On-Device Meeting Transcription: How It Works on a Mac

On-device meeting transcription means the recording, the transcript, and the summary are all produced on your own Mac, and none of the audio or text is sent to a cloud service to get there. The meeting never leaves the machine. That’s the whole definition, and for a growing number of people it’s the only acceptable way to take notes on a call — because the alternative is uploading every word your customers, candidates, and colleagues say to a third party’s servers.

For most of the last decade, “transcribe my meeting” meant exactly that upload. Otter, Fireflies, Granola, and the rest record or ingest your audio, ship it to their backend, transcribe it there, run a cloud LLM over it, and store the result on their infrastructure. It works, and for low-stakes meetings the tradeoff is invisible. But it’s a real tradeoff, and on a modern Mac it’s no longer a necessary one. macOS 26 Tahoe ships the frameworks to do the entire pipeline locally, at speeds that make the cloud version look slow.

This post explains what on-device meeting transcription actually is, how it works on Apple hardware, where it beats the cloud and where it doesn’t, and how to set it up.

What “on-device” actually means

The phrase gets used loosely, so it’s worth being precise. On-device meeting transcription means all three of these stages run locally:

Capture — the meeting audio is recorded by your Mac, from your Mac’s own audio, with no bot dialing into the call and no audio streamed to a server.
Transcription — the audio is converted to text by a speech-recognition model running on your Mac’s Neural Engine, not by an API call to a transcription service.
Summarization — the transcript is turned into decisions, action items, and notes by a language model running on the device, not a cloud LLM.

If any one of those three happens in the cloud, it’s not fully on-device — it’s a hybrid, and the audio or transcript left your machine at some point. Most “private” notetakers are hybrids: they may encrypt in transit, but the processing still happens on their servers, which means the data is decrypted and present there. Genuinely on-device means the data is never on anyone else’s computer at all.

Why it matters

Three reasons people specifically want the local version, in rough order of how often they come up:

1. Confidentiality you can actually attest to. When a customer’s security team asks “where does this recording go,” the on-device answer is “nowhere — it’s on this Mac’s SSD, behind FileVault, and it auto-deletes in seven days.” There’s no data-processing agreement to negotiate, no subprocessor list, no breach surface at a vendor. For sales calls under NDA, candidate interviews, internal strategy, or anything a competitor would want, that’s not a nice-to-have; it’s the requirement.

2. It works with no network. A flight, a customer site with locked-down Wi-Fi, a SCIF-adjacent office that blocks outbound traffic — none of it stops local capture. A cloud notetaker needs the network up the whole time; an on-device one doesn’t care.

3. No bot, no per-meeting cloud cost, no vendor lock-in. Because nothing is being streamed to a server, there’s no “Notetaker” attendee joining your meeting, no per-minute cloud transcription bill passed back to you as a subscription, and no transcript archive held hostage on a platform you’d have to export from.

The privacy argument is the same one that pushes people off cloud tools in the Otter alternative and Granola alternative write-ups — the difference here is the focus on why the local architecture is now technically viable, not just preferable.

How it works on a Mac, framework by framework

The reason this is possible in 2026 and wasn’t in 2022 is that Apple shipped three pieces that close the gap between “record” and “summary” entirely on-device. On macOS 26 Tahoe:

Capture: ScreenCaptureKit

The hard part of meeting transcription is the far-end audio — the other people’s voices, which come out of your Mac’s speakers as system audio, not into your microphone. ScreenCaptureKit with an audio-only content filter records the system audio output of a specific running process. Point it at the Zoom client, the Teams client, or the browser tab running Meet or Webex, and it captures that app’s audio, scoped to just that process, with no virtual audio driver (no Soundflower, no BlackHole) and no bot in the call. Mix in the microphone and you have both sides. The full mechanics are in the guide on recording system audio on Mac.

This is the step cloud notetakers solve by sending a bot into your meeting. On-device, the Mac just records its own audio output — no participant added, nothing for a security team to flag.

Transcription: SpeechAnalyzer

SpeechAnalyzer, the on-device speech-recognition framework Apple introduced at WWDC 2025, transcribes the captured audio on the Neural Engine. Per Apple’s benchmarks it runs roughly 55% faster than Whisper v3 Turbo on the same chip, which is the number that makes local transcription practical: an hour of meeting transcribes in a few minutes, not in real time, with no upload. Because it’s running on the raw audio you recorded — not on a meeting platform’s live captions — the transcript quality is yours to control, and you can always re-listen to check it.

Summarization: Foundation Models

Apple’s Foundation Models framework exposes an on-device LLM. It takes the transcript and produces the summary, the decisions, the action items, and the owners — the part you actually read. This is the stage most “local” tools still send to the cloud, because running a capable LLM on-device only became practical recently. On Apple Silicon with macOS 26, it runs locally, which is what lets the entire pipeline stay on the machine.

