# Local transcription

**Categories:** Getting started

---

Local transcription turns your speech into text on your own Mac, using the Whisper model, without sending the audio anywhere. It is an option in Retain for Mac. By default, Retain Cloud transcribes your recordings on Retain's servers. This page covers what stays on your Mac, how to turn it on, what it needs, and how it differs.



### What stays on your Mac, and what does not



With Local Whisper selected, the code does this:



* **Audio stays here.** Each recorded chunk is transcribed on the Mac and is not uploaded for transcription. The finished recording is not uploaded either. It is saved on the Mac, and you play it back from this Mac.
* **The transcript goes to Retain.** When you stop, the transcript lines are sent to your Retain account. The summary, the Threads and search are built from that text on the server, as they are for any recording.
* **Timing data goes with it.** The app also sends how loud your microphone and the system audio were in each 50 millisecond frame, the start time and pauses of the recording, and when each word was said. Retain uses these to work out whether each line came from your microphone or from the other people on the call. This is not audio.

So local transcription keeps your voice off the server, not your words. For how Retain handles stored data, see [Privacy and security](https://retainapp.ai/help/article/privacy-security).



### Turn it on



* Open Settings and select **Models**. Under Transcription, choose **Local Whisper (On-Device)**.
* Pick a model from the list: **Large v3 Turbo** (best quality, about 800 MB, the default), **Small** (balanced, about 466 MB) or **Base** (fast, about 142 MB). A smaller model suits a slower Mac.
* Select **Download**. A progress bar shows the percentage, and **Cancel** stops it. When it finishes the row reads Downloaded, with **Delete** to remove the file. If a download fails, the button reads **Retry**.

Downloads go one at a time, and a download keeps going if you leave Settings. The model is stored in the app's folder under Application Support.



### What it needs



* A Mac. The option is not offered on iPhone, where recordings use Retain Cloud.
* Disk space for one model, from about 142 MB to about 800 MB.
* An internet connection for the one-time download of the model, which comes from Hugging Face. Transcribing a recording does not need the network, though saving the transcript to your account and writing the summary do.
* A downloaded model. If you start a recording without one, Retain says "Whisper model not downloaded. Please download a model in Settings first."

### How it differs from the default



* **Retain Cloud** is the default. It uploads the audio in 10-second chunks, the server returns the text, and the finished recording is uploaded too. The settings page says "Cloud offers highest accuracy."
* **Local Whisper** transcribes in 3-second chunks on the Mac, so live text appears in shorter steps, and nothing is uploaded for transcription.
* **Apple Speech (On-Device)** recognizes speech on the Mac for the live text, but the app still uploads each chunk of audio to Retain in the background to keep it. It is not the option that keeps audio local.

You choose the engine for the next recording, and you can switch back at any time in the same place.



Next, see [How to capture meetings with Retain](https://retainapp.ai/help/article/retain-app-how-to-capture-meetings-with-retain).
