---
title: "Speaker Labels and a Local Summary Model on Mac — Mumr"
description: "Mumr 1.13.0 tells the voices on a call apart and can write summaries with a local model powered by Mirai. Both run on your Mac, with no bot in the call."
canonical: "https://mumr.app/blog/speaker-labels-local-model"
language: "en-GB"
last-updated: "2026-11-10T00:52:40+00:00"
source: "https://mumr.app — Markdown rendition of the HTML page at `canonical`"
---

Product news

# Speaker labels and a summary model that run on your Mac

Mumr 1.13.0 for Mac tells the voices on a call apart and lets you name them. It can also write your summaries with a language model that runs on your Mac, [powered by Mirai](https://trymirai.com). Neither needs a bot in the call or a server of ours.

Last updated: 9 November 2026

![A dark grey three-legged conference speakerphone on a slate table, a small violet light on its edge, lit from behind by a violet rim light.](https://mumr.app/changelog/img/1-13-0.webp)

Until now a Mumr transcript knew two speakers: You, from your microphone, and Others, from the call. A meeting with four people read as you against a wall of “Others”. Note takers that name every speaker often get there by joining the call as a bot, or by uploading the recording to their own servers. Mumr does neither. It already records the call from your Mac’s own audio, so the work of telling voices apart can happen in the same place.

## Who said what

![Three folded white name cards on a dark slate table, reading Anna, Mark and You, lit from behind by a violet light.](https://mumr.app/changelog/img/1-13-0-speakers.webp)

Speaker labels need a separate model of about 221 MB, for Macs with Apple silicon. Mumr downloads it by itself once the speech model is ready, with progress and Cancel in Settings → Transcription; Cancel or Delete there stops automatic downloads for good. Mumr checks every file against a fixed size and checksum before it installs it.

From then on, when a meeting ends and its transcript is final, Mumr separates the voices on the call on your Mac. The transcript reads Speaker 1, Speaker 2 and so on, in the order they first spoke, and your own lines stay marked You.

-   **Names.** Click a speaker in the transcript and type a name. If the meeting has a calendar event, its organizer and attendees are offered as suggestions.
-   **Corrections.** Merge two voices that turn out to be the same person, or move a single line to someone else. Changes apply to that meeting only, and they survive Regenerate transcript.
-   **Where names go.** Exports, Copy Transcript, the PDF, Send to Participants and the hand-off to Claude and GPT all use the names. The summary uses only names you gave, never “Speaker 2”.
-   **Your voice.** Mumr can learn your voice from meetings on headphones, or from a one-minute recording you make in Settings. After that, the people your microphone hears through speakers or across the table get speaker numbers of their own, so a meeting in a room gets labels too. Reset makes Mumr forget your voice.

## A summary model on your Mac

![A space-black laptop open on dark slate, its screen dark with a soft violet glow, a pair of white wired earbuds beside it.](https://mumr.app/changelog/img/1-13-0-local-model.webp)

Mumr has always transcribed on your Mac. The summary was the part where you chose: Quick highlights, which picks sentences from the transcript; Apple Intelligence on macOS 26; or your own Claude, GPT or OpenRouter key, which sends the transcript to that provider under your account.

The local model is a fourth option. It writes the same kind of summary as Claude or GPT, from the same prompts, including any you customised, and it titles each meeting. The transcript does not leave your Mac, and it works with the network off.

-   **Which model.** Qwen 3.5 4B by Alibaba, about 2.9 GB, on Macs with 16 GB of memory or more. On a Mac with 8 GB, Mumr offers Qwen 3.5 2B instead, about 1.1 GB. It is faster but can get names, owners and dates wrong, and Settings says so beside it.
-   **Which Mac.** Apple silicon and macOS 26.4 or later.
-   **Turning it on.** Open Settings → AI and click the Local model tile. A sheet names the two companies involved and what each one sees, and nothing downloads until you agree. When you set up Mumr on such a Mac, the AI page offers the local model first.
-   **When it cannot finish.** If a meeting is too long for the model, or the model stops, Mumr writes Quick highlights instead and says why on the summary.

## [Powered by Mirai](https://trymirai.com)

The local model runs on [uzu](https://github.com/trymirai/uzu), the inference engine [Mirai](https://trymirai.com) builds for Apple silicon and publishes as open source under the MIT licence. Making a language model run well on a Mac is the hard part of a feature like this, and [Mirai](https://trymirai.com) did it. Thank you to the [Mirai](https://trymirai.com) team for building uzu and for making it open.

If you build apps for Apple devices and want a model to run on them, [Mirai](https://trymirai.com) is worth a look: [trymirai.com](https://trymirai.com).

## What leaves your Mac

Recording, transcription, speaker labels and local-model summaries all happen on your Mac. Mumr runs no server that receives your audio, transcripts or summaries, and you need no account.

Two outside services take part, and neither sees a meeting. The model files download from Hugging Face. Each time the engine starts, it sends [Mirai](https://trymirai.com) your macOS version, chip and memory size so [Mirai](https://trymirai.com) can pick the right build for your Mac. The [privacy policy](https://mumr.app/privacy) lists both.

## Also in 1.13.0

-   Settings has a new design, with a sidebar like System Settings. Options sit on cards, each pane has its own coloured badge, and switches, progress bars and your chosen engine take Mumr’s violet.
-   Settings → AI shows every summary engine as a tile that says where the transcript goes and what the engine writes, with the one in use marked.
-   Automatic summaries have their own switch, Summarize after each meeting.
-   If labelling the speakers of a meeting fails, Separate speakers in the Regenerate menu runs it again.
-   Your lines and the call’s lines now appear in the order they were spoken.

Every detail, setting name and requirement is in the [1.13.0 release notes](https://mumr.app/changelog/1-13-0).

## Name the voices on your next call

Mumr is free on the Mac App Store. Version 1.13.0 arrives as an update.

[Download for Mac](https://mumr.app/download)

On the Mac App Store · macOS 15 or later