The result: an audio-anchored summary

When all three run locally, you can do something cloud tools structurally can’t do well: tie every summary bullet back to the exact moment in the audio it came from. Because the audio, the transcript, and the summary all live together on the device, clicking a summary line can scrub the recording to that second and highlight the transcript. That’s the difference between a summary you hope is right and one you can verify in two seconds.

On-device vs. cloud, honestly

Dimension	Cloud notetaker	On-device (Mac)
Where audio is processed	Vendor servers	Your Mac
Where transcript is stored	Vendor cloud	Local SSD (FileVault)
Bot in the meeting	Often yes	No
Works offline	No	Yes
Data-residency answer	Vendor’s subprocessors	”This Mac”
Speaker labels	Usually yes	Not in v1.0 of most local apps
Team integrations (CRM, Notion)	Usually rich	Usually export-only
Cross-OS (Windows/ChromeOS)	Yes	Mac-only for system audio
Per-seat cloud cost	Yes	No (one device, one app)

The honest read: cloud notetakers win on speaker labels, deep SaaS integrations, and cross-OS coverage, because those lean on server-side processing and a web backend. On-device wins on confidentiality, offline capability, no-bot capture, and cost. Which set matters more depends entirely on what your meetings contain. For a routine internal standup, the cloud convenience may be worth it. For anything confidential, the local architecture is the one you can defend.

What you need to run it

On-device meeting transcription has real hardware requirements, because Apple Intelligence and these frameworks need recent silicon:

A Mac on macOS 26 Tahoe with Apple Silicon (M-series). The Neural Engine does the transcription and summarization work.
An app that wires the three frameworks together. The frameworks are public, but you need software that orchestrates capture, transcription, and summarization into one loop.
For mic-only capture on iPhone or iPad: iOS 26 / iPadOS 26 on an A17 Pro or M-series device. System-audio capture of a meeting’s far end is Mac-only — that’s a ScreenCaptureKit constraint, not an app limitation.

Setting it up with Dictanta

Dictanta is the Mac app that implements exactly the pipeline above — ScreenCaptureKit capture, SpeechAnalyzer transcription, Foundation Models summarization, audio-anchored notes, all on-device. The setup is short:

Install and grant Screen Recording permission. macOS gates ScreenCaptureKit behind the Screen Recording permission in System Settings → Privacy & Security. Grant it once. (The permission is named for screen recording, but Dictanta uses the audio-only path — it isn’t capturing your screen.)
Start the meeting in any app. Zoom desktop, Teams desktop, a browser tab — it doesn’t matter, because Dictanta records system audio, not a specific client. The same on-device path drives the Zoom, Teams, Google Meet, and Webex flows.
Press ⇧⌘R when the call starts. Dictanta records the system audio and your mic. Nothing is uploaded.
Read the summary when it ends. Transcription and summarization run on the Neural Engine in the minutes after the call. You get a summary with audio-anchored bullets — click any one to scrub to the moment it came from.

Everything stays on the SSD. Transcripts and summaries can sync across your own Apple devices via CloudKit if you want them on your iPhone; the audio stays local unless you explicitly enable iCloud Drive backup. If you prefer a tool that keeps dictation separate from meetings, that boundary is covered in the Superwhisper alternative write-up.

When on-device is the wrong choice

It’s worth naming the cases where you should just use a cloud tool:

You need automatic speaker labels today. Most on-device apps, Dictanta included, don’t diarize raw audio in their current versions. If “who said what” tagging is a hard requirement now, a cloud notetaker that reads platform captions has it.
Your team automates summaries into a CRM or Notion. Deep, automatic SaaS integrations are a cloud strength. On-device tools mostly export Markdown or JSON for you to route.
You’re on a mixed-OS team standardizing on one tool. System-audio capture is Mac-only. A cross-platform cloud extension is the simpler common denominator for Windows + Mac teams.

If none of those is load-bearing and the meetings are confidential, on-device is the better default.

Bottom line

On-device meeting transcription isn’t a privacy slogan bolted onto a cloud product — it’s a specific architecture where capture, transcription, and summarization all run on your Mac and the audio never leaves it. That architecture became practical on macOS 26 because Apple shipped the three frameworks it needs: ScreenCaptureKit to record the meeting’s system audio with no bot, SpeechAnalyzer to transcribe it on the Neural Engine faster than Whisper, and Foundation Models to summarize it locally. The payoff is a summary you can verify against the audio, an answer to the data-residency question that’s just “this Mac,” and a workflow that works on a plane.

Dictanta ships that pipeline for Mac, iPhone, iPad, and Vision Pro — free for your first three recordings with no length cap, so you can run a real meeting through it and confirm nothing leaves the device. Paid tiers are $9.99/mo, $79.99/yr, or $149.99 lifetime. If your meetings are worth keeping confidential, the local pipeline is the one you can stand behind.