Apple Silicon · Neural Engine
The whole pipeline,
on your Mac.
Capture, voice-activity, ASR, diarization, embeddings, summary. Five stages, all on the Neural Engine. The only network call Mac Note Taker makes is to verify your license once a day.
01
Capture
AVAudioEngine + ScreenCaptureKit
02
VAD
FluidAudio voice-activity
03
ASR
Parakeet TDT v3 · ANE
04
Diarization
pyannote-seg 3.0 · ANE
05
Embeddings
CAM++ · 192-d fingerprint
06
Summary
Ollama / your OpenAI key
What never leaves your Mac.
- Raw meeting audio~/Library/Application Support/MacNoteTaker/Recordings/
- Transcript text + timestampsLocal SwiftData store
- Speaker labels + colorsLocal SwiftData store
- Voice fingerprints (CAM++ embeddings)Local SwiftData store
- AI summary (when using Ollama)Local SwiftData store · LLM runs locally
- Calendar attendee namesRead live from EventKit · not stored on our servers
- Settings / preferencesUserDefaults
A closed system.
When you record a meeting, the audio buffer never crosses a network socket. The transcription model runs in CoreML on the Neural Engine. The summary model runs in Ollama on your Mac (or, optionally, via your own OpenAI key - your choice, not ours).
Mac Note Taker makes exactly one outbound call: POST /api/licenses/auth/validate, once per day. That call sends your email + a Mac fingerprint. No audio. No transcript. No summary.
Recording · 00:14:32
Audio buffers in RAM · zero network sockets open
Neural Engine
Parakeet · pyannote · CAM++
Disk
~/Library/Application Support/MacNoteTaker/
Apple Silicon performance
Real numbers, M3 Pro.
ASR latency
120 ms
Parakeet TDT v3 streaming
Diarization throughput
60× real-time
pyannote-seg 3.0 on ANE
Battery / hour of meeting
6-8%
M3 Pro, 16-inch
Ollama 7B summary
8-15 s
30-min transcript → 5 bullets
Audio inside, audio stays.
$149 lifetime · 3 Macs · code FOUDNER for $79.