Local dictation processes speech on your device, while cloud dictation sends audio or text to remote servers. Local systems can reduce cloud dependence and work offline once models are downloaded. Cloud systems can offer cross-device services, centrally managed models and lower hardware needs. Many apps are hybrid, choosing local or cloud per mode, so check each step.
Local, cloud and hybrid dictation: what each means
Turning speech into text takes a speech model. Where that model runs, and where the optional cleanup step runs, define the architecture. Speech recognition is only one step: an app can also check a licence, look for updates or send analytics without sending your audio anywhere.
- Local (on-device) dictation: Microphone → local model → text → destination app. Speech recognition runs on your device and the model is stored there, so recognition needs no internet once the model is downloaded.
- Cloud dictation: Microphone → network → cloud model → text → device. Audio goes to remote servers that run the model and return text, so a connection is required.
- Hybrid dictation: Microphone → local or cloud model, depending on the mode → text. You or the app choose per mode, so the architecture depends on your settings.
| Architecture | Speech recognition runs | Formatting and cleanup runs | Internet required? | Where the models live | What may still go online |
|---|---|---|---|---|---|
| Local | On your device | On your device, or an optional cloud step | Not for recognition, once models are downloaded | On your device, after a download | Model downloads, licence checks, update checks, optional analytics and any optional cloud features |
| Cloud | On the vendor's or a provider's servers | Usually on the same servers | Yes | On the vendor's servers | Audio, text and context for processing, plus account and usage data |
| Hybrid | On your device or on servers, by mode | Local or cloud, by mode | Depends on the mode | On your device and on servers | Whatever the cloud modes send, plus the local list |
Does local dictation mean private?
Not automatically. Local speech processing can reduce how much raw audio goes to a vendor's servers, which is a real difference. Privacy also depends on everything else the app does.
A cloud service is not automatically unsafe either. Vendors publish retention and training terms, and some offer settings that limit what is stored. The useful question is which steps leave your device, and what the vendor says happens to them.
- Telemetry: an app may send usage or crash events, and whether they are opt-in or on by default differs by app.
- Licensing: activating or checking a licence usually makes a network request, even in a local app.
- Update checks: apps look for new versions online.
- Accounts: sign-in ties usage to an identity, and some local apps need none.
- Optional AI and cloud features: a local recogniser can pair with a cloud step that rewrites the text, which sends the text out.
- The destination app: see the next section.
Local dictation does not make the app you are typing into local
Dictation architecture is separate from how the destination app handles your text. If a dictation app inserts text into Claude, ChatGPT, Gmail, Slack or Cursor, those apps may send that text to their own servers, under their own terms.
- The dictation app decides where your speech is recognised and what else it sends.
- The destination app decides what happens to the text once it is inserted: it may store, sync or send it to its own servers.
- Check both. Read the dictation app's data-flow statements and the privacy terms of the app you dictate into.
Is local transcription more private?
For the recording itself, usually yes. Local transcription turns audio into text on your own device, so the recording is not uploaded to a server to be transcribed. That removes one exposure, not all of them. The app can still go online for models, licences, updates or optional cloud features, and you are responsible for storing and securing the files.
Built-in operating-system dictation can be mixed: it may process some languages on the device and send others to the vendor's servers, so check its settings for your language.
Pepys makes one product of each kind, and they handle your words differently. Pepys Transcribe is cloud software for recordings you upload. Pepys Dictate is a Mac app for live dictation and for files and memos transcribed on the Mac.
Neither approach suits everyone. A cloud service can offer sharing, AI summaries and access from any browser; a local app keeps the audio on one machine. For recordings that must stay confidential, see the guide to confidential transcription.
| Topic | Pepys Transcribe (cloud) | Pepys Dictate (local) |
|---|---|---|
| Where speech is processed | In the cloud: audio is split into chunks and sent to a third-party speech-to-text provider | On the Mac: recognition and formatting run locally, with no cloud transcription |
| The recording | Deleted automatically about 30 days after upload; you can delete it sooner | Not uploaded; kept on the Mac for History, then deleted on the schedule you choose |
| The text | The transcript is kept until you delete it | Stored in a local history database on the Mac |
| Training | Pepys does not use your content to train its own models | Audio and text are not uploaded; an optional setting, off by default, can send a transcript to an AI provider you configure |
| Account | A free account is required | No account |
Does local dictation work offline?
Often, but not by default and not for every step. Offline here means speech recognition keeps working without a connection, after any required models are on the device.
- Models usually need a download first. Superwhisper's model page lists one-time downloads of 75 MB to 3 GB for local models, and Pepys Dictate's speech and formatting models are about 1.3 GB on first launch.
- Licence checks differ. Pepys Dictate activates a paid licence key online once and checks it at launch when online. Other apps' pages in the table below do not describe it.
- Cloud formatting can fail even when local transcription works. VoiceInk's cloud options and AI enhancement need a connection, while its local transcription does not.
- Hybrid tools lose some features offline. Superwhisper says local models work offline, so a mode that uses a cloud model does not.
- Read the app's privacy page or security page. Look for the words "processed", "servers" and "retained", and note whether it names providers.
- Look for a data or privacy setting that lets you opt out of your data being used to improve models.
- Turn off Wi-Fi, open a text field and dictate. If text still appears, recognition ran on the Mac. If it fails, the app likely needs a connection for that step.
- If the policy and the offline test disagree, ask support in writing which steps run on the device.
Is local dictation faster, and is cloud dictation more accurate?
Neither is true in general. Local processing can avoid a network round trip. Cloud processing can use larger managed models and powerful remote compute, but depends on the network. Actual latency depends on the model, hardware, network and implementation.
