optional)
Nilo includes an optional Android keyboard (input method). If you enable it and select it as your keyboard, it composes the text you dictate. It does not log keystrokes or upload what you dictate. Dictated text follows the history settings described below, and the keyboard contains no advertising or tracking code. You can switch back to another keyboard at any time.
5. Data stored on your device
Transcript history (optional)
If history is on, the text of your dictations is stored in the app's private storage so you can review, search and export it. History can be turned off in Settings, individual entries can be deleted, and "Clear history" removes all of it. It is never synced and never leaves your device unless you explicitly export it.
Usage statistics
Nilo keeps counts on your device, such as words dictated, speaking pace and day streak. These are numbers only, stored locally, and are not linked to your identity.
Dictionary, snippets, presets and preferences
Everything you add or configure is stored locally. You can export dictionary entries to the clipboard and transcript history to a file location you choose.
6. Network access and model downloads
Models are downloaded over HTTPS from Hugging Face and its download hosts. These providers receive IP addresses, request times, requested model paths and technical headers. Hugging Face documents approximate country/region information and hashed IP addresses in download analytics, and uses IP addresses for rate limiting. This metadata supports delivery, analytics and platform security under the Hugging Face privacy policy. Model requests contain no audio, transcripts, dictionary, snippets or local dictation statistics. You can avoid this download flow by importing a model file from your device.
Nilo also communicates with Google Play for prices, purchases and ownership checks on startup and resume, even when the Pro purchase UI is disabled. The Play Billing library sends diagnostic events and client-session identifiers to Google, with information such as response/error codes, SDK/app version and device/network context. These events support diagnostics and service improvement under Google's privacy policy. The current app has no switch to disable Billing's diagnostic path. These transfers use encrypted connections and do not include your dictated audio or text.
Feedback you send. “Send feedback” in Settings and “Report” on a transcript open an email to support@sorens.dev in your own email app; Nilo does not send anything itself. The draft contains the app version, speech model, language, insertion method, Android version and device model, and for a report the text of that transcript and the app it went to. You can read, edit or delete any of it, or not send it at all. If you send it, we receive your email address and what you wrote, use them only to reply and to fix problems, and delete them on request.
7. Purchases
If you buy a paid feature such as Nilo Pro, the purchase is processed by Google Play under Google's privacy policy. We receive only a confirmation that a purchase exists, its product identifier and order information needed to grant and restore the feature. We never receive your name, card number, billing address or other payment details. Purchases are tied to your Google account, which is how they can be restored after reinstalling the app or moving to a new phone.
8. What Nilo never does
- No Nilo account creation or sign-in.
- No collection of personal data on our servers — we do not operate a service that receives it.
- No advertising SDKs or tracking pixels. Provider download analytics and Billing diagnostics are described above.
- No selling or renting of personal data. Technical data is transferred to the providers described in section 6; user-requested insertion, export and feedback emails transfer text to the destination you choose.
- No advertising.
- No microphone use while dictation is off.
9. Permissions and why they exist
| Permission | Purpose |
|---|---|
| Microphone | Hear you while you dictate, and hear speaker audio you choose to transcribe. Used only during those sessions. |
| Display over other apps | Show the floating dictation pill above your keyboard. |
| Accessibility service | Optional; insert dictated text into the field you are using. |
| Notifications | Show the microphone-in-use notification and dictation controls. |
| Internet | Download speech models and communicate with Google Play for purchases. |
| Vibrate | Haptic feedback while dictating. |
Each permission is requested in context, and you can deny any of them and still use the parts of Nilo that do not need it.
10. Security
Local data is kept in the app's private storage, protected by Android's app sandbox. Network traffic uses HTTPS. Because we do not run a backend that stores your data, there is no server-side copy of your speech, transcripts or dictionary to breach or to request.
11. Retention and deletion
Transcript history retains at most the latest 500 entries. Older entries are removed as new ones arrive. Other saved data remains on your device until you delete it. You can delete individual history entries, clear all history, disable history, remove dictionary entries and snippets, or uninstall the app, which removes all app data from the device. Nilo opts out of Android cloud backup, so none of this data is copied to Google Drive. If you move to a new phone with a direct phone-to-phone transfer, Android may copy it to the new phone. Because there is no account and no server-side storage, there is nothing further for us to delete. If you believe we hold data about you and want it removed, email support@sorens.dev.
Clearing local data does not delete provider-held download logs or Billing diagnostics. Those providers set their retention periods and handle applicable access/deletion requests under their own policies; we do not control or promise deletion of their records. For Hugging Face privacy requests, contact privacy@huggingface.co. Google provides privacy controls.
