Local vs cloud dictation: where does your text go?

Local dictation turns your speech into text on your own device. Cloud dictation uploads the audio and does it on someone else's server. If privacy, flaky Wi-Fi or a subscription bothers you, local is the better default on a modern Mac; cloud makes sense when you need the same tool on every platform or a feature only a server model offers.

That label only covers one step, though. Cleanup, AI rewriting, history and the app you type into each have their own data path, and a tool can be "local" for one and "cloud" for another. This page shows you how to trace all of them.

We make Sayburst, a Mac dictation app that recognizes speech on the Mac. Facts about other products come from their own sites and docs, checked October 31, 2026.

#Local vs cloud at a glance

The price row has a simple cause. A cloud service pays for compute every time anyone speaks, so it bills monthly. A local app uses your Mac's chip, so many of them charge once: Sayburst is $29, VoiceInk is $25 for one Mac, and MacWhisper Pro is €64.

#Real examples, from the vendors' own pages

• Wispr Flow is cloud only. Its data controls page says "Transcription always occurs on the cloud," and its help docs say a new dictation can't start without a connection. More in our Wispr Flow privacy explainer.
• Superwhisper is both. Its free plan runs local Whisper models, and its docs say nothing you dictate leaves your device on that plan. Pro adds cloud voice models, and the Mac setup creates a first mode that uses cloud models, so check which one your modes use (plans).
• Apple Dictation depends on your settings. Apple says the text under Keyboard settings tells you whether general dictation is processed on your device, and that dictating into a search box is excluded from that (Apple's guide).
• MacWhisper is local first, with optional cloud providers you connect with your own API keys.

The pattern: "local" and "cloud" describe a setting as often as a product. The same app can do both.

#Speed and accuracy

The old assumption was that local meant worse. On Apple silicon that's no longer true for everyday dictation. In our benchmark on an M4 Max, using a clean read-speech sample from LibriSpeech, Parakeet v3 had a 0.46% word error rate and took a median 82 ms to transcribe a dictation; Whisper Large v3 Turbo scored 0.83% at 320 ms. That's clean audiobook speech, not your kitchen or your accent, so treat it as a sign that local models are good enough, not as a promise.

Cloud services can run very large models and update them without you downloading anything. Local apps avoid the round trip and keep working on a plane. The only accuracy test that matters is yours: test dictation accuracy on your Mac with your own voice, vocabulary and microphone.

#Follow your words through five stages

A tool can be private at stage two and still send text at stage three. That isn't necessarily bad. A cloud writing model on your text can be useful, as long as you chose it and know what's sent. Text and audio are also different disclosures: sending a cleaned sentence to a language model is not the same as uploading your voice.

#How Sayburst splits the stages

• Capture: recordings are deleted after transcription unless you choose to keep them.
• Recognition: on your Mac, with Parakeet or Whisper models downloaded once in the app and checksum-verified. Sayburst never uploads audio.
• Cleanup: on your Mac. Fillers, repeats, capitals, punctuation and spoken "new paragraph" or "bullet point" are handled instantly, and the optional Sayburst Writer model runs on the Mac's GPU with no account or key.
• Optional writers: Ollama or LM Studio stay on your Mac. If you add your own key for a cloud provider such as OpenAI or Anthropic, only the transcribed text (plus any context you switch on, all off by default) goes to it, and Sayburst shows exactly what will be sent first.
• History: searchable, with retention from forever down to one day, or off.
• Network that isn't dictation: model downloads, activation and license checks. A purchased license keeps working offline for up to 60 days between checks.

Our privacy page and the full privacy policy document the details.

Here's what local looks like in practice. With Wi-Fi off on a train, using Sayburst's notes shortcut:

"call the bank about the, uh, rate lock, it expires on the twelfth, no, the thirteenth, and ask about the appraisal fee"

What gets appended as a timestamped bullet under today's date in your Markdown notes file:

Call the bank about the rate lock. It expires on the 13th. Ask about the appraisal fee.

Nothing about your mortgage went anywhere but a file on your Mac.

#The destination is a separate decision

Dictating into a local text file and dictating into a shared web document can use the same recognizer. What happens next is up to the receiving app. Once text lands in Google Docs, Slack or an AI chat, that product stores and syncs it under its own rules. Local recognition keeps your voice off the internet; it doesn't make a cloud document private.

#An offline test you can repeat

1. Install the app, download its model, and finish setup while online.
2. Turn off Wi-Fi and unplug Ethernet.
3. Dictate some harmless sample text into TextEdit and check that recognition and cleanup both work.
4. Separately try any feature you rely on, such as rewriting or file transcription, and note which ones stop.
5. Reconnect.

This tells you what works without a network. It doesn't prove what an app sends when it is online; that takes a network monitor and reading the vendor's documentation. For more on dictating without a connection, see offline dictation on Mac.

#When cloud is the right call

Choose cloud dictation if you need the same tool and settings on Windows, iPhone and Android, if your Mac is an Intel model that local apps don't support well, or if your organisation has signed an agreement with a specific provider. Choose local if your words are sensitive, your connection is unreliable, or you'd rather not pay monthly for compute your Mac can do itself.

Sayburst is the local option for Mac only: it needs Apple silicon and macOS 14 or later, and has no Windows or phone version. For more head-to-heads, see the compare page.

Our take

On an Apple silicon Mac, local dictation is fast and accurate enough that cloud is no longer the default. Pick cloud for reach across platforms; pick local for privacy, offline work and a one-time price. Either way, trace all five stages, not just the label.

Questions people ask

What is the difference between local and cloud dictation?

Local dictation runs the speech model on your own computer, so the audio never has to leave it. Cloud dictation uploads your audio to a server, which turns it into text and sends it back, so it needs a connection.

Is local dictation as accurate as cloud dictation?

It can be. Current local models run well on Apple silicon; in our own benchmark on clean read English, Parakeet v3 had a 0.46% word error rate. Accuracy on your voice, accent and vocabulary is what counts, so test both on your own speech.

Does offline dictation mean no data ever leaves my Mac?

No. Local recognition covers one step. Optional cloud rewriting, license checks, model downloads, synced history and the app you type into can each involve the network, so check each stage.

Does Sayburst upload my recordings?

No. Sayburst recognizes speech on your Mac and never uploads audio. If you connect a cloud writing provider with your own key, only the transcribed text and any context you switch on is sent, and Sayburst shows exactly what before you enable it.