Everything we get asked.

MeetPing matches against confirmed Parakeet output (not the volatile partials), with three strategies any-of: Unicode-aware word-boundary regex (catches exact ASR output, no substring false-fires), Soundex phonetic equality (so "Ogtay" still triggers when Parakeet hears "octay"/"october"), and bounded Levenshtein for typo-style mutations. Each match carries a confidence band so the popover tells you which strategy fired.

Yes. MeetPing reads from the system mic and (with screen recording permission) the system audio output, so it doesn't care which app produced the audio. Auto-arm detects all common meeting apps via bundle id + Chromium tab URL.

Parakeet TDT v3 supports 25 European languages (English, Spanish, German, French, Italian, Portuguese, Dutch, Polish, Czech, Hungarian, Romanian, etc.). All run on the same on-device model — no language pack downloads.

Approximately 1.4–2.0 seconds end-to-end on M1+: that's the chunk window (3s with 1s hypothesis chunks) plus the keyword scan plus the alert hop. Tight enough to glance at the popover and respond without breaking flow.

Parakeet on the Apple Neural Engine is designed for sustained streaming. Expect ~3–5% additional CPU on M1 Pro during an active meeting; the ANE itself is power-efficient and doesn't burn the main cores.

Likely yes. v0.1 has no backend, no telemetry, and zero external network calls. The sandbox is off only because Core Audio Taps + FluidAudio's BNNS dylib loading require it; Hardened Runtime entitlements are documented and minimal.

Not in v0.1.0 — we're starting opinionated to validate the UX. If your meetings have a different cadence (say, you need 60s of past context for product reviews), open an issue and we'll add it.

v0.1 covers the at-desk case (notification, sound, screen flash). Remote alerts ship in v0.2 — we're picking the channel (Telegram, Pushover, Slack DM, ntfy.sh) based on what users actually ask for in the first month. Email [email protected] if you have a strong opinion.