Weesper Neon Flow is the only dictation app in our 2026 audit that makes zero network calls during transcription, verified by packet capture. Apple Dictation and Windows Voice Typing process locally on supported hardware but still ship diagnostic telemetry. Every cloud-based competitor we tested — including Wispr Flow, Otter.ai, Google Docs Voice Typing, and Dragon Anywhere — transmits audio to third-party servers governed by privacy policies you cannot independently audit.
Why a dictation app privacy audit matters in 2026
A dictation app privacy audit matters because voice is now classed as biometric personal data under GDPR and as personally identifiable information under CCPA. The European Data Protection Board treats voice recordings as biometric identifiers because they reveal gender, ethnic origin, and even potential health conditions. That elevates voice data well above ordinary text — and most users have no idea their dictation app may be storing, training on, or human-reviewing those recordings.
The stakes have risen sharply in the last twelve months. Otter.ai is defending a US federal class action over alleged recording without consent. Wispr Flow’s prior SOC 2 audits were caught up in the March 2026 Delve compliance scandal, where 99.8% of 494 reports shared identical boilerplate. Cloud speech-to-text providers including Google, Amazon, Microsoft, and Apple all retain audio by default unless you actively opt out.
This audit examines nine of the most-used dictation apps on Mac and Windows. We tested each one for outbound network traffic, parsed their published privacy policies, and checked their local data footprint. The methodology is reproducible — you can run the same tests yourself in under an hour.
How we ran the privacy audit
The audit follows a three-layer methodology that any privacy-conscious professional can replicate. We tested each app under Little Snitch on macOS Sonoma and Wireshark on Windows 11. Every test was run with a fresh user profile and default settings — what a typical buyer would experience out of the box.
Layer 1 — Network traffic: We captured every outbound connection during a 60-second dictation session. We flagged any connection to an audio-processing endpoint, telemetry collector, or third-party AI service.
Layer 2 — Privacy policy audit: We parsed each app’s current privacy policy for four signals: what is collected, where it is stored, who has access (subprocessors), and whether it is used for model training. We treated “may” clauses as enabled unless the user must opt in.
Layer 3 — Local data footprint: We inspected ~/Library/Application Support, ~/Library/Caches, and the Windows equivalents after a session. We looked for raw audio caches, transcript logs, and analytics queues.
We also ran an airplane mode test: if the app continues to transcribe with all network interfaces disabled, the model is local. If it fails or falls back to a degraded mode, it is cloud-dependent.
Privacy comparison: 9 dictation apps ranked
The table below summarises the voice data security comparison across the apps we tested. Apps are ranked by the strictness of their default privacy posture, not by feature set.
| App | Audio leaves device | Used for training (default) | Audio retention | Privacy verdict |
|---|---|---|---|---|
| Weesper Neon Flow | No | No | None (not stored) | Strongest — provably offline |
| Apple Dictation (Enhanced) | No (M-series Macs) | No | None | Strong — local on supported hardware |
| Windows Voice Typing (offline) | No (Copilot+ PCs) | No | None | Strong — local on supported hardware |
| Dragon Professional v16 | No (Windows desktop) | No | Local profile only | Strong but Windows-only, £600+ |
| Apple Dictation (Standard) | Yes | No | Aggregated diagnostics | Moderate — falls back to cloud |
| Dragon Anywhere | Yes | Unspecified | Cloud, Microsoft-owned | Moderate — cloud-only mobile |
| Wispr Flow | Yes | Yes (unless Privacy Mode on) | Cloud, AWS us-east-1 | Weak by default, configurable |
| Google Docs Voice Typing | Yes | Yes (Google Workspace terms) | Unspecified | Weak — opaque retention |
| Otter.ai | Yes | Yes (de-identified) | “As long as necessary” | Weak — active litigation |
A few patterns stand out. First, every truly offline option is platform-restricted: Weesper covers both Mac and Windows, but Apple and Windows native dictation require recent hardware, and Dragon Professional is Windows-only. Second, every cloud app uses “may” language for training and retention, which means the default behaviour is opt-out, not opt-in. Third, “encrypted in transit” is universal and meaningless as a differentiator — the provider still holds the keys.
Weesper Neon Flow: zero network calls, verified
Weesper Neon Flow made zero outbound connections during transcription in every test run. The model runs entirely on your Mac or Windows PC using Whisper-derived local inference. There is no audio endpoint, no telemetry queue, no subprocessor chain. We confirmed this with three independent checks.
- Little Snitch capture (macOS): No outbound traffic from the Weesper process during a 60-second dictation session.
- Airplane mode test: Weesper continued to transcribe with all network interfaces disabled. Latency was unchanged.
- Local footprint: No raw audio files written to disk. Transcripts are inserted into the active app and discarded.
This is the architectural reason Weesper appears at the top of the offline dictation privacy test rankings: there is no privacy policy interpretation needed. If audio never leaves your device, no policy clause matters. For a deeper look at why this matters, see our guide to offline voice dictation and privacy.
Wispr Flow: cloud-first with a configurable privacy mode
Wispr Flow transmits audio to cloud subprocessors by default. Audio passes through Baseten for inference, with text processed by OpenAI, Anthropic, and Cerebras at various stages. Storage is in AWS us-east-1. With Privacy Mode disabled, Wispr Flow may retain dictation data “for debugging model failures or improving transcription services.”
Privacy Mode, when enabled, blocks retention and training use entirely. The mode is on by default for Enterprise tiers and off by default for individual users. Wispr Flow also offers a Context Awareness feature that captures content from the active window — useful for accuracy, but a meaningful expansion of the data the app sees. The Wispr Flow privacy posture is therefore configurable, but the safe configuration is not the default.
