Mac Dictation App Privacy: Where Does Your Voice Actually Go?

Every time you use a voice dictation app, you're sending audio of your voice — your words, your meetings, your private messages — somewhere. The question most people never ask is: where, exactly? Who processes it? How long is it kept? Could it be used to train models?

These aren't hypothetical concerns. They determine whether your dictated medical notes, legal drafts, or private messages stay private. We went through the privacy policies, network behavior documentation, and architecture disclosures of every major Mac dictation app in 2026. Here's what we found.

---

The Three Privacy Tiers

After analyzing each app, they fall into three distinct categories based on how they handle your audio:

• Tier 1 — Direct API (most private): Your audio goes from your Mac directly to a single known API endpoint (OpenAI, Apple). No third-party proxy server. No intermediate storage. You control the API key.
• Tier 2 — Vendor-proxied (partial privacy): Your audio travels through the vendor's own servers before reaching a transcription provider. The vendor can see your audio, may retain it, and may use it to improve their product.
• Tier 3 — Proprietary cloud (least private): Audio is processed entirely on the vendor's infrastructure. Terms of service may permit use for model training. Data retention periods are often indefinite or vague.

Most people assume the apps they pay a subscription for sit in Tier 1. Most don't.

---

App-by-App Analysis

ParlaParla — Tier 1 (Direct API)

ParlaParla uses a bring-your-own-key (BYOK) model. You provide your own OpenAI API key, stored locally in your Mac's app preferences — not on any server. When you dictate, your Mac sends audio directly to api.openai.com. ParlaParla's servers are never in the path because there are no ParlaParla servers.

This architecture eliminates an entire class of privacy risk. There's no vendor database of your audio. There's no intermediary who could be hacked, subpoenaed, or acquired. The data path is: you → OpenAI. Full stop.

OpenAI's API data handling policy (as of 2026): audio submitted via the API is not used to train models by default, and is retained for up to 30 days for abuse monitoring before deletion. You can request zero-day retention by contacting their enterprise team. This is OpenAI's standard policy, the same one that applies when developers call the API directly.

ParlaParla collects no analytics, no crash reports, and no usage data. The app is App Sandbox compliant, meaning its file system and network access are restricted to what Apple's sandbox allows.

Apple Dictation (Built-in) — Tier 1 (Direct Apple)

Apple's built-in dictation has a meaningful privacy feature most users overlook: Enhanced Dictation mode. When enabled in System Settings → Keyboard → Dictation, it downloads a language model to your Mac and processes everything locally using the Neural Engine. No audio ever leaves your device.

The tradeoff is accuracy. Apple's on-device model is excellent for standard prose and everyday language but falls behind OpenAI's Whisper on technical vocabulary, accented speech, and non-English languages. For many users — especially those dictating in English without specialized terminology — it's the most private option available at zero cost.

Cloud mode sends audio to Apple's servers. Apple's privacy policy is strong: they don't sell data or use it for advertising, and the data handling is governed by Apple's privacy framework, which includes GDPR compliance for European users. But the audio does leave your device.

Wispr Flow — Tier 2 (Vendor-Proxied)

Wispr Flow is transparent that audio travels through their infrastructure. From their privacy policy: "We collect audio recordings and transcriptions to provide our services and may use this data to improve our products." There is an opt-out mechanism for training data use, but it's not opt-in by default.

The app routes audio through Wispr's servers, which perform AI enhancement and context-awareness features. These server-side features — the "AI understanding" of what you're doing on screen — require your audio to be processed in their cloud, not just forwarded to a transcription API.

Wispr Flow is distributed outside the Mac App Store and does not carry an App Sandbox entitlement. This means it can access more of your file system and system resources than sandboxed apps. This is necessary for some of its context-aware features, but it's worth knowing.

None of this makes Wispr Flow unsafe — they're a well-funded startup with a credible privacy policy. But you are trusting a third party with your audio, not just OpenAI.

Superwhisper — Tier 2 (Vendor-Proxied with Local Option)

Superwhisper offers something genuinely valuable for privacy-conscious users: a local processing mode. If you select one of their on-device Whisper model tiers, your audio never leaves your Mac. Accuracy and speed vary with your Mac's hardware — M-series Macs handle local models well; Intel Macs may find the experience sluggish.

