Where does your meeting audio actually go?
Essay · 2026-09-13
Every meeting note-taker now says it is private. The word covers three different architectures, and the difference matters more than the marketing. Here is how to read the privacy page of any of them — ours included — in five questions.
Three kinds of private
- Fully local. A speech model runs on your machine; audio never leaves. Pure form: MacWhisper. Cost: the model has to fit on your Mac, which means Apple silicon and lots of memory, and the model is whatever fits — usually weaker on Cantonese and on switching languages mid-sentence than the large hosted ones.
- Local-first. Your notes, transcripts and everything derived from them live on your machine as files; audio is streamed to a speech provider and comes back as text; text goes to a language model to be summarised. Nothing is stored server-side. This is Audiee.
- Cloud with a good client. A desktop app records, but the record of truth is the vendor's database; deleting the app deletes nothing. Most bot-free note-takers are this, and the bot-based ones (Otter, Fireflies, Fathom, tl;dv) are this by construction.
The trouble is that all three use the same sentences: "private by design", "your data belongs to you", "no bot". So ask five questions instead.
The five questions
- Which company's servers hear the audio? Not "is it encrypted" — everything is encrypted in transit. Which named company runs the speech model, and in which region? A vendor that will not name the provider is telling you something.
- Where is the transcript stored after it is written? On your disk as a file you can open without the app, or in the vendor's database with an export button? "You can export" and "it is a file" are different answers.
- What text reaches the language model, and does it stay there? Summaries are written by an LLM. Does the vendor keep a copy, does the provider retain the text, and will the vendor tell you who the provider is when you ask? Some providers retain by default; most privacy pages do not say.
- What else is sent? Crash reports, usage analytics, the queries you type into search. The honest page lists these; the rest hope you will not ask.
- What happens when you leave? If the app is deleted today, what do you still have, and what does the vendor still have?
Audiee's answers
- 1. Meeting audio and Mandarin, Cantonese and Mixed dictation go to BytePlus's Seed ASR in Singapore; dictation in other languages goes to Deepgram. Both through our gateway, which holds the keys and stores nothing.
- 2. Transcripts and notes are Markdown under ~/.audiee on your Mac. There is no server copy to export from.
- 3. Transcript text goes to a hosted language model to be summarised; questions you ask go there too. Our gateway keeps no copy, and the trust page says what reaches the model and what never does. We name the provider to anyone evaluating Audiee, rather than on the page, because it can change between releases — the day it does, the answer to this question must not go stale.
- 4. A usage log of MCP tool calls (tool name, timing, status, client, and for search tools the first 200 characters of the query) is sent to our gateway, which currently discards it. A screenshot you ask a question about is sent to a vision model; screenshots saved to Knowledge are OCR'd on device.
- 5. Delete Audiee and you keep a folder of Markdown that opens in anything. We keep your email, tier and any shares you have not revoked, deleted within 30 days on request.
Why we did not go fully local
Because of one sentence: 「呢個 proposal 我哋 next week 再 confirm」. Models that fit on a laptop still do badly on Cantonese as spoken, and worse when English arrives every fourth word. The hosted CJK models do not have that problem, and they run on an Intel Mac from 2019. We chose accuracy on the meetings our users actually have, and we pay for it by having to be precise about where the audio goes. The next post is about that sentence.
What we would ask of every vendor
Name the speech provider and region. Say what text reaches the language model and whether either provider retains it, and name the model provider to anyone who asks. List every other outbound flow, including analytics. Then say what is not local, in the first paragraph, not the footnote. We wrote our trust page to that standard; hold us to it, and hold the others to it too.