We covered Wispr Flow’s wider trade-offs in our Wispr Flow review and in our Wispr Flow alternatives roundup. The short version: it is a competent tool that you should treat as a cloud service, not a private one.
Otter.ai: training on your recordings by default
Otter.ai’s privacy policy is explicit about dictation app data collection: the company trains its proprietary AI on de-identified audio recordings, and transcripts containing personal information may also be used for training. Audio is stored in AWS infrastructure. Subprocessors include data labelling vendors and AI service providers.
Otter.ai is also defending a US federal class action — In re Otter.AI Privacy Litigation — over alleged recording without proper consent in two-party-consent jurisdictions. The case centres on the visibility of bot consent in virtual meetings, but the broader signal is that Otter.ai’s default data handling does not always align with user expectations.
For confidential dictation, Otter.ai’s combination of cloud processing, default training use, and active litigation makes it a high-risk choice. The opt-out path exists but requires careful policy reading and account configuration.
Dragon Professional, Anywhere, and Medical One
Microsoft acquired Nuance in March 2022 for $19.7 billion, and Dragon now ships in three distinct variants with very different privacy postures.
- Dragon Professional v16 (£599 one-time, Windows-only): Processes audio locally on your PC. Your voice profile stays on the machine. This is the privacy-strong option but the licence is expensive and macOS users are excluded.
- Dragon Anywhere (£12/month, mobile): Cloud-only. Audio leaves the device to Microsoft Azure. Retention terms are not user-configurable in the consumer flow.
- Dragon Medical One (£60–80/user/month): Cloud-only on Microsoft Azure with a signed Business Associate Agreement (BAA) for HIPAA covered entities.
If you need Dragon-class accuracy without Windows-only restrictions or a per-seat cloud fee, see our Dragon alternatives 2026 guide and our on-device vs cloud transcription comparison.
Built-in dictation: Apple, Windows, and Google
Built-in dictation tools sit in a middle ground: free, convenient, and partially private — but each one ships diagnostics that most users never see.
Apple Dictation runs locally on M-series Macs when Enhanced Dictation is enabled. On Intel Macs or with Enhanced Dictation disabled, audio is sent to Apple servers, anonymised with random identifiers, and retained for up to six months for improvement, then up to two years in anonymised form. Apple now requires explicit opt-in for human review.
Windows Voice Typing runs locally on Copilot+ PCs (Snapdragon X, Intel Core Ultra Series 2). On older Windows 11 hardware, audio is processed by Microsoft cloud services. Microsoft retains audio for up to 30 days by default for troubleshooting, and custom speech models may retain training data indefinitely.
Google Docs Voice Typing is fully cloud-based. Audio is governed by Google Workspace terms, which permit use for service improvement. Retention duration is not specified in user-facing documentation. There is no offline mode.
For a wider comparison of native versus dedicated tools, see our Apple Dictation alternatives guide.
What the GDPR and CCPA require for voice data
Voice is biometric personal data under GDPR, full stop. Since May 2018, any organisation handling EU residents’ voice recordings must inform users about collection, allow opt-out of data sales, provide data access, and enable deletion under the right to be forgotten. Maximum GDPR fines reach €20 million or 4% of global annual revenue, whichever is higher.
CCPA covers Californian residents at organisations above $25 million in annual revenue or handling data of 50,000+ consumers. It treats data sharing arrangements that provide “valuable considerations” as sales, which captures most ad-tech integrations.
The practical implication for buyers is simple: any cloud dictation app you use for European or Californian customer conversations needs a documented lawful basis, a Data Processing Agreement, and a retention schedule you can defend. Offline dictation removes all three obligations because no personal data is processed by a third party in the first place. See our GDPR voice dictation compliance guide for a deeper walkthrough.
How to choose: a 30-second decision tree
The right answer to which dictation app is most private depends on your threat model and platform. Use this short decision tree.
- Does your work involve regulated data (health, legal, finance, trade secrets)? Pick a provably offline option. On Mac or Windows, that is Weesper Neon Flow. On Windows-only with budget for an enterprise licence, Dragon Professional v16 also works.
- Do you have a Copilot+ PC or M-series Mac and only need light dictation? Windows Voice Typing (offline) or Apple Dictation (Enhanced) are credible free options.
- Do you need cloud features like meeting transcription and shareable notes? Otter.ai or Wispr Flow are competent — but treat the output as recorded and reviewable, not private.
- Do you handle EU customer voices? Default to offline. Cloud paths require a DPA, lawful basis, and retention policy you can defend in audit.
For a broader buyer’s view, our enterprise security guide and best offline dictation software 2026 cover the procurement angle in more depth.
Frequently Asked Questions
The FAQ items above answer the most common questions about dictation app privacy audits in 2026. If your question is not covered, the Weesper Help Center has additional documentation on data handling, GDPR posture, and self-hosting options.
Conclusion
The most private dictation app is the one that never sees your audio. In our 2026 audit, only Weesper Neon Flow combined provable offline operation with cross-platform support across Mac and Windows. Apple Dictation and Windows Voice Typing are credible local alternatives on recent hardware, but both require platform compromises. Every cloud-first competitor — including Wispr Flow, Otter.ai, Google Docs Voice Typing, and Dragon Anywhere — sits on a privacy policy you cannot independently verify and a subprocessor chain you cannot inspect.
If your dictation contains anything you would not want a stranger to read, the architectural answer is offline. Download Weesper Neon Flow for a free 15-day trial and run the audit yourself — Little Snitch will show you the same zero outbound traffic we saw.