tl;dr
Local speech recognition runs the speech-to-text model on your own device, so audio never leaves it for the transcription step. The claim is narrower than it sounds: most dictation tools also clean up the transcript afterward, and that second step can use a server even when the first does not. 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).
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 transcription runs on hardware you own, which is why it keeps working on a plane.
- The model file lives on disk. Auri Voice's local engines run from about 141 MB to about 1.5 GB.
- 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.
- The first session after a launch can lag while the model loads into memory, then settles.
You also pick the engine yourself, which turns the privacy claim into a setting you can point at. A setting you chose is easier to verify than a sentence on a marketing page, and it does not change when the vendor updates their servers.
Local transcript and local cleanup are separate claims
Modern dictation tools do two jobs. First they transcribe. Then they clean up the transcript: filler removed, false starts dropped, spoken corrections applied, punctuation fixed. Each job can run locally or remotely, and vendors often describe only the first one. The second job is worth locating because it receives the full text of everything you dictate:
You said: "hey can you, um, take a look at the deck before the call, sorry, before end of day"
Auri typed: "Hey, can you take a look at the deck before end of day?"
Three arrangements exist: both steps in the cloud, both on the device, or local transcription with cloud cleanup, where the audio stays home and the transcribed text travels. Auri Voice runs the split by default. Recognition uses the Voice engine you pick, local out of the box. Cleanup uses the provider you pick in AI settings: Automatic resolves to Auri Cloud on every Mac, Auri Local (Apple Silicon) is the on-device option, and you can bring your own key for OpenAI, Anthropic, or Google.
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 | 141 MB to 1.5 GB per engine | Almost none |
| Battery and heat | Noticeable on large models | Server side |
| Latency | Depends on your hardware | Depends on the network |
| Languages | Set by the model you downloaded | Broad, often with auto-detect |
| 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.
Engine choices on Mac
The picker is in Auri Voice's settings under Voice engine. Seven production engines ship, each shown as a card with its name, model, download size, and languages, grouped on-device and cloud:
- Fast Multilingual Dictation (Parakeet TDT v3, ~461 MB, 25 languages): the default on Apple Silicon.
- Fast English Dictation (Parakeet TDT v2, ~443 MB, English only): higher accuracy if you never switch languages.
- Low-Latency Multilingual (Nemotron 3.5 ASR, ~668 MB, 32 locales): when turnaround matters more than peak accuracy.
- Lightweight, Any Mac (Whisper Base, ~141 MB, 99 languages): the smallest download and the default on Intel Macs.
- High-Quality Local (Whisper Large Turbo, ~1.5 GB, 99 languages): for noise, accents, and hard vocabulary; the app warns it needs 8 GB of RAM or more.
- Built-In, No Download (Apple On-Device, 9 languages): Apple's own speech model, requires macOS 26.
- Fast, Accurate, No Download (Auri Cloud, 99 languages with auto-detect): nothing to download, but audio goes to Auri's servers and it stops offline.
A Labs section below them holds two experimental engines, Live English Dictation (Parakeet Live, ~428 MB, streaming English) and High-Accuracy Multilingual (Cohere Transcribe, ~1.5 GB, 14 languages). The engine guide has the full decision logic. Download local engines on Wi-Fi before a trip and confirm one offline session while you still have time to fix it. Custom vocabulary lives under Settings > Dictation > Text > Vocabulary, works with every engine, and the custom words guide covers building the list.
iPhone works differently
iOS runs keyboards inside a small extension, so large local models are hard to host there, and cloud transcription is the common default on iPhone. A custom keyboard also cannot reach the network at all until you allow Full Access, and Auri Keyboard's dictation needs Full Access on.
Auri Keyboard ships six dictation engines under Keyboard Dictation > Recognition: Ultrafast (streams text while you speak, the recommended default), On-Device, Legacy, Azura, Supernova, and Wizper. On-Device is the local one: English only, a roughly 460 MB download, and audio stays on the phone. The rest transcribe in the cloud. Apple's own dictation runs on device for many languages and uses servers in other cases.
Cleanup on the keyboard is its own setting. The default output mode, Raw, types the literal transcript and sends nothing for cleanup; Clean, Concise, and Professional send the transcript to Auri Cloud and return finished text. Hold the keyboard mic to switch modes. For an English message that must stay local, the On-Device engine with Raw output keeps both steps on the phone.
How to verify a local claim
Four checks separate a real local mode from a marketing adjective, and none needs technical tooling.
- Find the engine picker. A genuine local mode names the model and shows a download size; in Auri Voice that is Settings > Voice engine.
- Run the ten-minute offline test in the verification guide: dictate with the network off (Wi-Fi off in Control Center on a Mac, airplane mode on a phone). If dictation fails, recognition is remote; if the text arrives but stays rough, cleanup is.
- 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.
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?
Keep the default: Fast Multilingual Dictation (Parakeet TDT v3) on Apple Silicon, or Lightweight, Any Mac (Whisper Base) on Intel. If you only dictate in English, Fast English Dictation (Parakeet TDT v2) is more accurate. Move to High-Quality Local (Whisper Large Turbo) when noise or accents still force edits; it needs 8 GB of RAM or more.
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. That is Auri Voice's default arrangement: audio stays on your Mac with a local engine, and the transcribed text goes to Auri Cloud for cleanup. Choose Auri Local in AI settings to keep both steps on the Mac, or use a bring-your-own-key provider.
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.