Disclosure: this article is published by Relay, and author Raymond Zhu is a founding member. It is a documentation-based explainer, not a hands-on benchmark. Relay is a cloud-based product and is not recommended here for offline use. Third-party features and prices were checked against official pages on October 29, 2026 and can change. Sources: Apple Dictation (macOS 14 guide), Apple Dictation (current guide), Relay Privacy Policy.
Offline is two questions, not one.
Most dictation tools have at least two stages. Speech recognition turns audio into raw text. An optional second stage, AI cleanup, formatting or rewriting, turns that raw text into something you would send. Some tools add a third input, context such as a screenshot or window text.
A product can run the first stage locally and the second in the cloud. That combination can be perfectly reasonable, but it is not offline, and your words still leave the computer as text. Check every stage before you describe a setup as local, to yourself or to a client.
Apple Dictation: the built-in local option.
Apple’s current Dictation guide says Keyboard settings show whether your dictation is processed on your device. Its earlier macOS 14 guide stated that requests are processed on the device for supported languages with no internet connection required, and that requests in languages without on-device support are sent to Apple. Check Keyboard settings on your own Mac rather than relying on an older page. Apple’s feature-availability information notes that on-device dictation requires downloading speech models.
On macOS 27, Apple’s current guide adds an on-device model with Apple Intelligence that improves spelling, punctuation and capitalization, for English on Macs with M3 or later and at least 12GB of memory. For setup steps, see how to dictate on a Mac.
Sources: Apple Dictation (macOS 14 guide), Apple Dictation (current guide).
Local-model apps, and where they use the cloud.
Superwhisper’s plans page says it is free with local Whisper models, which its voice-model page says run entirely on your device. Its language-model page lists one local rewrite model, S1-mini, on Pro for English Message and Mail modes; other rewrite models are cloud.
MacWhisper processes transcription locally, and its AI features can use a local model through Ollama or LM Studio, or cloud providers with your own keys. VoiceInk is open source and processes transcription locally; its optional Cloud Enhancement sends text. Handy is free and open source and says it works completely offline; check where any optional post-processing runs in its settings. Spokenly offers free unlimited local transcription and a Local Only Mode.
All of these were checked on October 29, 2026. Prices and features change, so confirm the edition you install.
Cloud tools that don’t work offline.
Wispr Flow’s documentation says an internet connection is required for transcription. Aqua Voice’s FAQ says cloud dictation needs an internet connection and that offline availability varies by platform, app version and mode, so check the desktop app before relying on it offline.
Relay is also cloud-based. Its Privacy Policy explains that audio is compressed on your Mac and sent to an AI provider for transcription, and that screenshots and on-screen context, when used, go to AI providers with the request. Relay doesn’t store them on its servers, but that is a different promise from offline. How Relay handles context and privacy has the details.
Sources: Relay Privacy Policy.
How to check a local setup yourself.
Turn off Wi-Fi and unplug Ethernet, then dictate into the apps you use. If it still works in every mode you rely on, the core path is local. If one mode fails, that mode uses the network.
Do this after the app has been set up and its models downloaded. Some local apps need the internet once to fetch a model, which is fine; what matters is whether everyday dictation needs it.
Then read the settings for each mode: the speech model, any AI enhancement or rewrite model, context features such as screenshots, and history or cloud sync. A product can be local by default and cloud in a single mode you forgot you set up. An outbound firewall that shows connection attempts gives you a further check if your work requires it.
When Relay is still the right choice, and when it isn’t.
If your organization’s policy, or your own judgement, says audio must never leave the computer, Relay is the wrong tool and we would rather you knew that up front. Use a local setup and verify it as above.
If cloud processing is acceptable and you want context-aware text and Relay Skills in your Mac apps, Relay is worth a trial on its own terms. Read the provider list in the Privacy Policy before using it with sensitive material.