private meeting transcription — your audio never uploads

who actually needs on-device transcription

Most note-taking tools optimize for convenience and treat audio as something to ship to a data center. That trade is unacceptable in a lot of real work: attorney–client conversations, therapy and medical consultations, HR investigations, journalism with confidential sources, M&A and board discussions, security reviews, and any meeting covered by an NDA that doesn't have a "we may share recordings with our AI vendor" carve-out.

As of July 2026, the risk is no longer hypothetical. Otter faces a consolidated federal class action over recording meeting participants without consent and processing their voiceprints, and Fireflies has drawn two BIPA voiceprint suits. Even "bot-free" doesn't mean private: Granola sends recorded audio to Deepgram or AssemblyAI for transcription, passes transcripts to OpenAI and Anthropic, and trains on user data by default. And the one competitor that does offer on-device transcription — Krisp — gates it behind Enterprise pricing.

For all of these, the cleanest answer to "is the vendor's cloud secure enough?" is to make the question disappear: if audio never leaves the machine, there is no vendor cloud to evaluate, no retention policy to negotiate, and no breach scenario that includes your recordings.

your old recordings deserve the same privacy

The recorder page also accepts existing recordings: drop an MP3, M4A, WAV, or OGG file — or an MP4, whose audio track is extracted locally — and it is transcribed on your device with timestamps, then becomes a full meeting with a summary, speaker labels, exports, and search. The file never uploads; the network tab stays as quiet during an import as it does during a live recording.

That matters because file import is where cloud services are at their most cloud: an uploaded file is the vendor's copy by definition. It's also where free tiers pinch — Otter's free plan allows 3 file imports for the lifetime of the account, and Fathom has no file upload at all. Here it's free and unlimited, because your device is doing the work.

what this means for GDPR and compliance

This is not legal advice, but the structural difference matters for compliance reviews: with cloud transcription services, the vendor processes your meeting audio, which typically makes them a data processor and triggers DPA, sub-processor, and data-residency questions. With on-device transcription there is no audio processing relationship at all — the recording and transcript live in your browser's local storage, on hardware you already control.

Three optional features do send data, and all three are text-only and explicit: AI summaries send the transcript text to an LLM when you click Summarize — and every action item and decision in the result cites the exact transcript turns it came from, so claims stay checkable; "Ask your meetings" answers questions across your synced meetings with cited sources, for signed-in users; and signing in syncs meeting text — never audio — to your account. Everything is currently free for everyone in early access (planned Pro: $8/mo). All of it is skippable; recording, transcription, speaker stats, and search over local meetings work without an account. Deletion is final, too: removing a meeting from your account deletes it, with no silent resurrection on the next sync.

private for you, consented for everyone else

On-device processing protects the audio; it doesn't replace consent — and the current wave of recording lawsuits is precisely about participants who never agreed to be recorded. So the consent flow ships alongside the privacy: a visible indicator shows whenever recording is active, and one click copies an "I'm recording this meeting" disclosure ready to paste into the meeting chat. Share pages carry a published promise as well: we never email your attendees.

verify it in 10 seconds

Open nowrecording.com, press F12 to open DevTools, and select the Network tab. Start a recording and talk. You'll see the speech model download to your device (one time, cached afterwards) — and no audio leaving it: no WebSocket streams, no chunked uploads, nothing. Stop the recording and the transcript is already there, because it was produced on your machine while you spoke. The same check works for imports: drop a file on the page and watch the transcript appear with nothing going out.

If you'd rather read than test, the full "what leaves your device" comparison matrix — with sources — is published at nowrecording.com/trust, and the commitments behind it (never meter minutes, never paywall history, never train on your data) at nowrecording.com/pricing-promise.

where cloud services are still better

Honesty about the trade: cloud ASR running on server-grade models is still more accurate than on-device Whisper, especially on far-field audio, heavy crosstalk, and strong accents. now recording narrows the gap with an opt-in high-accuracy re-transcribe pass after the meeting, but if you need the last word in raw accuracy and your compliance posture allows uploads, a cloud service transcribes better.

Bot-based services can also attend meetings you don't: send the bot, read the notes later. now recording only captures meetings you're actually in. (That advantage is narrowing from the other side, though — Microsoft Teams begins auto-blocking third-party notetaker bots in August 2026.) And because audio never uploads, the original recording lives on the device that captured it: you can export it from there anytime, but you can't stream it from another laptop the way you can with a cloud library.

the rest of the pipeline is local too — and exportable

Transcription isn't the only stage that stays on your machine. Speaker statistics — per-speaker talk-time share, turn counts, longest monologue, questions asked — are computed on-device, and meetings you never sync are searchable on-device too, so the tool is fully useful without an account.

When you want your data elsewhere, every meeting page has an Export menu: Markdown notes, a plain-text transcript, .srt/.vtt subtitles with speaker-labeled cues, full-fidelity JSON, and the original audio from the device that holds it — or everything in one click. The pledge behind it is published at nowrecording.com/your-data: your data outlives your subscription.

frequently asked questions

How can transcription work without a server?

The browser downloads OpenAI's Whisper speech model once (about 75 MB) and runs it locally via WebAssembly — no GPU needed, so it works in any current Chrome or Edge, including locked-down managed work profiles, and the model is ready in about 2 seconds. On iPhone, the same model family runs on the Neural Engine. Modern devices transcribe faster than real time, so the text streams live while you speak.

Is the transcription accurate without the cloud?

It's the same Whisper model family that many cloud services run server-side — published benchmarks put whisper-small within a couple of WER points of commercial cloud ASR on meeting audio. A higher-accuracy on-device pass can re-process the recording after you stop.

Can I transcribe an existing recording without uploading it?

Yes. Drop an MP3, M4A, WAV, OGG, or the audio of an MP4 on the recorder page and it's transcribed on-device with timestamps — free, unlimited, and the file never leaves your machine. It becomes a regular meeting with a summary, speaker labels, exports, and search. For comparison: Otter's free plan allows 3 file imports for the lifetime of the account, and Fathom has no file upload at all.

Where are my recordings stored?

In your browser's local storage (IndexedDB) on your device, or in the iOS app's local storage. They never upload. If you sign in (currently free for everyone in early access), you can back up the meeting text — title, transcript, summary — to your account; the audio stays local even then. And deleting a meeting from your account is final — it doesn't silently reappear on the next sync.

Do AI summaries break the privacy model?

Summaries and "Ask your meetings" are optional and text-only: clicking Summarize sends the transcript text (never audio) to an LLM, and every action item and decision in the result cites the exact transcript turns it quotes. Cross-meeting Q&A works the same way over synced meeting text, with cited sources. If your content is too sensitive even as text, skip both — recording, transcription, speaker stats, and local search all work without them.

Is private recording the same as secret recording?

No. A visible indicator shows whenever recording is active, and one click copies an "I'm recording this meeting" disclosure to paste into the meeting chat. On-device processing protects the audio; the consent flow protects the people in the room.

What about languages other than English?

Live on-device transcription supports 29 languages. English is the most accurate; major European and Asian languages work well on the high-accuracy setting. AI summaries come out in the transcript's language — a Turkish meeting gets Turkish headings.