Inside MinuteAI: Transcription That Never Leaves Your Mac or iPhone

Where Transcripts Usually Go

Ask most transcription apps to turn a meeting recording into text, and the actual work happens somewhere else entirely: your audio gets uploaded, a server processes it, and a transcript comes back a few seconds or minutes later. For a casual voice memo that’s a minor inconvenience. For a client meeting, a legal recording, or an interview with someone who assumed the conversation was private, it’s a much bigger question — where did that audio just go, and who else can see it?

MinuteAI is built around a different answer: nowhere. It’s a privacy-first local AI transcription app for macOS and iOS that turns meetings, interviews, and voice memos into text processed entirely on your device — no cloud round trip, no tracking, and no server that ever sees your audio.

What MinuteAI Actually Does

MinuteAI is native to macOS and iOS, built with SwiftUI, and runs on-device AI through CoreML — Apple’s framework for running machine learning models efficiently on Apple Silicon. In practice that means:

• On-device transcription — audio is transcribed locally, not uploaded to a transcription API
• CoreML-powered inference — models run through Apple’s Neural Engine instead of a remote GPU cluster
• Zero cloud dependency — the app works without an internet connection, because the processing never needed one
• Native macOS and iOS apps — not a wrapped web view, a proper SwiftUI app on both platforms

It’s free to download, and the whole pitch is in that one sentence from its own product description: no cloud, no tracking, just pure privacy.

The Pipeline Underneath

We wrote about the architecture behind this kind of app in more general terms in Building Privacy-First AI Apps on macOS, and MinuteAI is the concrete product that architecture describes. The short version: a Model Manager downloads, caches, and loads ML models locally; a processing pipeline chains audio into transcription and then into post-processing steps like summarization; and a result store — SQLite-based, with full-text search — keeps everything queryable without a network call. Core ML handles the Whisper-based transcription model, while GGUF (via llama.cpp) powers the LLM-based features like summarization, since those two workloads have different performance profiles and benefit from different runtimes.

Memory Is the Actual Constraint

Running AI models locally trades a privacy problem for a resource-management problem. A Whisper Large model needs roughly 3GB of RAM to run, which is not nothing on a phone and not free even on a Mac with plenty of memory to spare. The practical response to that constraint is what shapes a lot of the app’s behavior under the hood: models load lazily — only when you actually trigger a transcription — and unload automatically after 60 seconds of inactivity rather than sitting resident in memory indefinitely. Model weights are read via memory-mapped files instead of being fully loaded into RAM up front, and the app watches os_proc_available_memory() to degrade gracefully instead of crashing when memory gets tight.

None of that is visible to a user opening the app to transcribe a voice memo. It’s the difference between an app that works reliably on a base-spec iPhone and one that only performs well on the developer’s own high-end hardware.

Why This Matters Specifically for Transcription

Transcription is a category where the privacy stakes are higher than most AI use cases, because the input is almost always someone’s actual voice, saying something they didn’t necessarily plan to have transcribed by a third party. Meeting recordings often include information nobody meant to put in an email. Interviews capture things people say off the record. Voice memos are, by definition, personal. Running that processing on-device isn’t a nice-to-have feature flag — it changes what MinuteAI is allowed to be used for, because there’s no upload step where a company’s confidentiality policy would otherwise have to weigh in.

Where MinuteAI Fits in the Portfolio

MinuteAI shares its core discipline — data never leaves the device — with the rest of the AityTech portfolio. PrivateAI applies the same principle to a set of browser-based tools instead of a native app. Legacy Dragon applies it to parsing legacy source code locally rather than sending it to a cloud service. Yomite applies it to reading Japanese text aloud offline. Different problems, same non-negotiable: don’t ask for more trust, or more data, than the job actually requires.

---

MinuteAI is available on the App Store for macOS and iOS, or you can learn more at getminute.app. Questions about how it fits your workflow? Reach out at [email protected].