Does VoiceScroll Send Your Voice to a Server?

Any app that listens while you talk is worth asking one question about: where does that audio actually go? For a teleprompter, the microphone is on for the whole take, not just a few seconds, so the answer matters more than it does for a quick voice command.

The short version

On iPhone and iPad, VoiceScroll's speech recognition stays on your device by default: the app forces on-device processing and never switches to a server on its own, not even as a fallback. The one exception is Online recognition, which is off until you turn it on. With it on, a language whose speech model isn't on your device is recognized by Apple's servers and needs an internet connection; languages whose model is installed stay on the device. On Android, VoiceScroll asks for on-device recognition by default and doesn't pick a server-based recognizer by itself. If your device can only recognize the language you picked by sending the audio away, the app tells you instead of quietly doing it. For that language you can try Online recognition, which sends your audio to Google and needs an internet connection.

What "on-device" actually means

Speech recognition can run two ways. Cloud recognition streams your audio to a server, which does the transcription and sends text back. On-device recognition runs the model locally: your voice never has to leave the phone in order to be understood. The tradeoff is usually accuracy and language coverage, cloud models tend to be bigger, but for a teleprompter that's already reading a script you wrote, on-device is normally accurate enough, and it doesn't depend on your connection at the venue.

iPhone and iPad: on-device by default

By default, VoiceScroll's recognition request on iOS is configured to require on-device processing. That's not a preference the app can silently override if the network happens to be available, it's a hard requirement set on the request itself. This default applies to both of VoiceScroll's recognition engines, the classic one and the newer one introduced with iOS 26. If a language doesn't have an on-device model on your device, the app tells you rather than quietly switching to a server, and you can download the model or turn on Online recognition. Whether a session uses Online recognition is decided when it starts, so a session never moves to a server partway through.

Android: on-device by default, and the one limit worth stating

Android's speech APIs work differently from iOS's. Android does offer a strictly on-device recognition API, but restricting VoiceScroll to that one alone would rule out devices that do recognize speech offline through a different provider, so the app keeps the wider set of recognition services and, by default, refuses everything that isn't offline. Providers that only work online are dropped from the candidate list entirely, both in the first pick and in the mid-session failover chain, and every request is pinned to prefer offline. If a language is only served remotely on your phone, the app blocks Start and opens the voice settings sheet explaining why, rather than switching to a server on its own to make it work. That sheet also offers Online recognition, which you can try for that language; it sends the audio to Google and needs an internet connection.

An earlier release had a real gap, and it's worth stating plainly. The app preferred on-device recognition but did not refuse the remote path: when the leading offline recognizer died mid-session, on a start timeout or a no-transcription watchdog or an unsupported-language error, the failover chain could land on an online provider, rebuild the request with the offline flag turned off, and hand the rest of that take to Google's speech service with nothing on screen to say so. That was fixed in version 1.10.1, released in July 2026. Online providers are now rejected everywhere the app picks a recognizer, so the failover chain has nowhere to fall to. Android's Online recognition setting, added in September 2026, is different: it stays off until you turn it on, it applies only to languages your phone can't recognize offline, and it's chosen before a session starts, so a session never switches to a server partway through.

And here is the limit worth stating: because VoiceScroll uses that wider set rather than the strict API alone, the offline preference it pins on every request is a hint those recognition services can ignore, not a platform guarantee. Excluding online providers and pinning that flag is the strongest lever the app has on Android. It is an app-level guarantee, not the categorical promise iOS can make.

What about the script itself?

Separately from audio, your script text isn't collected or uploaded either. The one exception is Online recognition on iPhone and iPad: words in your script identified as another language are sent to Apple with the audio, as recognition hints. Scripts are stored locally on your device, and VoiceScroll doesn't require an account to use, so there's no server-side profile your scripts could be attached to. If you record video, the same applies: video and audio are written to a file on your device, not streamed anywhere.

How to check this for any app, not just VoiceScroll

This isn't specific to teleprompters, it's worth checking for any app that listens to you regularly:

• Look for the words "on-device" specifically in the privacy policy or app description. "We take your privacy seriously" without a technical claim usually means nothing.
• Check whether the app works with the network turned off. If it does, that's a strong sign recognition is actually local.
• Watch what an app does when it can't recognize locally. Refusing the language and saying so, or disclosing a fallback while it happens, are both defensible. Switching to a server without telling you is not.

Bottom line

On iPhone and iPad, your voice stays on the device by default. If you turn on Online recognition, Apple's servers recognize only the languages whose speech model isn't on your device. On Android, VoiceScroll asks for on-device recognition by default and doesn't pick a server-based recognizer by itself, though on Android that is an app-level guarantee rather than a platform-level one. If your device can only handle a language remotely, the app tells you instead of switching to a server, and you can choose to try Online recognition, which sends that audio to Google. Either way, your scripts aren't collected, and there's no account tying any of it back to you.