Voice profile persistence
Each voice,
a profile.
Mac Note Taker builds a voice profile for every person you record. The profile starts thin after one meeting and strengthens with every new recording. By meeting ten, re-identification is locked in.
- P
Priya
Founder47 turns · first Jan 18 · last May 28
0.94
cosine
- M
Marko
Eng lead31 turns · first Feb 02 · last May 27
0.88
cosine
- A
Anya
PM22 turns · first Feb 12 · last May 26
0.85
cosine
- L
Lina
Design18 turns · first Mar 04 · last May 24
0.82
cosine
Stored at ~/Library/Application Support/MacNoteTaker/voices.db
What a profile actually stores.
No audio. Not the recording. Not even compressed audio. A profile is a numeric fingerprint that grows tighter each time the model hears more of someone speak.
192-dim embedding
A CAM++ voice fingerprint is 192 floating-point numbers. It captures vocal tract resonance, pitch contour, and speech rhythm - not the words.
Rolling centroid
Each new turn pulls the profile centroid toward the speaker's average. Background noise and bad audio get downweighted automatically.
Sample count + confidence
We track how many turns built the profile and the current intra-cluster cosine spread. Both numbers are visible per speaker.
Confidence curve
Thin after one meeting. Locked in by ten.
The first time we hear a new voice, the profile is a single 90-second average. Misidentifications happen. By the third meeting, the centroid stabilizes. By the tenth, you can rename a 4-second voicemail and we will still match it.
One meeting
Bare profile. Susceptible to room noise.
Three meetings
Centroid stable. Most matches above threshold.
Seven meetings
Different rooms, different mics. Profile generalizes.
Ten+ meetings
Production-grade re-id. Short clips match reliably.
The profile library is yours.
Lives in your app's SwiftData store, on disk, under FileVault. Never replicated, never synced, never our cloud.
Forget any voice in one click
Settings → Voice profiles → Forget. The embedding is deleted. New recordings will treat the speaker as new.
Export and bring across Macs
A profile library is a single binary file. Copy between your 3 activated Macs to share recognition without re-training.
No name leaves the device
Speaker names live in the local store next to the embedding. Mac Note Taker has no telemetry, no analytics, no account sync of voice data.
Where persistent profiles change the workflow
Customer-facing roles
Account managers and CSMs who talk to the same 30 customers on rotation get named transcripts without ever touching the speaker labels.
Recruiting and HR
First-round screens get auto-attributed. By the third candidate touchpoint, the profile recognizes the voice even on a noisy phone line.
Founders + investors
If you talk to the same five LPs every quarter, all four meetings auto-label them and the LP intro line ends up tagged to the right partner.
Legal + deposition prep
Cross-meeting recognition makes after-the-fact transcript review faster - one search by name pulls every utterance across the case file.
Name once. Recognized in every future meeting.
$149 lifetime · 3 Macs · code FOUNDER for $79.