The gap that matters is no longer accuracy. On-device and cloud transcription now produce comparable text for ordinary meetings in major languages. What differs is structural: cloud transcription creates a copy of your meeting on someone else's infrastructure, governed by terms you don't set and reachable by legal process aimed at them, not you.
That distinction survives every marketing cycle, which is why it is worth understanding once rather than re-litigating per vendor.
What actually changes when your audio leaves the Mac?
Three things change, and none of them are "the cloud is insecure." Reputable providers encrypt in transit and at rest and staff real security teams. The changes are about custody, not competence.
A second copy exists. Your recording now lives in at least two places. Deleting your local file no longer deletes the meeting.
The rules can be revised without you. Retention windows, training policies, and sub-processor lists are terms of service, and terms of service get updated. A guarantee that depends on a policy is only as durable as the next policy version. Otter's privacy policy, as published in September 2026, commits to storing personal information "for as long as necessary to fulfill the purposes set out in this Policy" — a purpose-bound window, not a fixed number of days.
A third party becomes a target. Any subpoena, warrant, or preservation order aimed at the provider now reaches your meeting, and you are not necessarily party to it.
Apple, which sells both on-device and server-side AI and therefore has no incentive to overstate the local case, puts the underlying asymmetry plainly in its Private Cloud Compute writeup:
When on-device computation with Apple devices such as iPhone and Mac is possible, the security and privacy advantages are clear: users control their own devices, researchers can inspect both hardware and software, runtime transparency is cryptographically assured through Secure Boot, and Apple retains no privileged access. — Apple Security Research, Private Cloud Compute (as of September 2026)
The same document is candid about why Apple still built a server tier: "to process more sophisticated requests, Apple Intelligence needs to be able to enlist help from larger, more complex models in the cloud." Both halves of that are true, and the second half is the honest case for cloud.
Is cloud transcription actually more accurate?
For meetings in major languages, no — not by enough to decide on. The speech models that run locally on Apple silicon are frequently the same open models cloud services run server-side. NVIDIA's Parakeet TDT 0.6B v3 and OpenAI's Whisper large-v3 are both downloadable and both fit on a laptop.
The numbers that matter are the ones measured on real meeting audio, not read audiobooks. On the AMI meeting corpus, NVIDIA reports 11.31% word error rate for Parakeet V3 and OpenAI reports 15.95% for Whisper large-v3 — the same models either architecture would use. We took that apart in Whisper vs Parakeet on real meeting audio.
Where cloud still wins on quality is narrower and worth naming: rare or low-resource languages, heavily accented or badly-miked audio, and very long batch jobs where server compute simply finishes sooner. If your calls are trilingual, or recorded through a laptop mic in a noisy room, a large proprietary cloud model may genuinely read them better.
What on-device costs you is the option of a much larger model. What it buys you is everything in the next section.
Who can be compelled to hand over your meeting recording?
This is the part most comparisons skip, and it is the one that doesn't depend on trusting anybody.
Since 2018, 18 U.S.C. § 2713, added by the CLOUD Act, has read:
A provider of electronic communication service or remote computing service shall comply with the obligations of this chapter to preserve, backup, or disclose the contents of a wire or electronic communication and any record or other information pertaining to a customer or subscriber within such provider's possession, custody, or control, regardless of whether such communication, record, or other information is located within or outside of the United States.
Read the operative phrase twice: possession, custody, or control. Storing data in an EU region does not remove a US provider's obligation. And the reverse is the whole point — a provider that never received your audio has nothing in its possession to produce, so the request has to come to you, where you can see it and contest it.
This is not hypothetical volume. Zoom's own transparency report — covering July 1 to December 31, 2024, the most recent period published as of September 2026 — records 175 US government requests for user data, of which 66 arrived with a valid delayed-notice order, meaning the user could not be told at the time.
Otter's policy, likewise, reserves disclosure "if we are legally required to do so, or if we have a good faith belief that such use is reasonably necessary to comply with a legal obligation" (as of September 2026). That is a normal, lawful clause. It is also unavoidable once a copy exists.
Separately, some providers train on what you upload — Otter's policy describes training its AI "on de-identified audio recordings and on transcriptions." We checked that tool by tool in Do AI notetakers train on your meetings?
On-device vs cloud transcription, side by side
| Dimension | On-device | Cloud (as of September 2026) |
|---|---|---|
| Where audio is processed | Your Mac | Provider's servers |
| Copies of the recording | One, on your disk | Yours plus the provider's |
| Reachable by subpoena to the vendor | No — nothing in their custody | Yes, under 18 U.S.C. § 2713 |
| Meeting-audio accuracy | 11.31–15.95% WER (Parakeet / Whisper) | Comparable; better on rare languages and rough audio |
| Works with no connection | Yes, after model download | No |
| Training on your data | Not possible — data never arrives | Vendor-dependent; check the policy |
| Team-wide shared archive | No | Yes |
| Cost model | Typically one-time (Speechmark: $79) | Per-seat subscription |
Where cloud transcription is the better choice
Plenty of the time, honestly.
Teams that need one archive. If eight people must search every call from a laptop, a phone, and a browser, with permissions and shared workspaces, that is exactly what cloud tools are built for. A single-Mac app cannot do it, and pretending otherwise helps nobody.
Integration-driven workflows. Pushing notes into Salesforce, HubSpot, or a ticketing queue is cloud-native work.
Mixed or non-Mac fleets. On-device quality depends on the silicon in front of you. A Windows-and-Chromebook company gets more consistent results from a server.
Hard language or audio requirements. Covered above — sometimes the bigger model is the right answer.
A reasonable split many people land on: route confidential calls to the on-device pipeline and everything routine to the team's cloud tool. The categories are not a loyalty test.
What to check before you trust an "on-device" label
"Local" is a marketing word as often as an architectural one. Three questions separate them:
Does transcription itself run locally, or just the note editor? Several apps cache notes on disk for fast offline reading while sending audio to a transcription API. Both are honestly described as "works offline," and only one keeps the audio home. We tested which Mac apps actually survive airplane mode in Offline meeting notes: which Mac apps work with no internet.
What happens to the summary? Transcription and summarization are separate stages. An app can transcribe locally and still ship the transcript text to a cloud model for the summary. That is a defensible design — text is far less sensitive than a voice recording — but you should know it is happening.
Who keeps the original audio? Some tools transcribe and then delete the recording; some never store it; some keep it indefinitely. The full tool-by-tool breakdown is in our roundup of private Mac notetakers.
Speechmark's own position, stated so it can be checked: recording, transcription, speaker separation, and — by default — summarization all run on the Mac; the audio file stays on your disk next to the transcript; there is no account and no server-side copy to subpoena. If you deliberately plug in your own OpenAI, Anthropic, or Gemini key, transcript text goes to that provider for the summary, and nothing else does. We are not HIPAA or SOC 2 certified, and we don't claim to be — the argument here is architectural, not a certification.
See the difference on your own calls: download Speechmark free, record a meeting, then pull the Wi-Fi. The transcript and speakers arrive anyway, because nothing was ever waiting on a server.