Offline dictation on Mac
Seven apps, read against each vendor’s own privacy policy or docs on 2026-10-13, plus a test you can run yourself.
Mac dictation apps that work without sending your audio to the cloud, by default: MacWhisper, VoiceInk, Superwhisper with a local model, Voice Type, and Apple Dictation on an Apple silicon Mac in a supported language. Wispr Flow and Typeless send audio to their servers every time; neither documents a local mode.
Most local apps can also send audio or text to a cloud provider once you add a key or pick a cloud model. The table lists both, in each vendor’s own words, checked 2026-10-13.
| App | Audio by default | Optional paths off the Mac | The vendor’s own words |
|---|---|---|---|
| Apple Dictation | Depends on the Mac and the language. On Apple silicon, in a language Apple lists for on-device dictation, general text stays on the Mac. On an Intel Mac or in other languages, audio goes to Apple. | Audio is stored by Apple only if you turn on Improve Siri and Dictation. Dictating into a search box can send the text to the search provider. | When you dictate on an Intel-based Mac or in a language that doesn’t support on-device dictation, your dictated utterances are sent to Apple to process your requests.Apple: Dictate messages and documents on Mac (macOS 14 guide), Apple: Ask Siri, Dictation and Privacy |
| Wispr Flow | Leaves the Mac. Recognition runs on Wispr’s servers. | No local recognition mode is documented. | Transcription always occurs on the cloud. This is the best way for us to provide accurate, low latency transcription.Wispr Flow: Data controls, Wispr Flow: Security and compliance FAQ |
| Typeless | Leaves the Mac, along with the name of the app you are in and relevant text from it. | The policy says audio may also go to third-party AI providers, naming Microsoft and Google. Typeless says audio is discarded after the result returns. | Both your audio and contextual data are processed in real time on our cloud servers.Typeless: Privacy policy (updated 2026-10-13), Typeless: Data controls |
| Superwhisper | Stays on the Mac when a local voice model is selected. | Cloud voice models, chosen per mode, receive the audio plus your vocabulary words as recognition hints. | This can happen entirely on your device using a local voice model, or via a cloud voice service if you choose one.Superwhisper: Sensitive data docs |
| MacWhisper | Stays on the Mac. | A cloud transcription provider (Groq, ElevenLabs, Deepgram) receives the audio. DeepL translation and AI prompts send the transcript text. | By default, transcriptions (including Speaker Identification) is all performed locally on your Mac or iOS device.MacWhisper: Keeping transcriptions private |
| VoiceInk | Stays on the Mac. | If you add a key for a cloud provider, the audio file goes to that provider. | By default, all transcription processing happens entirely on your device using local AI models.VoiceInk: Privacy policy (updated 2026-09-27) |
| Voice Type | Stays on the Mac. There is no code path that uploads audio. | Optional text rewrite, off by default, sends the finished transcript text to the provider you pick. Never audio. | Checked in the app’s Swift source. Details below. |
Apple’s macOS 14 guide gives the rule. On a Mac with Apple silicon, “Dictation requests are processed on your device for supported languages” and “no internet connection is required.” On an Intel Mac, or in a language without on-device support, “your dictated utterances are sent to Apple.” Dictating into a search box may send the text to the search provider.
The supported list is on Apple’s macOS Feature Availability page under On-Device and Modeless Dictation. It includes English for the US, UK, Canada, Australia and India, and carries a note that it “requires download of speech models.” The current macOS guide no longer spells the rule out; it tells you to read the text under Dictation in Keyboard settings, which says whether your dictation is processed on the Mac or sent to Siri servers.
When audio does go to Apple, Apple’s privacy notice says it is not stored unless you opt in to Improve Siri and Dictation. The request history, including transcripts, is tied to “a random, device-generated identifier that rotates multiple times per hour” and Apple “may retain and use this data for up to two years.” Apple Dictation is free. If your Mac and language qualify, it is the local option you already have.
Thirty minutes of dictation is this much audio on the wire, depending on the codec the app uses:
At a 1 Mbit/s uplink that is 41 seconds of upload as Opus and 4.6 minutes as WAV. A local app uploads zero bytes. The WAV figure is exact arithmetic; the Opus and AAC sizes were measured on synthetic speech and will run somewhat larger on a real voice. We do not know which codec each vendor uses, so read these as a range, not a figure for any one app. Script and raw samples are on the measurements page (bench/payloadsize.py).
nettop -P -p "Voice Type" (use the app’s process name, as shown in Activity Monitor). It lists that process’s bytes in and out. Dictate for a minute. A cloud app’s bytes out climb with every sentence; a local app’s stay flat.Recognition runs on the Mac with whisper.cpp. Audio is never sent anywhere; the app has no code that uploads it. This was checked in the shipping Swift source, which makes exactly these network calls:
huggingface.co when you first set up the app or switch models. The app refuses any other host for model files (ModelArtifactManifest.swift). After that, recognition needs no network.files.carelesswhisper.app/review-gate.json, which we host. The request has no body and no identifier beyond the default macOS user agent; the response is a build number and a version field (Defaults.swift). Run curl -s https://files.carelesswhisper.app/review-gate.json to see all of it.llmRewriteEnabled = false in WhisperState.swift) and pick a cloud provider with your own key, the app sends that provider your rewrite prompt, the model name and the transcript text (LLMRewriter.swift). The Rewrite Selected shortcut sends the text you selected instead, capped at 50,000 characters by default. The request goes straight from your Mac to the provider. Apple On-Device (macOS 26), Ollama and a local script keep the rewrite on the Mac, though a script you write can call anything.And these are absent:
Where Voice Type loses
Cloud recognition can run models far larger than a laptop holds. We have not measured word error rate for Voice Type or any other app here, so this page makes no accuracy claim in either direction. A local app also costs disk space and memory for the model, and the first run needs a download. If your audio is not sensitive and the best possible transcript matters more than where it is processed, a cloud app may serve you better. If you are on an Apple silicon Mac in a supported language, free Apple Dictation already keeps general text on the Mac.
Every vendor page was read on 2026-10-13. Policies change; the date is part of the claim.
Try it on your own work
Voice Type is on-device dictation for macOS. Press a shortcut, speak, and the text goes into the app you are typing in. Your audio never leaves your Mac. Free for 7 days, then $19.99 once on the Mac App Store. No subscription and no account.
Start the free trial