tl;dr
On-device speech-to-text means the model that converts audio into text runs on your computer or phone rather than on a server. The phrase covers that one step and says nothing about the cleanup pass most dictation tools apply afterward, which often calls a server. You do not have to take anyone's word for either step: a ten-minute offline test shows where both run. A fresh Auri Voice install passes the transcription half offline; the cleanup pass needs the network until you set it to Auri Local.
What the phrase covers, and the step it skips
On-device means the model that turns your voice into words runs on your hardware: after a one-time download, recognition works with the network off, latency depends on your machine rather than your connection, and audio does not travel for that step. Local speech recognition explained covers the models, the downloads, what they cost, and why iPhone keyboards default to cloud more often than Mac apps.
The step the phrase skips is cleanup. After recognition produces text, most tools strip filler, resolve corrections you spoke aloud, and fix punctuation, and that second model has its own location. An app can transcribe on your device and still send every transcript to a server for cleanup.
The ten-minute offline test
The test needs no tools beyond your network switch. On a Mac, turn Wi-Fi off in Control Center, since macOS has no airplane mode; on an iPhone, airplane mode works.
- Finish setup and download whatever local model the app offers.
- With the network on, dictate the sample paragraph below and save the result.
- Turn the network off and dictate the same paragraph. Save that too.
- Still offline, try each AI feature the app has (rewrite, cleanup, formatting) and note which ones fail.
- Reconnect and check the settings screen for anything that switched back on its own.
Dictate this paragraph both times, fillers included:
"Hi Dana, um, I looked at the draft this morning, I mean yesterday morning, and the intro feels too long. Can we cut it to two paragraphs and, uh, move the pricing table to page three? I can review again on Tuesday, sorry, Wednesday, at 3pm."
It carries what cleanup acts on: filler, two spoken corrections, a number, and a time.
How to read the result
The comparison has three possible outcomes, and each one tells you where a step runs.
Dictation fails offline. Nothing appears, or the app shows a connection error. Recognition is remote: your audio goes to a server every time you dictate.
The text arrives but stays rough. The ums survive and "Tuesday, sorry, Wednesday" lands exactly as spoken. Recognition is local and cleanup is remote: audio stays home, and the transcribed text goes out whenever the network is up.
The text arrives clean. The offline output matches the online result: "Hi Dana, I looked at the draft yesterday morning and the intro feels too long. Can we cut it to two paragraphs and move the pricing table to page three? I can review again on Wednesday at 3pm." Both steps run on your machine.
What each outcome looks like in Auri Voice
Auri Voice transcribes on your Mac by default. The cleanup pass sends the transcribed text (never audio) to the provider you pick — Auri Cloud unless you change it; choose Auri Local in AI settings to keep both steps on the Mac (Apple Silicon). The test shows that split directly:
- A default install lands on the second outcome. The default engine (Fast Multilingual Dictation on Apple Silicon, Lightweight, Any Mac on Intel) transcribes the paragraph offline, and the transcript arrives without the cleanup pass, because Automatic cleanup resolves to Auri Cloud and cannot reach it.
- Pick Auri Local in AI settings and rerun the offline paragraph: the third outcome. The fillers are gone and only Wednesday remains.
- Switch the Voice engine to Fast, Accurate, No Download (Auri Cloud) and offline dictation fails with "Auri Cloud transcription could not be reached": the first outcome, for that engine only.
If you want the third outcome, two settings get you there: any local Voice engine plus Auri Local cleanup. The offline dictation guide covers working that way full time.
Which network traffic is your speech
Even an app with a fully local writing path talks to the internet. The table sorts the traffic you might see:
| 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 |
| Cloud cleanup or rewrite | Yes, your transcribed text is sent |
| Cloud transcription engine | Yes, your audio is sent |
Only the last two rows carry your words. When you audit an app, find the switch that controls each of those rows; in Auri Voice the Voice engine picker controls the audio row and the AI provider list controls the text row.
What the test cannot show
The test proves where things run today, on your machine, with your settings. It cannot show what a server does with text it legitimately receives: how long the text is kept, who can read samples, whether it trains models. Those are policy questions, and the retention guide lists five worth sending to any vendor, along with the answers that should worry you. Keep the vendor's replies next to your test note, and rerun the ten-minute test after major app updates, since defaults can change.
FAQ
Does on-device mean nothing at all leaves my machine?
No. It describes where the recognition model runs. Model downloads, license checks, and crash reports still use the network, and a cloud cleanup pass sends your transcribed text out after local recognition. Test each step separately.
How do I confirm cleanup is local too?
Dictate the same paragraph online and offline, then compare. If the offline version keeps the filler and spoken corrections that the online version removed, the cleanup pass needed a server.
What does the test show on a fresh Auri Voice install?
Transcription passes offline, because the default engine runs on the Mac. Cleanup does not, because Automatic resolves to Auri Cloud. Choose Auri Local in AI settings (Apple Silicon) and the offline paragraph comes back clean.
Will I notice the difference between a local and a cloud model?
On everyday writing, most people cannot. The gap shows up on unusual accents, heavy background noise, and rare terminology. Judge with your own paragraphs rather than with published benchmark scores.
Why dictate a fixed paragraph instead of speaking freely?
So you can compare like with like. Seed the paragraph with filler and a spoken correction; whether those survive offline tells you if the cleanup pass ran.