On-device vs cloud speech recognition: what changes

What is the actual difference?

A speech recognizer is a model that takes audio and returns text. In a cloud service, that model runs on someone else's server: your Mac records you, sends the audio, and waits for the words to come back. On-device, the same kind of model is downloaded once and runs on your own processor, so the audio goes from the microphone into memory and no further.

Everything else follows from that one fact. The model is not magic in either place, and a recognizer small enough to run on a laptop is now good enough for everyday writing.

Where does my voice go?

This is the difference that does not shrink. A cloud recognizer has to receive your audio to do its job. What happens to it after that is set by the provider's policy: how long it is kept, whether people or training pipelines ever see it, and which country the server is in. Those policies can be good, but they are a promise you read rather than a fact you can check.

On-device, there is nothing to promise, because the audio is never sent. Dictation hears things that were never meant to be written down: a name you said while thinking, a phone call in the background, a sentence you cancelled. Keeping that on the Mac is simpler than deciding who you trust with it.

In Dictera, every speech model runs locally, and the audio is transcribed as it arrives and never written to disk as a recording, either. The security page lists every host the app contacts, and what Dictera sends over the network shows how to watch the connections yourself.

Is cloud recognition more accurate?

Sometimes, and it is worth being specific about when. A large server model has more capacity, so it tends to hold up better on the hard cases: heavy accents, noisy rooms, crosstalk, and long-tail vocabulary such as drug names or obscure place names. Some cloud services also let you send a list of expected words, which nudges the recognizer toward them.

For a clear voice, a decent microphone and ordinary sentences, the gap is small, and much of what people notice as quality is not recognition at all. It is punctuation, capitalization and the "um" that stays in. Those are a second step, and they can be done well locally. Two models per dictation explains how Dictera splits hearing the words from writing them down.

The fairest test is your own voice. Dictate a paragraph you actually write, with the names and numbers in it, and compare. A benchmark measured on audiobooks tells you little about your meeting room.

Is on-device dictation faster?

It removes the network from the wait, and that matters more than the raw speed of the model. A cloud round trip adds upload time and a trip to the server; on a train or a busy café connection it adds a lot, and sometimes the request just fails. On your Mac, the only wait is the computation.

What shapes the wait most, though, is not where the model runs but whether it streams. A streaming recognizer transcribes while you speak, so the text is mostly written by the time you let go of the key. A batch recognizer starts on the whole recording after you stop, so the pause grows with what you said: barely there for a sentence, noticeable for three paragraphs. Cloud services come in both kinds, and so do local models.

Of Dictera's four models, Apple Speech Analyzer and Nvidia Nemotron stream; Parakeet Ultra 0.6B and Alibaba SenseVoice transcribe after you stop. Choosing a speech model goes through the trade-off model by model.

Does it work without internet?

On-device, yes, once the model is on the Mac. That covers the plane and the train, and it also covers the quieter case: the office network that blocks unfamiliar services, or the client site where you would rather not ask.

A cloud recognizer does nothing without a connection, and a weak one shows up as slow or missing text rather than as an error you can plan around.

In Dictera, the network is needed to download a model, once. The models from NVIDIA, Moondream and Alibaba come from Hugging Face; Apple's language files come through macOS itself, the way it installs any system language asset. After that you can turn off Wi-Fi and dictation, cleanup and translation carry on.

What does it cost my Mac?

Running the model yourself is not free, and this is the honest cost of the local approach.

• Disk. A downloaded speech model takes roughly 630 MB to 1.5 GB. Apple's engine is part of macOS, so it takes nothing extra in Dictera.
• Memory. A model has to be loaded to run. Dictera's Keep Model Ready setting holds the active one in memory so the first dictation starts straight away; turn it off if memory is tight and accept a short wait instead.
• Hardware. Local recognition leans on the chip. Dictera needs a Mac with Apple silicon and macOS 26 or later; an older Mac without Apple silicon cannot run it.

A cloud service moves all of that to the server. Your Mac only records and uploads, which is why cloud dictation can run on almost anything with a browser.

What about languages?

Here the two approaches are closer than they look. Large cloud services cover a long list of languages; so, now, do local models. Dictera covers up to 100+ spoken languages depending on the model: Parakeet Ultra 0.6B, Moondream's model built on NVIDIA's Parakeet v3, handles 25 European languages, SenseVoice is built for Chinese, Cantonese, English, Japanese and Korean, and Nemotron covers the rest.

The detail that matters more is how a model decides which language it is hearing. Some follow a setting, some work it out from the audio, and some let you pin one. Apple's engine, for instance, transcribes in the language your Mac is set to. That choice changes your results far more than whether the model sits on a server.

Can I teach it my words?

This is one place where cloud services often have a real feature that Dictera does not: a vocabulary hint sent with the request, so the recognizer is more likely to hear "Kubernetes" than "Cooper Nettie's".

Dictera does not bias the recognizer. Its custom replacements work after it: when the model writes a phrase you have listed, the text that lands in your document is your spelling instead. That fixes a name the model gets wrong the same way every time. It does not help a word the model mishears differently each time.

Who decides when the model changes?

A cloud provider can update its model on the server, and your results change the next morning without anything on your Mac changing. Usually that is an improvement. Occasionally a word it always got right starts coming out differently, and there is nothing to roll back.

A local model changes when you change it. Dictera's downloaded models stay as they are until an app update or your own choice replaces them, and you can keep more than one and switch between them from Quick Controls in the notch. Apple's engine is the exception in between: it is local, but macOS keeps it up to date.

When is cloud recognition the better choice?

When the hard cases are your everyday cases: a language or accent no local model handles well, audio recorded in difficult rooms, or vocabulary so specialised that a hint list is the only thing that helps. Also when the machine cannot run a model at all. And when your organisation already has a contract with a provider that covers what happens to the audio.

On-device is the better fit when what you dictate is private, when you work where the connection is poor or not allowed, and when you want the same sentence to come out the same way next month. For most people dictating email, notes and messages, those are the conditions that apply.