tl;dr
Some voice typing tools upload your audio to a server and some transcribe it on the device; it depends on the engine behind the mic button. Apple's dictation runs on device for many languages and falls back to servers in other cases. Google-style and cloud AI dictation upload by design. Auri Voice on Mac transcribes locally by default and uploads audio only if you pick its cloud engine, while its cleanup pass sends transcribed text to Auri Cloud unless you switch it to Auri Local. Auri Keyboard on iPhone uses cloud engines by default, with an English-only On-Device option; its default Raw mode sends nothing for cleanup.
The matrix
Nothing on screen tells you where audio goes: the same microphone icon can mean a downloaded model on your device or a stream heading to a data center. Transcription and cleanup are also separate questions, because a tool can transcribe locally and still send the transcribed text to a server; local speech recognition explained covers that split. The table answers both questions for the common paths, so check the row that matches the tool and setting you are on today.
| Path | Audio uploaded for transcription? | Transcript sent for cleanup? |
|---|---|---|
| Auri Voice on Mac | No with the default local engines; yes if you pick the Auri Cloud engine | Yes, to Auri Cloud by default; Auri Local (Apple Silicon) keeps cleanup on the Mac |
| Auri Keyboard on iPhone | Yes on the cloud engines; no with On-Device (English only) | No in Raw mode, the default; Clean, Concise, and Professional send it to Auri Cloud |
| Apple Dictation (iPhone and Mac) | On device for many languages, servers in other cases | No separate cleanup pass; punctuation is added during dictation |
| Google-style keyboard dictation | Typically yes, with some on-device support on Android hardware | Varies by keyboard |
| Cloud AI dictation apps | Yes, that is the design | Yes, on the same servers |
| Local Mac engines (Parakeet, Nemotron, Whisper, and similar) | No, after the model download | Set by whichever app hosts them |
The two Auri rows carry the same brand and answer differently. Auri Voice and Auri Keyboard are separate apps on separate platforms, and the defaults reflect what each device can do.
The two Auri rows, spelled out
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 Voice engine picker in settings controls the audio column: five local downloads, Apple's built-in model, and Auri Cloud. The AI provider list controls the text column, including bring-your-own-key OpenAI, Anthropic, and Google.
Auri Keyboard's audio column depends on the dictation engine, set under Keyboard Dictation > Recognition: the cloud engines (Ultrafast, the recommended default, plus Legacy, Azura, Supernova, and Wizper) upload audio to transcribe it, while On-Device keeps audio on the phone and works in English only. The text column depends on the output mode. Raw, the default, types the literal transcript and sends nothing for cleanup; Clean, Concise, and Professional send the transcript to Auri Cloud and return finished text:
Raw mode types: "hey sorry i missed this can we do thurs instead"
Clean mode types: "Sorry I missed this. Could we move it to Thursday instead?"
Hold the keyboard mic to switch modes. To check any row on your own machine, run the ten-minute offline test in the verification guide: dictation failing offline means recognition is remote, and text arriving rough means cleanup is.
Apple and Google, with the nuance intact
Both platforms have shipped on-device speech recognition, and both still use servers in some situations. Summaries that flatten this into “Apple is local, Google is cloud” will mislead you at the moment it matters.
- Language packs and device generation change whether the local path is available.
- Since iOS 16, Apple Dictation punctuates automatically as you speak; that happens as part of dictation rather than as a separate cleanup request.
- The switch between local and server processing happens quietly, with no interface change to warn you.
You can see the setting on your own device: on iPhone, Settings > General > Keyboard, under Dictation; on a Mac, System Settings > Keyboard, in the Dictation section. Read Apple's dictation privacy page for the OS version you are running. If you need a guarantee rather than a general tendency, use a tool where you pick and download the engine yourself.
Cloud AI dictation apps
The apps that clean up rambling speech into finished writing usually run large models on servers. That is how they get their quality and their speed on modest hardware, and it is a reasonable trade as long as you make it knowingly.
- Expect microphone permission plus an account or an API key.
- Expect two requests per message: one carrying audio for transcription, one carrying text for cleanup.
- Retention varies widely. Some discard audio immediately; others keep samples for quality work.
Before you dictate client material into one of these, read the audio section of the policy rather than the homepage.
Rules for the drafts that matter
Not every message needs the same treatment. Sort by content and the decisions get easy.
- Health details, legal matters, unreleased plans, and client names: local transcription and local cleanup. On a Mac that means Auri Voice with a local engine and Auri Local selected in AI settings.
- Ordinary work email: cloud is fine if you know the retention terms.
- Anything you cannot verify: type the sensitive part by hand and dictate the rest.
- Passwords and card numbers: type them, and only on the system keyboard.
If you never look this up again, keep one habit: check which engine is selected before you dictate something you would not email to the vendor.
FAQ
Does Apple Dictation always stay on device?
No. Apple documents on-device dictation for many languages and devices, with server processing in other cases. Check the Dictation section under Settings > General > Keyboard on iPhone, or System Settings > Keyboard on a Mac, and Apple's dictation privacy page for your OS version.
If it works in airplane mode, is it local?
That is strong evidence for the transcription step. Some tools cache a local model for offline use and switch back to a cloud engine once you reconnect, so recheck the engine setting after the test.
Do rewrite features upload my audio?
Usually not. They send the text, which is a separate request from transcription. An app can transcribe on your device and still send that transcript to a server for cleanup, so ask about both steps.
What should I ask a vendor?
Whether raw audio leaves the device, how long anything is stored, whether content is used for training, and whether a fully local mode exists. Ask about the transcription step and the cleanup step by name.
Is cloud transcription a bad choice?
Not by itself. It is often more accurate and works on older hardware. It becomes a problem when you did not know you were using it, so check the default engine before you rely on a tool.