What runs on the device
Local speech recognition means the model that turns sound into words sits on your machine. You download it once, then inference happens on hardware you own, which is why it keeps working on a plane.
- The model file lives on disk, typically a few hundred megabytes to several gigabytes.
- Transcription uses your CPU, GPU, or Neural Engine rather than a data center.
- No connection is needed for the transcription step once the download finishes.
- You choose the engine, which makes the privacy claim something you can point at in settings.
- The first session after a launch can lag while the model loads into memory, then settles.
That last point about choosing the engine is the practical difference from a hosted service. A setting you selected yourself is easier to verify than a sentence on a marketing page, and it does not change when the vendor updates their infrastructure.
Local transcript and local cleanup are separate claims
Modern dictation tools do two jobs. First they transcribe, then they improve the text: filler removed, false starts dropped, punctuation and paragraphs added. Each job can run locally or remotely, and vendors often describe only the first one.
- Local transcription, cloud cleanup: audio stays home, your draft text does not.
- Cloud transcription, local editing: audio leaves, the text stays.
- Both local: nothing leaves for either step.
Auri Voice on Mac can keep both steps on-device. Transcription uses the Voice Engine you pick, and text improvement follows your AI Processing setting: Automatic prefers on-device AI when available and otherwise uses Auri Cloud; you can also choose Auri Cloud or bring-your-own-key providers. Cloud speech and cloud cleanup are choices, not requirements.
What it costs you
Local processing trades convenience for control in ways that are easy to predict. Pick based on the machine you have and the work in front of you.
| Factor | Local engine | Cloud engine |
|---|---|---|
| First run | Model download, then ready | Immediate |
| Storage | Hundreds of megabytes and up | Almost none |
| Battery and heat | Noticeable on large models | Server side |
| Latency | Depends on your hardware | Depends on the network |
| Languages | Limited to what you installed | Usually broader |
| Offline | Works | Stops |
| Audio privacy | Audio stays on the device | Audio is uploaded |
If your Mac is a few years old, test one long session before you standardize on the largest model. Fan noise during a call is its own kind of cost.
Engine choices on Mac
On Mac you can pick the engine per session rather than living with one compromise. Auri Voice exposes seven Voice Engine choices in settings.
- Fast English (Parakeet TDT v2): the Auri Voice pick for English-only work with higher accuracy and custom vocabulary.
- Fast Multilingual (Parakeet TDT v3): when you switch languages in a normal day.
- Low-Latency Multilingual (Nemotron 3.5 ASR): when turnaround matters more than peak accuracy.
- Lightweight Any Mac (Whisper Base): a smaller local download for older or shared machines.
- High-Quality Local (Whisper Large Turbo): when noise, accents, or hard vocabulary still force edits.
- Auri Cloud: no download, 100+ languages, needs a network.
- Apple On-Device: built-in speech with no extra download.
Download local engines on Wi-Fi before a trip and confirm one offline session while you still have time to fix it. Custom vocabulary teaches names and uncommon words so you stop retyping the same corrections.
iPhone works differently
Phone constraints are real. Battery, thermal limits, and the way iOS sandboxes keyboards all push voice features toward servers, which is why local-first is more common on Mac than on iPhone.
- Custom keyboards cannot reach the network at all until you allow Full Access.
- On Auri Keyboard, users can choose among selectable cloud transcription engines, or On-Device English-only transcription.
- Cloud transcription is the common default on iPhone; local is an explicit option.
- Apple's own dictation runs on device for many languages and uses servers in other cases.
A workable split for private work: capture on Mac with a local engine, and treat the phone as the tool for messages you are comfortable processing in the cloud.
Verify a local claim in five minutes
Four checks separate a real local mode from a marketing adjective. None of them requires technical tooling.
- Find the engine picker. A genuine local mode names the model and shows a download.
- Turn on airplane mode and dictate a paragraph after the download completes.
- Try the cleanup or rewrite feature offline. If it fails while dictation works, that step is remote.
- Read what the docs say about the text step as well as the audio step.
- Check whether an account is required. A tool that will not start without a login is doing something over the network, even if transcription is local.
Record the result in one line: engine name, offline transcription confirmed, cleanup local or cloud. That note is what you will want the next time you dictate something you would rather keep to yourself.
FAQ
What does “local after download” mean?
You fetch the model files once over the network, and every session after that runs on your own hardware. The first download needs a connection; the dictation itself does not.
Which local engine should I start with?
On Mac, start with Fast English Dictation (Parakeet TDT v2) if you mostly speak English, or Fast Multilingual (Parakeet TDT v3) if you switch languages. Move to Whisper Large Turbo when noise or accents still force edits, or Whisper Base if you want a lighter download. Auri Cloud covers 100+ languages without a download; Apple on-device is the built-in fallback.
Do local models slow my Mac down?
Larger models use more CPU and GPU while they transcribe, which shows up as heat and fan noise in long sessions. Smaller engines cut that cost and are usually good enough for clear speech.
Can I mix local transcription with cloud cleanup?
Yes, and plenty of tools do it by default. The audio stays on your machine and the transcript goes to a server. Decide whether that trade is acceptable per draft rather than once for everything.
Is on-device recognition less accurate?
Sometimes, especially on unusual accents and noisy audio, because hosted models can be much larger. The gap has narrowed enough that most people notice storage and battery before they notice accuracy.