Diarization + Embeddings
Pyannote based speaker extraction separates overlapping voices with high precision.
Persistent Profiles
Vector database (Chroma/FAISS) stores voice signatures locally. Recognized across sessions.
Consent-First
Strict enrollment flows. "Who is speaking?" answered locally without biometric cloud upload.
Context Linking
Queryable API allows agents to recall: "What did Speaker A say about the budget last week?"
