javimosch percept-speaker-id
Speaker identification and management for multi-person conversations.
What it does
Tracks who said what in conversations. Maps anonymous speaker labels (SPEAKER_0, SPEAKER_1) to real names, maintains speaker profiles, and gates voice command authorization.
When to use
- User asks "who said that?" or wants speaker-attributed transcripts
- User wants to configure which people can trigger voice commands
- Agent needs to know who is speaking in a multi-person conversation
Requirements
- percept-listen skill installed and running
- Omi pendant (provides
is_userflag for primary speaker)
How it works
- Omi sends transcript segments with speaker labels (SPEAKER_0, SPEAKER_1, etc.)
- Percept resolves labels to names using the speakers registry
is_userflag from Omi identifies the pendant wearer as the primary speaker- Speaker profiles track first/last seen timestamps and authorization status
Speaker registry
Located at percept/data/speakers.json:
{
"SPEAKER_00": {
"name": "David",
"is_owner": true,
"approved": true
},
"SPEAKER_01": {
"name": "Rob",
"is_owner": false,
"approved": true
}
}
Manage via Percept dashboard (port 8960) → Settings → Speakers.
Authorization levels
- Owner (
is_owner: true): Full command access, always authorized - Approved (
approved: true): Can trigger wake word commands - Unknown: Logged only, commands not executed
Future: Voice embeddings
Planned: pyannote speaker diarization with 192-dim voice embeddings for automatic speaker recognition via cosine similarity. Currently speaker mapping is manual.