Cloud mode routes through Superwhisper's infrastructure, similar to Wispr Flow. Their privacy policy permits using transcription data for improving their services unless you explicitly opt out.

The local model option makes Superwhisper the strongest choice for offline-first privacy, particularly for users who don't want any network dependency for transcription.

Whisper Transcription — Tier 2 (Direct API, Partial)

Whisper Transcription on the Mac App Store supports BYOK (bring your own key), making it Tier 1 when configured that way. If you use the app's built-in API credits instead, you're on the developer's key and the privacy model shifts: the developer's account sees your audio, not just OpenAI.

Worth noting: Whisper Transcription is primarily a transcription recorder, not a system-wide push-to-talk dictation tool. It's excellent for transcribing long recordings or files, but it doesn't integrate into other apps the way ParlaParla or Wispr Flow does.

Otter.ai — Tier 3 (Proprietary Cloud)

Otter.ai is a collaboration-focused transcription service, not a push-to-talk dictation tool — but it appears in enough "best dictation app" searches to warrant inclusion. All audio is processed on Otter's proprietary infrastructure, and their business model includes enterprise upsells that depend on stored transcriptions being searchable and shareable.

If your use case is meeting transcription where you want recordings stored and accessible, Otter is reasonable. For private, real-time dictation of sensitive content, Tier 3 is the wrong choice.

---

The Questions Worth Asking Before You Choose

Before selecting a dictation app, these five questions will tell you most of what you need to know about its privacy posture:

1. Does audio leave my device, and if so, where does it go? "Our servers" is less informative than "OpenAI's API" or "Apple's servers." The more specific the answer, the better.
2. Is my audio used to train models? Look for an explicit "no" or an opt-out mechanism. Vague language like "improve our products" typically means yes.
3. How long is audio retained? Short, defined windows (30 days, then deleted) are better than indefinite retention. "As long as necessary" is a red flag.
4. Is the app sandboxed? Mac App Store apps are required to be sandboxed — their file and network access is restricted. Apps distributed outside the App Store may have broader system access.
5. What happens to my data if the company is acquired? Startup acquisitions routinely include data assets. Your audio library becoming part of an acquisition is a real risk with any cloud-dependent service.

---

Summary Table

---

What "No ParlaParla Servers" Actually Means

We want to be direct about something: we built ParlaParla, so you should weigh our claims about it accordingly. But the BYOK architecture is verifiable. If you're technically inclined, you can monitor network traffic from ParlaParla using a tool like Little Snitch or Charles Proxy and confirm that outbound connections go only to api.openai.com — no other hosts, no analytics endpoints, no telemetry beacons.

The code path is: microphone input → local audio buffer → HTTP POST to https://api.openai.com/v1/audio/transcriptions using your API key. The response comes back as JSON with the transcription text, which is inserted at your cursor. Your audio is not stored locally after the transcription completes. It's not cached. It's not logged.

The privacy architecture isn't a marketing claim — it's a consequence of the BYOK model. There are no ParlaParla servers to store audio on because there are no ParlaParla servers at all.

---

Choosing Based on Your Threat Model

Privacy isn't binary, and the right choice depends on what you're actually worried about:

If you're dictating highly sensitive content (legal, medical, financial, journalistic sources), go Tier 1 with BYOK, or use Apple's Enhanced Dictation for fully on-device processing. Nothing leaves your Mac.

If you're a developer who wants fast, accurate, system-wide dictation and you already have an OpenAI API key, ParlaParla or Whisper Transcription (BYOK) are the obvious choices. Same transcription engine as the $15/month apps, direct API call, one-time purchase.

If you want zero internet dependency and you have an M-series Mac, Superwhisper's local mode or Apple's Enhanced Dictation are your best options. Both run models on-device using the Neural Engine.

If privacy is a secondary concern and you want the richest feature set with AI context awareness, Wispr Flow delivers that — but understand that your audio flows through their infrastructure, and their business model depends on that arrangement.

The best dictation app for privacy is the one whose architecture matches what you're willing to trust.

---