about 600 MB
Speech model
- Model
- Parakeet TDT 0.6B v3
- By
- NVIDIA; CoreML conversion by FluidInference
- License
- CC BY 4.0
- Used for
- Studio: transcripts, read-along, search inside episodes, bookmark quotes
SwatCast is a native Mac app with no server behind it, no account and no analytics. Transcripts, summaries, search and Ask are made on your Mac and kept there.
Transcription
Neural Engine
Summaries and Ask
GPU (Metal)
Finding passages for Ask
Your library database
Voice, Silence Skip and focus sounds
CPU
No model ships with the app. Playback, playlists, bookmarks, ordinary search, voice settings, Silence Skip and focus sounds need none. Studio and Ask show what their model is for, its size and memory use, and download it only when you click. Settings offers two language models; only one stays loaded.
about 600 MB
2.7 GB
Local AI doesn’t mean the app never goes online. A podcast player has to fetch podcasts. These are all the places SwatCast connects to, as listed in Settings → Privacy.
Feeds, artwork and audio, to follow, refresh, stream and download. Refreshes are conditional, so an unchanged feed sends nothing new.
Needed for podcasts to work at all.
Searching for shows and the charts in Discover. If it fails or finds nothing, fyyd and gpodder.net are tried. This is the only connection that sees what you type.
Turn off directory search in Settings → Privacy, and follow by RSS address or OPML instead.
Downloading a model, once, when you choose to. Each file is checked against a pinned SHA-256 before it’s used.
Skip the models and this connection never happens.
Nothing else. No account, no analytics, no crash reports, and nothing about your library, transcripts or questions is uploaded.
Your library, progress, bookmarks, transcripts, downloads and models live in one folder:
~/Library/Application Support/com.swat.swatcast/
Unfollowing a show keeps your progress, bookmarks and transcripts. You can export subscriptions as OPML and any episode as Markdown.