Accuracy works the same way. Where recognition runs does not by itself decide it. Model size, training data, the domain you dictate in, your microphone, language, accent, background noise and decoding settings all matter. The published comparisons we found are written by vendors that sell one side, and they disagree on the size of any gap. This page ranks neither, and independent benchmarks are still missing.
What local and cloud dictation need from your hardware
Running a model on your own machine moves the work from a vendor's servers to your device. Cloud dictation lowers local compute needs but depends more on the network and the service being available.
- Storage: local models are downloads, from 75 MB to 3 GB in Superwhisper's docs to about 1.3 GB for Pepys Dictate.
- Chip and system: Pepys Dictate needs an Apple silicon Mac on macOS 13 or later. Whisper Notes needs Apple silicon and macOS 14 on the Mac. VoiceInk recommends Apple silicon and warns that local models do not work reliably on Intel Macs.
- Memory and speed: Superwhisper's model page gives no specific RAM requirement, and recommends cloud models for older machines because they are fast on any hardware.
- Model size is the core constraint. A review of on-device recognition research says conventional recognisers could need a language model of up to 100 GB, which made them impractical on a device, and that fully neural designs need much smaller memory footprints.
How local, cloud and hybrid dictation fail
Each architecture fails in its own ways. No probabilities are claimed here, because none have been measured.
- Local failure modes: the model is not downloaded yet, there is not enough memory or storage, the hardware is too old for the chosen model, or the local app crashes.
- Cloud failure modes: no internet, a service outage, high latency, service or rate limits, or an account or authentication problem. Cloud speech services publish limits: Google's Cloud Speech-to-Text lists 900 recognition requests per 60 seconds per project, and says requests beyond a limit produce an error.
- Hybrid failure modes: a hybrid app can inherit both lists, depending on which mode you use.
Where some apps sit today
These rows come from each vendor's own pages, read on 8 October 2026. They describe architecture, not quality, and they are not ranked or priced. If local, offline or privacy is your buying criterion, see our comparison of private, local and offline dictation apps.
| Product | Speech processing | Text and formatting | Offline mode | Cloud option | Account | Main architecture |
|---|---|---|---|---|---|---|
| Pepys Dictate | Local: speech recognition runs on your Mac | Local: formatting runs on your Mac | Yes, once its models are downloaded; a paid licence key is activated online once | No cloud transcription. An optional advanced setting can send a transcript to an AI provider you configure; it is off by default | No account | Local-first dedicated dictation |
| VoiceInk | Local by default: its FAQ says transcription runs locally; optional cloud transcription uses your own API key | Your choice: VoiceInk Refine runs on-device; AI enhancement sends text to the provider you select; Ollama can keep it local | Yes after setup: its README says processing is 100% offline; cloud options need a connection | Yes, optional: cloud transcription and AI enhancement with your own API key | Not stated in public docs | Local-first dedicated dictation |
| MacWhisper | Local: its homepage says it transcribes locally on your Mac with local AI models | Optional AI clean-up uses your own AI key (cloud providers) or a local model such as Ollama or LM Studio | Yes: its homepage lists an Offline feature for local processing | Yes, optional: remote models and providers with your own API keys | Not stated in public docs | Local transcription suites with dictation |
| Whisper Notes | Local: its site says transcription runs on the device and audio never leaves it | Not stated in public docs | Yes: its privacy page says no internet connection is required | None stated | No account, per its FAQ | Local transcription suites with dictation |
| Superwhisper | Your choice per mode: local voice models or cloud models | Your choice per mode: a local language model (S1-mini), a cloud language model, or no AI processing | Yes with local models: its model page says local models work offline; cloud models need a connection | Yes: cloud voice and language models (Pro), or your own keys | Not stated in public docs | Hybrid and model-selectable |
| Wispr Flow | Cloud: its data-controls page says transcription always occurs on the cloud | Cloud service: it uses the app name and surrounding text for formatting and punctuation | No: transcription always occurs on the cloud, per its data-controls page | Yes: cloud only | Sign-in is a setup step, per its setup guide | Cloud and cross-platform dictation |
| Willow Voice | Cloud: its help center says Willow uses cloud-based AI models | Cloud models; optional Context Awareness reads small bits of text around the cursor | Paid plans: its help center lists offline dictation under the Individual Pro plan, and its homepage says it is powered entirely by your device | Yes: cloud models by default | Sign-in required, per its help center | Cloud and cross-platform dictation |
A worked example: Pepys Dictate
Pepys Dictate is local-first AI dictation software for Mac. Its privacy page says speech recognition and formatting run on your Mac, with no account and no cloud transcription, and that audio is never uploaded. The speech and formatting models are about 1.3 GB and download on first launch.
Pepys Dictate goes online to download models, activate a licence key, check the licence at launch when online, and check for updates. Usage events are opt-in and contain no text or audio. An optional advanced setting can send a transcript to an AI provider you configure, and it is off by default.
Pepys Dictate does not claim that nothing ever crosses the network. It also does not control what the app you dictate into does with the text.
Which architecture fits you?
Choose by your constraints, not by a label. Each of these is a fair choice for some people.
- Choose local if offline use matters, you prefer speech to be recognised on your own device, you work on one device, or you want less dependence on a vendor's servers.
- Choose cloud if you need the same dictation across devices, team administration, centrally managed models, hosted history or sync, or lower local hardware needs.
- Choose hybrid if you want a local default with optional cloud models, a different quality or speed trade-off per workflow, or the freedom to switch per mode. Superwhisper is one example.
- For how Apple's own Dictation behaves offline, see offline dictation on Mac. For wider comparisons, see the roundup of dictation apps, the one-time purchase options or Wispr Flow as a cloud example.
Last updated November 8, 2026.