What the phrase covers
On-device speech-to-text means the acoustic model that turns your voice into words is running on your hardware. The audio is captured, processed by a downloaded model, and written into a text field without a server involved in that conversion.
What that gets you:
- Recognition works with the network off, once the model is downloaded.
- Microphone audio does not travel for the transcription step.
- Latency depends on your machine rather than your connection.
- An outage or a bad hotel network does not stop you writing.
Those are real benefits and they are narrower than the phrase implies in casual use. On-device describes one step in a pipeline that usually has more than one. When you read the term on a product page, treat it as an answer to the audio question and as silence on everything else, which is why the next section matters more than this one.
The step the phrase usually leaves out
After recognition produces text, most voice typing tools do something to that text. They strip filler words, drop false starts, resolve corrections you spoke aloud, add punctuation, and format lists. That work is done by a second model, and the second model has its own location.
Three arrangements are common:
- Local recognition, cloud cleanup. Your audio stays home and your sentences travel.
- Cloud recognition, cloud cleanup. Everything travels, usually for quality and convenience.
- Local recognition, local cleanup. Nothing needs to travel for the writing path.
Auri Voice on Mac is built for the third arrangement by default, so the polished paragraph that lands in your document was produced by your Mac. Wispr Flow and Monologue are known for turning speech into clean writing, and that polish commonly involves a cloud model; Monologue also has a local mode where the output sits closer to raw transcription. Knowing which arrangement you are running changes what you can safely dictate.
A downloaded model does not mean silence on the wire
Even with a fully local writing path, apps talk to the internet for reasons unrelated to your speech. Confusing those calls with transcription leaks leads people to distrust local tools for the wrong reasons, or to trust them for the wrong ones.
| Traffic | Related to your dictation? |
|---|---|
| Model download and updates | No, one time per engine |
| License and account checks | No |
| Crash reports and analytics | No, though check what they contain |
| Optional cloud rewrite you enabled | Yes, your text is sent |
| Cloud transcription mode | Yes, your audio is sent |
The distinction you care about is whether your words leave, and the last two rows are the only ones where they do. When you audit an app, note which switch controls each of those rows, then decide with the switch positions in front of you rather than from a general impression of the vendor.
How to test any on-device claim in ten minutes
This test needs no tools beyond airplane mode, and it separates marketing language from behavior on your machine today.
- Complete setup and download whatever model the app offers.
- Dictate a fixed paragraph with the network on. Save the result.
- Turn on airplane mode and dictate the same paragraph. Save that result.
- Compare them for filler removal, punctuation, and formatting.
- Turn each AI feature on and off, repeating the offline test each time.
Interpret the comparison carefully. If offline fails entirely, recognition is remote. If offline succeeds but comes back rough, recognition is local and cleanup is remote. If offline output matches the online output, both steps are running on your machine. That third outcome is what Auri Voice aims for on Mac, and it is easy to confirm in a couple of minutes rather than taking anyone's word for it.
Why platform limits shape the defaults
The reason on-device is more common on Mac than in an iPhone keyboard is architectural. Apple runs keyboard extensions in a constrained environment, and a large speech model plus a cleanup model is a heavy thing to load into that space every time you tap the mic.
What follows from that:
- Mac apps can reasonably ship multi-gigabyte engines and run them for every dictation.
- Keyboard extensions favor smaller models or a cloud call, which is why cloud transcription is a common default on iPhone.
- Local options on iPhone exist and involve tradeoffs in speed or model size. Auri Keyboard offers local transcription.
- Full Access is the iOS permission that makes any cloud feature possible in a keyboard.
Use the platforms for what each does well. Sensitive drafting on the Mac with local recognition and local cleanup, quick replies on the phone with the default, and a deliberate switch to local transcription on the phone for the occasional message that deserves it.
Questions worth asking a vendor
If documentation leaves you unsure, ask directly. These questions are specific enough that a straight answer is possible, and a non-answer tells you something as well.
- Which steps run on device by default on each platform you support?
- Does the text cleanup step require a network connection?
- What is sent when I enable each cloud feature: audio, text, or both?
- How long do you retain anything sent, and can staff review samples?
- Where is transcript history stored and does it sync?
Keep the replies with your test notes. Together they cover what a test cannot see, such as retention, and what a policy cannot promise, such as behavior in the version you installed. A vendor that answers all five plainly is easier to defend to a security reviewer than one with a page full of privacy adjectives and no per-feature detail.
FAQ
Does on-device mean nothing at all leaves my machine?
No. It describes where a model runs. Accounts, licensing, crash reports, and any feature you enable can still use the network. Test each feature separately.
Is on-device recognition less accurate?
Sometimes, and the gap has narrowed enough that most people cannot tell on everyday writing. Judge with your own paragraphs rather than with published benchmark scores.
Why is a downloaded model so large?
Because the whole speech model has to be present locally. Sizes run from a few hundred megabytes to several gigabytes depending on the engine you pick.
How do I confirm cleanup is local too?
Dictate the same paragraph online and in airplane mode, then compare quality. If the offline version is much rougher, the polish step needed a server.
Why do iPhone keyboards default to cloud more often?
iOS runs keyboard extensions under tight memory limits, which makes large local models hard to host there compared to a Mac app.