Halopen vs the leading dictation app

Privacy

The category leader runs cloud-only. Your audio leaves the device on every dictation, gets transcribed in a data center, and (per their terms) is discarded after transcription. The contract relies on them keeping it.

Halopen ships both. Cloud mode is our default — same architectural shape, our own servers. On-device mode is an explicit toggle. When it's on, your audio never leaves your Mac. Apple Silicon runs the transcription locally. The audit log inside Halopen records every cloud call the app makes; in on-device mode your dictations don't appear there, because there's nothing to log. The contract is a mechanical fact instead of a promise.

For most uses, either is fine. For dictation that touches regulated data, NDAs, sensitive personal material, or anything you wouldn't say out loud in a coffee shop, on-device mode is a different category of guarantee.

The audit log

The category leader does not publish a way to verify their privacy claims from inside the app. The promise that "audio is deleted after transcription" is something you have to trust on faith from outside their infrastructure.

Halopen ships an audit log. Open Settings → Privacy → Open audit log… and you'll see every cloud call the app has made from this Mac, with timestamp, endpoint, status code, duration, and (for transcription calls) audio seconds. Metadata only — no audio, no transcripts, never leaves the device. Capped at 500 entries on a ring buffer so it can't grow without bound.

The audit log is what makes the on-device-mode promise falsifiable: dictate in on-device mode, open the log, confirm no Transcription entry appeared. The privacy claim isn't something Halopen asks you to trust; it's something you can prove from your own machine in under a minute. Read more about the audit log →

Native vs Electron

The category leader ships an Electron-based Mac app. Electron apps carry the runtime overhead of an embedded browser, which shows up in Activity Monitor as elevated idle memory and CPU vs the equivalent native app.

Halopen is native Swift. AppKit, AVAudioEngine, AXUIElement — the same APIs Apple writes its own apps with. Idle RAM is in tens of megabytes; idle CPU is near-zero. The menu-bar pill is a real macOS panel, not a web view. The hotkey is a CGEventTap, not a JavaScript shim. The text injection talks to the AX tree directly.

This isn't an ideological point. It's a feel point. Native apps respond differently than Electron apps under load, integrate with macOS conventions cleanly, and don't show up at the top of Activity Monitor when your battery is dying. Halopen feels like a Mac app because it is one.

Verbatim vs cleanup

The category leader's flagship feature is a cleanup pass: their model rewrites your dictation in a "professional" register by default. Many users like this; many others have complained — loudly, on Reddit and in support tickets — that the cleanup sands off contractions, code-switches, intensifiers, and the rhythm of how they actually speak.

Halopen ships verbatim by default. What you said is what appears at your cursor. Polish, when you want it, will be opt-in — a separate toggle, never an unannounced rewrite. The speaker is the author. The app is the typist.

Transparency policy

The category leader's privacy disclosures refer generically to "third-party LLMs" and "industry-standard providers." The transcription model is not named on a discoverable page. The subprocessors are not enumerated. The retention windows are not published.

Halopen publishes a transparency policy that lists what we disclose openly (the transcription engine for both modes, every subprocessor, every retention window, the audit-log architecture, the on-device invariant) and what we hold as competitive intel (the token allocations behind vocabulary biasing, the per-app injection profiles, the edge-function prompt templates). The decision rule is published on the same page so you can hold us to it.

The category leader will not publish a document like this. They cannot — their disclosure default is opacity, and publishing the rule would force them to honor it. That is the structural difference: we compete on being the most-trusted dictation app, not the most-secret one.

Pricing

The category leader is subscription-only. There's no path to own the app outright; the relationship is monthly or yearly, forever.

Halopen offers a free tier (8,000 words per month, forever, no card), Pro Monthly at $19/mo, Pro Annual at $179/yr, and Pro Lifetime at $499 — one payment, all features for life, including future major versions. Lifetime exists because some people don't want a subscription relationship with their dictation app. We agree it shouldn't be the only option.

Accuracy

Both apps lean on state-of-the-art transcription models. The category leader uses a proprietary ensemble in their cloud and doesn't say which models or providers — their privacy policy refers only to "third-party LLMs" generically. Halopen names the engine: Cloud mode uses OpenAI's gpt-4o-transcribe; on-device mode uses Whisper Large v3 via WhisperKit. Head-to-head on clean English dictation, the gap between our engine and theirs is small enough that most users won't notice.

We name the model on purpose. The transcription engine is the floor; what we build on top is the ceiling. Naming the model means you know exactly what's processing your voice — and when something better ships, you'll know we shipped with it.

Where Halopen makes a different bet is vocabulary biasing: when you hold the hotkey, Halopen pre-feeds your personal dictionary to the transcription engine so proper nouns, brand names, and technical jargon land on the first pass. (Clipboard and selected-text biasing is also available as an explicit opt-in, off by default because clipboard contents can be sensitive.) The wedge is that the bias is applied at the transcription step, not as a post-processing rewrite — fewer rewrites means less risk of the engine quietly changing what you said.

Where they're better than us, today

We don't want to write a one-sided comparison. The category leader has a real product with real strengths Halopen doesn't yet match:

• Cross-platform: they ship Mac, Windows, iOS, Android. Halopen is Mac-only today; a native Windows version is in development. iOS is on the roadmap; Android isn't.
• Multilingual code-switching: their model handles mid-sentence language switching more cleanly than ours does today, particularly Hindi-English.
• Hands-free toggle mode for long-form dictation: they have it, we don't yet.
• Team / Enterprise tooling: their multi-seat admin panel is more developed than our Halopen for Business preview.
• Years of head start: they've been refining for two-plus years; we're a few months into shipping.

The right answer probably isn't a one-app-for-everyone story. Use what fits your work.