The actual difference

Cloud transcription uploads your audio to a server, runs a large model on it, and sends text back. On-device transcription runs a smaller model on the hardware in your hand. That is the whole difference, and everything else follows from it — the accuracy, the speed, the offline behaviour and the privacy claim are all consequences of that one choice.

The important word is required. On Apple's platforms, an app asking for speech recognition can ask for it on-device, or it can ask and accept whatever it gets — and the second one silently uses the server whenever the device cannot do the job. Both can be described as "on-device" in marketing copy. Only one of them is.

What it costs

Being honest about the trade is more useful than pretending there is not one:

Accuracy. An on-device model is smaller than a server one. On clear speech the gap is small; on names, jargon, heavy accents and overlapping voices, a cloud model usually wins.
Language coverage. Not every language has an on-device model on every device, and older hardware carries fewer of them.
Battery and heat. The work happens on your device, so a long clip costs battery. On a laptop you will hear the answer.

What you get back: it works on a plane, it works with no signal, there is no upload to wait for on a large file, and your audio never becomes someone else's training data or someone else's breach.

How to check whether an app means it

Marketing copy is not evidence. Three checks that are:

Turn off the network and try. A genuinely on-device feature works in airplane mode. A cloud one fails, or hangs, or produces an error that mentions a connection.
Watch what happens on an unsupported device or language. An app that truly requires on-device recognition will tell you the feature is unavailable. An app that says "on-device" and quietly succeeds anyway was never requiring it.
Read the privacy label for "Audio Data". An app that uploads audio has to declare it, and the declaration outlives the landing page.

What Contentbuilder does

Captions are requested with on-device recognition explicitly required, and the code checks the device supports it before starting. There is no cloud path in the app for it to fall back to — so on a device that cannot do it, the feature reports unavailable rather than uploading your clip. That is the honest version of the trade: sometimes you get no captions, and never a surprise upload.

The related tools work the same way. Silence removal is audio analysis on the file itself, nothing leaves. AI overlay text uses Apple Intelligence on device, and drops to a set of offline suggestions when the model is not available.

What we can prove lists every outbound request either app can make, including the ones that exist.