Pandecho blog

Offline meeting transcription on Mac and Windows: a setup guide

By Kamal Kannan Sankarraj · October 20, 2026 · 7 min read

Offline meeting transcription means the speech-to-text model runs on your own computer, so your audio is turned into text without being sent to a server. Done properly, you can record and transcribe a meeting in airplane mode. But "offline" is rarely absolute: almost every app still needs the internet for a few things, like downloading the model the first time. This guide explains what offline should mean, lists exactly what still touches the network in Pandecho, and walks through the setup on Mac and Windows.

The short answer

• Transcription offline: yes, once the speech model is downloaded. It runs on your machine.
• Summaries offline: yes, if you use the on-device summary model (the default). Not if you pick ChatGPT or another cloud provider.
• First setup: needs internet once, to download the models (roughly 2 to 3.5 GB in total, depending on which summary model you choose).
• Optional extras that need internet: Google or Outlook calendar sync, license activation, update checks, and anonymous usage data (which you can turn off).

What "offline" should actually mean

When you evaluate any tool that claims to work offline, check three separate steps, because they are often handled differently:

• Capture. Recording your microphone and the audio playing on your computer. This is local in every desktop app.
• Transcription. Turning audio into text. Many popular note takers send audio to a cloud transcription provider for this step, which means it stops when you lose your connection. Granola, for example, says in its own help center that transcription needs a live connection to its transcription providers.
• Summary. Turning the transcript into notes, decisions, and action items. This needs a language model, which is either a big model in the cloud or a smaller one on your computer.

A tool is only truly offline if all three run locally. If only capture is local, you get a recording, but no notes until you are back online.

What runs on your computer in Pandecho

Pandecho runs both models locally by default:

• Speech-to-text: Parakeet V3 (Multilingual), an on-device model of about 650 MB. It supports 25 European languages, including English, German, French, Spanish, Italian, Dutch, Polish, and Portuguese.
• Summaries and chat: a local Qwen3.5 model that runs inside the app. There are two sizes, described in the app like this:
  • Qwen3.5 2B Q4 (about 1.3 GB download): "recommended for most machines (~8 GB RAM)".
  • Qwen3.5 4B Q4 (about 2.7 GB download): highest quality and fewest hallucinations, "needs 16 GB+ RAM". Pandecho suggests this one automatically on machines with 12 GB of RAM or more.

Recordings, transcripts, and notes are saved on your computer. Speaker labels (who said what) are also worked out on the device.

What still needs the internet

To be precise about it, these are the parts of Pandecho that use the network, taken from the app's code rather than from marketing copy:

The free trial is counted per machine on our server, but if your computer is offline the check is skipped and you can keep recording. Meetings recorded offline are counted once you reconnect.

Step by step: set up offline transcription

This works the same way on a Mac with Apple Silicon (macOS 14 or later) and on Windows 10 or 11. Do it while you still have a connection.

1. Download Pandecho and open it.
2. Allow Microphone and System audio when asked. The first lets Pandecho hear you; the second lets it hear the other people on the call.
3. On the Choose how notes are written step, pick On-device model (marked Recommended) and click Download models. You can continue while it downloads; it keeps going in the background.
4. Wait until the transcription model and the summary model both show as ready. After this, you no longer need a connection for meetings.
5. Set your language under Settings → App → Main language. If your meetings mix languages, add the others under Additional spoken languages.
6. Optional: open Settings → Transcription and Settings → Intelligence to confirm that On-device is the provider in use. Under Intelligence you can also download or switch between the two summary models in the On-device models list.
7. Optional, for zero network traffic: turn off Settings → App → Share usage data, and skip connecting a Google or Outlook calendar.

To test it, switch on airplane mode, click Start recording, talk for a minute, and stop. The transcript, summary, and action items are generated on your computer a few minutes after you stop. More detail is in the recording guide and the summary guide.

Honest trade-offs

• Hardware matters. Local models use your computer's memory and processor. On an 8 GB machine you will be on the smaller summary model. Apple Silicon Macs and recent Windows laptops handle this well; very old machines are slower.
• Small models are not frontier models. A 2B or 4B model writes good working notes, but the largest cloud models are better at long, messy meetings. That is why Pandecho also lets you use your own ChatGPT account for summaries if you are comfortable sending the transcript to OpenAI. Either way, check important names and numbers.
• Languages. On-device transcription covers 25 European languages. If you need Japanese, Chinese, or Korean transcription, this model will not cover it.
• Backups are yours. Because your notes live on your computer, back them up the way you back up the rest of your files.

Related reading

• AI meeting notes without a bot: how it works and how to choose
• Pandecho vs Granola: an honest comparison
• KI-Meeting-Notizen lokal (in German)

← All posts