How to transcribe confidential client meetings without uploading the audio

What does "without uploading audio" actually require?

The speech-recognition model has to run on your computer — not just the recorder. A lot of tools market themselves as private because nothing joins your meeting, but a bot-free recorder that sends audio to a cloud server for transcription has not kept the meeting local; it has only changed where the upload happens. Notta's Privacy Mode documentation is explicit about the distinction: in Privacy Mode, "data is not uploaded to the cloud" — the model runs on your machine and the recordings stay in a folder you choose. That is the standard to hold every tool to. If you want the full background on the three architectures — built-in meeting summaries, cloud note takers and local transcription — see our guide to built-in vs cloud vs local meeting transcription. This page stays on the consultant's workflow: what to install, what to record, and what to check.

Which tools transcribe meetings on the device?

Only a handful of tools genuinely record and transcribe a live meeting on your computer. Notta Desktop's Privacy Mode is the most complete option for consultants: it works on both Mac and Windows, records online calls and in-person audio, transcribes on the device with unlimited local transcription on paid plans, and pauses cloud summaries, sync and sharing while it is on. It offers two engines — an offline model covering English, Japanese, Simplified Chinese and Cantonese, and Apple's Speech Analyzer engine (11 languages) on macOS 26 and later. MacWhisper is the best free option on the Mac, running local Whisper and Parakeet models and recording Zoom, Teams and Webex calls; its Pro licence (€64 one-off) adds more. Buzz is a free, open-source Whisper app for Windows, Mac and Linux that transcribes files offline — useful for recordings made elsewhere, but it does not capture live meetings. Krisp's feature matrix lists on-device transcription on Enterprise plans, but this is unverified, so confirm before committing.

For the full ranked list, see our guide to the best offline transcription software for Mac and Windows.

Are bot-free note takers and AI voice recorders local?

Usually not. Bot-free note takers avoid the visible meeting participant, which clients often prefer, but most still send audio to a cloud server for transcription. Granola is the clearest example: it transcribes in real time in the cloud and keeps no audio afterwards — a genuine privacy improvement, but not on-device transcription, and worth knowing before you tell a client nothing leaves your machine. AI voice recorders such as Plaud's Note Pro and Notta's own Memo can record while offline, but transcription and AI features happen after the device syncs to the cloud. Neither is a fit for a contract that requires the audio to stay local. If your concern is simply avoiding a bot rather than avoiding the cloud, our guide to transcribing sensitive meetings without a bot covers those options.

Can you get an AI summary without the cloud?

Only with a local language model, or by writing it yourself. Notta's cloud AI summaries are paused while Privacy Mode is on — that is the trade-off you accept for keeping everything on the device. MacWhisper can generate summaries, and its Pro version can switch to local models so the text never leaves your Mac, though local summarisation quality depends on your hardware. If you want to understand what that takes, see our guides on how much RAM you need for a local LLM and what an NPU is. For most consulting deliverables, the pragmatic route is a hand-written summary of a transcript you have already checked — you control the emphasis, and there is no model to trust.

How do you record an in-person client workshop locally?

Record on your laptop's microphone, or use an offline recorder and transfer by cable. For a workshop or in-person interview, open Notta Privacy Mode on the laptop and let it capture the room — the recording and transcript both stay on the device. If the room is large or the laptop would be intrusive, a dedicated voice recorder is more discreet, but check where it processes: many popular AI recorders, including Plaud's, transcribe only after syncing to the cloud. Copy files from the recorder to your computer by cable, not through the recorder's app, then transcribe locally with Notta, MacWhisper or Buzz. Our roundup of the best AI voice recorders has more options, with the cloud caveat noted.

How accurate is local transcription compared with cloud?

Local models can be good on clear audio, but they still mis-hear, and they can invent text in pauses. The Careless Whisper study (Koenecke et al., ACM FAccT 2024) found that roughly 1% of Whisper transcriptions contained hallucinated phrases, often clustered around silence — adding invented content to a confidential client transcript is a different kind of risk from a cloud provider seeing it. Cloud services generally hold an edge on accented speech, crosstalk and noisy rooms; local engines close the gap on clean, well-mic'd audio. Whichever you use, spot-check the transcript before it becomes a deliverable: names, numbers, commitments and anything you plan to quote. For most consulting work, a checked local transcript beats an unchecked cloud one.

What still needs the client's consent?

Recording itself. On-device transcription removes the third-party cloud from the picture, but it does not remove the legal and professional obligation to tell people they are being recorded — and for a consultant, the reputational obligation is at least as strict. Get consent in writing where the NDA requires it, and state it verbally at the start of the call otherwise. It is also worth telling the client what happens to the recording afterwards: where it is stored, that it is encrypted, and when it will be deleted. Our guide to recording consent laws covers the legal requirements in more detail.