Two ways to turn speech into text, and the differences that actually change your week. The rows that do not favour running it yourself are in here too.
NameSize
2026-08-03-1030-launch.m4a38 MB
2026-08-03-1030-launch.md24 KB
2026-08-04-0930-review.m4a31 MB
2026-08-04-0930-review.md19 KB
Vaultmeetings202608
Upload audio
2026-08-03-1030-launch.m4a
The same recording, on your Mac and in the cloud.
Where the audio becomes text and where the text becomes a note are two separate steps, and a tool can put them in two separate places. Most of the ones that call themselves local are local for the first one only.
A flight, a basement, a conference wifi that has given up. On your own machine the transcript happens anyway, which is the difference between a tool that works and a tool that usually works.
A four hour workshop costs what a ten minute standup costs, so the only meter in the room is the disk. That quietly changes what you bother recording at all.
Three people at once, a room mic across the table, words that exist only inside your company. The larger model in a datacenter is worth the most exactly here, and our own run says so.
Eleven rows, and they do not all go one way: five of them favour the machine on your desk, four favour the datacenter, and two are honestly a draw.
| What is being compared | On your own machine | Sent to a service |
|---|---|---|
| Where the audio goes | Plus. Nowhere. It is read on the disk it was recorded to | Limit. Uploaded, and kept as long as their policy says |
| Offline | Plus. Works on a plane, in a basement, on hotel wifi | Limit. Every minute needs a connection |
| What it costs | Plus. The machine you already own, and nothing per minute | Limit. Per minute, per seat, or a plan with a cap in it |
| Speed | Neither. Bounded by your Mac. Apple silicon does an hour in minutes | Neither. Bounded by the upload and their queue, usually fast |
| Size of the model | Limit. Whatever fits in your memory and your patience | Plus. As large as their hardware allows |
| Difficult audio | Limit. Strong accents, crosstalk and jargon cost you more here | Plus. The larger model is worth the most exactly here |
| Getting better over time | Neither. When you update the app or the model | Plus. They swap the model and you download nothing |
| Long or frequent meetings | Plus. Nothing metered. Disk is the only limit | Limit. The thing every pricing page counts |
| Setup | Plus. A download, once | Neither. An account, and often a key and a bill |
| When it goes wrong | Neither. Slow, or a warm laptop | Neither. An outage or a rate limit, on their schedule |
| Where you can use it | Limit. That one computer, awake, with the app running | Plus. A phone, a browser, Windows, a server |
A datacenter runs a model that does not fit on a laptop, on hardware that is not on your desk, and it can replace that model with a better one overnight without anyone downloading anything. That advantage is not evenly spread. On one person speaking clearly into a decent microphone in a common language, the two approaches land close enough that you would have to look for the difference. The gap opens on the hard parts: heavy accents, three people talking at once, a room mic six feet away, and vocabulary that only exists inside your company.
The cloud is also the only answer if the meeting is not on the machine doing the work. A phone, a Windows laptop, a call you are not in at all. No amount of local processing reaches those.
Because this is the row where our own product is not flattered, the run is written up rather than summarised: the benchmark page has the six meetings, the hand corrected reference, the three hosted services that were compared, and a section on what the run does not cover. Read the method rather than a headline number, including ours.
Offline and the meter are the two scenes at the top of the page. These three only show themselves over months.
Transcription and note writing are two different machines. Speech to text is one model, and reading that text to produce a summary, decisions and action items is another. A product can be entirely local for the first and entirely hosted for the second, and most tools that describe themselves as local, private or on-device are exactly that, because a summary is where a large language model earns its keep.
So the useful question is two questions. Where does the audio become text, and where does the text become a summary? A page that answers only the first one is answering the easier half.
The fully local path for both halves exists. It means pointing the summary step at a model running on your own machine, which is slower and weaker than a hosted one, and it is an honest choice rather than an equal one.
You record on one Mac, you deal with material you would rather not upload, you meet often enough for a meter to matter, or you want the thing to work with the wifi off. Also if the answer to where does this audio go has to be a short one.
Your meetings happen on phones and Windows machines as well, your audio is genuinely hard, you need somebody else to run it while your laptop is shut, or you need a team of people covered by one administrator.
Transcribe locally because it is free, offline and private, then send the transcript text to a good hosted model to write the note, and keep the choice of which one in your own hands.
Aside records and transcribes on your Mac. The speech model ships inside the download, so the recording and the transcript happen offline from the first second, in around 99 languages, with nothing metered and no account involved in that step.
The second machine is the hosted one, and the front page says so rather than burying it: the note is written by Claude Haiku through Aside cloud, which means the transcript text leaves your Mac for that one call. The same goes for asking questions and for chat. Bring your own API key and it goes to your account instead. Point it at Ollama and nothing goes out at all. The full list, including the uncomfortable part, is on the front page.
Whichever way you set it up, your meeting notes are files on your Mac, which is a separate argument worth making.
Last updated 13 August 2026.
The three routes from a recorded meeting to text without an upload, and what each one takes.
Read the guide
Six real meetings, a hand corrected reference, three hosted services and our own engine, with the method in full.
Read the method
One of the hosted tools, with what its free plan caps and where the transcript is made.
Read the comparison
Records and transcribes on your Mac, and every note is a Markdown file you keep.
Requires macOS 26 or later. Apple silicon, or Intel with much slower transcription.