Privacy
What Halen hears, what it keeps, and the only two things that ever touch the network.
Last updated 4 October 2026 · applies to Halen 1.0
What it hears
Your microphone, and only while a call is on (or you press Listen Now). It never saves or transcribes the other side of the call (see below).
If you set up the voice check, every second and a half of audio is compared with your voice print. Anything that isn’t you is wiped from memory before it’s analysed. If a moment of audio can’t be checked, it’s dropped rather than kept. If the voice model itself fails to load, Halen tells you, and other voices near your mic may be included in that call’s numbers.
Knowing when they’re talking
To measure talk share and interruptions, Halen watches the call app’s output volume: one loudness number at a time. The sound itself is thrown away instantly and never transcribed. macOS asks your permission for this; say no and Halen still works, just with fewer listening stats.
What it keeps
| Data | Kept | Where |
|---|---|---|
| Audio | Never. Analysed in memory, then discarded | — |
| Scores and stats (pace, talk share, …) | Until you delete them | Encrypted on your Mac |
| Your own words | Off by default. If you turn it on: 1, 7 or 30 days | Encrypted on your Mac |
| Your voice print | Until you delete it | Encrypted on your Mac |
| Your ratings and notes | Until you delete them | Encrypted on your Mac |
Files are sealed with AES-GCM using a key that lives in your Mac’s keychain and never leaves this device. Saving your words is only possible once the voice check is on, so other people’s words can’t end up in it. Settings › Privacy › Forget Everything deletes it all, key included.
Network traffic
Three things, and nothing else:
- Update check. Once a day Halen fetches
halen.dev/eq/appcast.xml. Updates are signed, and verified before they install. You can turn this off in Settings. - Voice model download. The first time you set up the voice check, a small speaker-recognition model (about 30 MB) is downloaded from Hugging Face. It runs locally from then on.
- Speech model. The first time, macOS may download Apple’s on-device speech model for your language. That’s Apple’s system download; your audio isn’t part of it.
No analytics, no crash reports, no telemetry. The code is open, so you can check every line.
Coaching text
The one-line takeaway after each call is written by Apple’s on-device model (Apple Intelligence), from the numbers only. If Apple Intelligence isn’t available, Halen uses a simple built-in template instead. Nothing is sent to a server.
Permissions, summarised
| Permission | Why |
|---|---|
| Microphone | Hear your side of the call |
| Speech Recognition | Turn your speech into text on-device, to measure pace and fillers |
| Audio capture (optional) | Measure how loud the other side is, never what they say |
Recording laws
Halen is designed to analyse only your own voice and to save no audio. Laws differ by place, and this isn’t legal advice. Speakers and shared rooms can let other voices reach your mic, so if you’re somewhere that requires everyone’s consent, tell people you use it.
Contact
Questions? Email the maintainer or open an issue at github.com/lukataylo/halen.