• Storage
  • System

Siri, dictation and dictionaries are several gigabytes of language data

On device language models are downloaded per language and kept. What each one is, which settings control them, and how much a language actually costs.

5 min read

Most of what makes a modern Mac feel clever runs on the device rather than on a server, and that means the models are on your disk. They arrive quietly after an update or when you enable a feature, and nothing announces the size.

See what your Mac has downloaded

du -sk /System/Library/AssetsV2/com_apple_MobileAsset_UAF_* \
  /System/Library/AssetsV2/com_apple_MobileAsset_Dictionary* 2>/dev/null \
  | sort -n | awk '{printf "%5d MB\t%s\n", $1/1024, $2}'

These scale with how many languages your Mac is set up for, not with how much you use Siri. A Mac configured with three keyboard languages carries three sets of most of this.

What each thing does

  • Siri understanding. The model that turns speech into intent on device, so requests do not need to leave the Mac. The largest of the group.
  • Speech recognition. Offline dictation. If you have ever enabled dictation without a connection, this is why it worked.
  • Text to speech. The voices. Downloading a high quality voice adds a few hundred megabytes on its own.
  • Linguistic data and dictionaries. Text analysis, spelling and definitions, per language.

The settings that actually change the size

  1. Remove languages you do not use. System Settings, General, Language and Region. This is the only lever that removes several of these at once.
  2. Turn off offline dictation. In Keyboard settings, if you never dictate without a connection.
  3. Remove downloaded voices. In Accessibility, Spoken Content, System Voice, where each voice lists its own size.
  4. Turn Siri off entirely. The models are removed when the feature is disabled, at the obvious cost.

Deleting the folders by hand is not the route. They live on system managed storage and macOS redownloads what a feature needs, so the space returns and you have gained a download.

Whether this is worth doing

For most people, no. Four gigabytes of language data is real but it is not why a disk is full, and turning off a feature you use to reclaim it is a poor trade. It is worth knowing about so you can rule it out, which is the same reason the sleepimage file has a page: both are large, both are system managed, and neither is your problem.

The rest of the asset folder, including the 25 GB of simulator runtimes on a developer Mac, is in what Apple's asset packs actually hold. If you want the categories that do move the number, the order to free 100 GB is the sequence.

Common questions

Why does Siri use disk space on a Mac?

Because the models that turn speech into intent run on the device rather than on a server. On a Mac measured in September 2026 they were 2.3 GB, and the total grows with each language your Mac is configured for.

How do I remove language files on a Mac?

Remove the language itself in System Settings, General, Language and Region. That removes its dictionaries, spelling and Siri data together. Deleting the asset folders by hand does not work reliably and macOS downloads them again.

Does offline dictation use storage?

Yes, around 400 MB for the speech recognition models on a measured Mac, plus a few hundred megabytes per high quality voice you have downloaded. Both can be turned off in Keyboard and Accessibility settings.

Is it safe to delete Siri assets?

Not by hand. They are system managed and will be downloaded again when a feature needs them. Turning off the feature or removing the language is the supported way and is the only one that lasts.

Read next