Downloading & Managing Models
Local transcription models run entirely on your Mac, but they have to be downloaded once first. This page covers the full lifecycle: downloading, selecting, switching, and deleting.
All of it happens in Settings > Dictation Models.
The catalog view
The catalog lists every available model grouped by engine family — Whisper Local, Apple Speech, NVIDIA Parakeet, and OpenAI. Your active model is pinned at the top with its accuracy, speed, and language details.
A filter bar narrows the list:
- Recommended — a curated set balancing speed and accuracy
- Local — on-device models (Whisper, Parakeet, Apple Speech)
- Online — cloud models (OpenAI, requires an API key)
- Fast / Accurate — filter by the model's speed or accuracy rating
Filters combine — selecting Local and Accurate shows only on-device models with top accuracy ratings.
Downloading a model
- Find the model in the catalog. Each card shows its download size, disk size, and RAM requirement (for example, Whisper Base: 142 MB download, 190 MB on disk, ~500 MB RAM).
- Click Download. A progress bar tracks the download; you can Cancel at any time.
- When the download finishes, click Select to make it the active model.
Murmur checks free disk space before downloading and warns if you are low — downloads briefly need about twice the model's size for temporary files. You can proceed anyway or free up space first.
The first time a newly selected model is used it warms up — loading from disk into memory, which takes a few seconds. After that it responds immediately. The recording indicator shows "warming up" when this happens.
Where models are stored
| Engine | Location |
|---|---|
| Whisper Local | ~/Library/Application Support/Murmur/Models/ |
| NVIDIA Parakeet | ~/Library/Application Support/FluidAudio/Models/ |
| Apple Speech | Managed by macOS — nothing to download in Murmur |
| OpenAI | Cloud — no local files |
Parakeet note: Parakeet V2 and V3 share their model files. Downloading either one fetches both, and deleting either removes both.
Switching models
Select a different downloaded model at any time from the catalog — the change applies to your next dictation. Notes and Meetings can each override the global model with their own choice; see the Transcription section in Settings > Notes and Settings > Meetings.
Deleting a model
Click Delete on a downloaded model's card to remove its files and reclaim disk space.
If you delete the model you were actively using, Murmur falls back automatically, in order: Apple Speech, then any other downloaded Whisper model, then OpenAI if you have an API key configured. The confirmation message tells you which engine took over — or warns you if no engine is available, in which case download a model or select Apple Speech before your next dictation.
When a download fails
A failed download shows a Retry button. Before retrying:
- Check your internet connection — model downloads are large
- Check free disk space (at least twice the model's download size)
Downloads resume cleanly on retry; a failed download never leaves a broken model selected.