12. Children
Nilo is not directed to children under 13. Audio and transcripts are processed locally for all users; technical provider data handling is described in section 6.
13. Third-party components
- OpenAI Whisper and whisper.cpp are MIT-licensed components executed locally on your device.
- Google Play handles purchases and ownership checks; its Billing library also sends diagnostics and session identifiers.
- Hugging Face and its download hosts serve models and process technical request metadata.
- GitHub hosts the desktop release feed and receives anonymous update-check requests.
14. Changes to this policy
If this policy changes, the new version will be posted at this URL with a new effective date. Material changes will also be noted in release notes for the Android and desktop apps.
15. Contact
KindaSoren
support@sorens.dev
Please do not send transcripts, payment details or other sensitive information by email.
16. Desktop app (macOS, Windows and Linux)
The desktop app follows the same principle as the Android app: your speech is transcribed where you are, and your text stays with you. The sections above that describe Android-specific features (Android's Accessibility service and keyboard, and Google Play purchases) do not apply to desktop. This section covers what the desktop app does instead.
Audio and transcription stay on your computer
Desktop dictation uses the whisper.cpp engine bundled with the app. It runs locally as a command-line tool or as a long-lived server that listens only on your computer's loopback address (127.0.0.1). No audio is sent to us or to a transcription service. Microphone audio is held in memory during a session; when the command-line engine is used, it is written to a temporary WAV file that is deleted when transcription ends. Audio files you choose or open with Nilo are read and decoded in memory and are never copied, renamed or uploaded. Only the resulting text is pasted, saved to history or copied.
What is stored locally, and how to delete it
The desktop app keeps its data in its own user-data folder on your computer:
settings.jsonholds your settings, including the dictionary, snippets, presets, the shortcut and hold-to-talk key, the automatic-updates choice, and the most recent transcript when "preserve last transcription" is on.history.jsonholds the optional transcript history — text only, never audio — and local usage counts. It keeps at most the latest 500 entries and is never synced.models/holds the speech-model files you download; transcription works from these local files without a network connection.
In Settings you can delete individual history entries, clear all history,
turn history off, and remove dictionary entries and snippets. Clearing
history also clears the retained most recent transcript. You can remove a
downloaded model by deleting its file from the models folder.
To remove all local data, delete Nilo's user-data folder. There is no
account and no server-side storage, so there is nothing further for us to
delete.
Network connections the desktop app makes
- Speech-model downloads. When you download a model, Nilo fetches it over HTTPS from
huggingface.co(theggerganov/whisper.cpprepository). As with any web request, Hugging Face receives request metadata such as your IP address and the model path. The download contains no audio or transcripts and is checksum-verified before it is installed. - Update checks. Nilo checks for updates on launch and about every six hours. It uses electron-updater's GitHub release feed; installations that cannot update themselves (such as a Linux
.debor an unsigned macOS build) read the same release information through an anonymous request to the GitHub API forKindaSoren/nilo-releases. As with any web request, GitHub sees your IP address and the request details; the API request sends aNilo-desktopuser agent and no account credentials. The "Install updates automatically" setting controls whether an available update is downloaded and installed — it does not disable checking. With it off, Nilo still checks, tells you when an update is available and offers a Download button (or, on installations that cannot update themselves, a link to the release page). - Nothing else. The desktop app contains no analytics, telemetry or crash-reporting service, and apart from model downloads and update checks it makes no outbound data connections. The local speech server talks only to Nilo itself on 127.0.0.1. Opening the release page or macOS Accessibility settings just opens your browser or System Settings; Nilo sends nothing through them.
Clipboard
To put dictated text into another app, Nilo writes it to the system clipboard, sends a synthetic Cmd+V (macOS) or Ctrl+V (Windows and Linux), and then, with "Restore clipboard" on (the default), puts your previous clipboard contents back if nothing else has changed them. If synthetic paste is unavailable or you choose clipboard insertion, the text is simply left on the clipboard for you to paste; "Copy last transcription" also writes it there.
Global keyboard listening
For the shortcut and hold-to-talk, Nilo registers your configured global shortcut and uses an optional global key hook to notice when you press and release the configured hold-to-talk key (for example AltRight or F8). The hook looks only for that one configured key. Nilo does not record, store or send your other keystrokes.
Permissions the desktop app asks for
- Microphone. Nilo asks for microphone access the first time you dictate. If macOS access is denied, only System Settings can change it. Windows controls microphone access for desktop apps in Settings → Privacy & security → Microphone; Linux has no system-wide prompt.
- Accessibility (macOS). Needed only for automatic paste. If it is not granted, Nilo asks once and otherwise leaves the text on the clipboard for you to paste. Windows and Linux have no equivalent gate.