Trust

Private dictation on Mac.

"Private" gets used loosely. This page lays out what each Mac dictation app actually sends where, so you can pick on facts rather than marketing copy. Apple Dictation, Murmur, VoiceInk, MacWhisper, SuperWhisper, Wispr Flow, Aqua Voice.

Verified May 2026 · written by the Murmur team

What "private" means

Three different things, often conflated.

"Private dictation" is a phrase that hides three separate questions:

  1. Does my audio leave my Mac? The audio file is the most sensitive thing. If it's uploaded for transcription, the vendor has a copy of you saying everything you said.
  2. Does my transcribed text leave my Mac? Some apps send the text to an LLM for polish. The text is less sensitive than the audio (no voiceprint, no tone, no pauses) but it's still your words.
  3. Is anything stored or used for training? Even if data is sent, vendors differ on retention. Some delete after processing; some retain for "service improvement"; some explicitly use it for model training.

An app can be private on one axis and not the others. Apple Dictation, for example: audio stays local (good), transcribed text stays local (good), nothing goes anywhere (great). But it has no AI polish step at all (which is why some people don't use it). Wispr Flow does the opposite: audio uploaded, text processed in cloud, polished output is excellent.

App by app

Where each app sends what.

Honest summary, no spin:

  • Apple Dictation: audio on-device, transcription on-device, no LLM step. Fully local once language pack is downloaded.
  • Murmur (Quick mode): audio on-device, Whisper on-device, no LLM step. Fully offline.
  • Murmur (AI mode / Silent mode): audio on-device. Transcribed text and screenshot sent to Claude using your API key. Claude API traffic is not used to train models by default.
  • VoiceInk: audio on-device, Whisper on-device. AI step calls whichever provider you configure (OpenAI, Groq, local Ollama). With Ollama, fully on-device.
  • MacWhisper: audio on-device, Whisper on-device. AI summarisation features call OpenAI / Claude if you opt in.
  • SuperWhisper: audio on-device, Whisper on-device. AI modes call whichever provider you configure.
  • Wispr Flow: audio uploaded to their servers. Transcription and AI polish in cloud. Retention covered in their privacy policy.
  • Aqua Voice: audio uploaded to their servers. Transcription and AI polish in cloud.

For sensitive material, the cluster you want is the first six (audio stays local). The split inside that cluster is whether the LLM polish step is acceptable to you.

Threat models

Pick by what you're worried about.

"I work for a company that processes regulated data." You want audio on-device. Murmur Quick mode, Apple Dictation, or VoiceInk + local LLM. Don't use Wispr Flow or Aqua Voice for work content.

"I dictate about clients but my polish step is fine on Claude." Murmur AI mode is built for this. Audio is local, your Claude API key handles the polish, and Claude API traffic is not used for training by default. You can also redact before triggering AI mode.

"I want zero LLM contact with anything I dictate." Murmur Quick mode, Apple Dictation, or VoiceInk with local Ollama. No outbound text traffic for the LLM step.

"I just don't want my audio in someone's training set." Any on-device app: Murmur, Apple Dictation, VoiceInk, MacWhisper, SuperWhisper. Avoid Wispr Flow, Aqua Voice.

"I trust the major cloud vendors and want the most polished output." Wispr Flow. Read their privacy policy first.

Verifying it

Don't take any of our word for it.

You can verify what an app sends with the Mac built-in Activity Monitor (Network tab) or with Little Snitch, which shows every outbound connection per process. Open the dictation app, dictate a test sentence, and watch what gets uploaded.

For Murmur, you'll see api.anthropic.com when you trigger AI mode or Silent mode (and only then). Quick mode produces no outbound traffic. Update checks hit GitHub Releases. That's the full surface area.

For cloud apps, you'll see traffic to their domain on every dictation. That's expected, that's how they work. The question is whether you're OK with it.

Quick reminder
Murmur is $29 once. Free Quick mode forever. 14-day money-back.
Decision tree

Which one fits.

Audio cannot leave my Mac for compliance reasonsMurmur (Quick)
Audio local, AI polish on my own API keyMurmur (AI)
I want everything on-device including the LLMVoiceInk + Ollama
Free, basic, fully localApple Dictation
Most polished output, OK with cloudWispr Flow
Common questions

Private dictation FAQ.

What's the most private dictation app for Mac?

Apple Dictation, Murmur (Quick mode), VoiceInk, MacWhisper, and SuperWhisper all transcribe on-device. Audio never leaves your Mac. Of these, Apple Dictation and Murmur Quick mode have no AI step at all, so nothing leaves your machine in any form.

Does Murmur send audio to the cloud?

No. Audio is transcribed on-device by Whisper. The audio buffer is held in RAM and discarded. AI mode and Silent mode send the transcribed text (not audio) to Claude via your own API key. Quick mode is fully offline.

Do Wispr Flow and Aqua Voice upload my audio?

Yes. Both transcribe in the cloud, which means the audio file is uploaded to their servers for processing. Their privacy policies cover retention and training-data use, but the upload itself is fundamental to how they work.

Is on-device transcription enough for private dictation?

Depends on what you mean by private. If "private" means "audio doesn't leave my Mac," then yes. If "private" means "no LLM ever touches what I dictated," you need Quick mode (raw transcription) and skip the AI polish step.

How do I verify what an app actually sends?

Use Activity Monitor (Window > GPU Dashboard > Network) or install Little Snitch, which shows every outbound connection per process. Open the app, dictate a test sentence, and watch the network calls. For Murmur, you'll see api.anthropic.com on AI mode triggers and nothing else.

Try Murmur

Local audio. Transparent AI. $29 once.

Murmur is the cheapest dictation app on Mac that keeps your audio on-device while still offering AI polish via your own Claude API key. Quick mode is fully offline; AI mode and Silent mode call your key transparently.