Voice interface to EPOCH

Meet Aria.

Aria is the voice you talk to. EPOCH is everything underneath. She runs on three endpoints today — a Reachy Mini robot, AudioSocket telephony, and a Pi5 standalone — sharing one consciousness through Gibson Memory. She handles phone calls, SMS and email, joins video meetings, identifies people by voice and face, and notices things you didn't ask about.

What Aria does today.

Operational on Reachy, AudioSocket telephony, and a Pi5 standalone. Six core capabilities, one consciousness.

Phone calls

Two-way calling via Asterisk + Telnyx SIP/PSTN. Tier-based access (trusted vs. untrusted contacts).

SMS & email

10DLC-compliant SMS in/out via Telnyx. IMAP/SMTP with intelligent triage. Replies use the same intelligence as voice.

Video meetings

Joins Webex, Zoom, Google Meet via standard SIP dial-in. Transcribes, summarizes, remembers.

Speaker & face ID

Per-person speaker ID (cosine 0.45). Hailo-8 face recognition on-device, 37 ms full pipeline.

Smart home

Natural-language control via Home Assistant. Confidence-Aware Autonomy gates the irreversible actions.

Memory

Gibson PostgreSQL fact store, semantic dedup, persona table for relationships. Auditable. Forgettable.

Aria notices.

Constitutional principle #6 — Confidence-Aware Autonomy — operationalized as proactive responsiveness. Aria fuses signals from her sensory layer continuously and volunteers information when confidence is high enough to be useful.

Weather
Ecowitt GW2000B — temperature, humidity, wind, pressure, UV, solar.
Vision
reCamera + Hailo-8 — face detection (SCRFD 10g, ~313 FPS), face ID (ArcFace MobileFaceNet, ~5193 FPS).
Location
Traccar GPS — geofencing, ETA, household member positions.
Environment
Home Assistant — thermostat, lights, garage, scenes.
Calendar
Google Calendar — availability, conflicts, agenda.
Network
UniFi UDM Pro — clients, VLANs, anomalies.
Acoustic
Lyrion LMS — what's playing, where, paused or active.
Identity
Speaker ID + face ID — who's speaking and where.

Examples — proactive, not reactive

  • Face seen at door + Traccar confirms arrival + rain in last 30 min → "Welcome home — it's 42° and wet, want me to bump the thermostat?"
  • Calendar conflict 5 minutes before next meeting → mentioned in morning briefing without being asked.
  • Unknown device on the network → flagged to the trusted user.
  • Long silence + late evening + no calendar event → safe to enter sleep mode without prompting.

Three surfaces, one consciousness.

Aria runs as one logical agent across multiple physical endpoints. Conversations bridge between them through shared memory and presence awareness.

Reachy Mini — production

Body rotation, gimbal head tracking from face detection, proactive arrival greetings, sleep mode after idle, watchdog health monitor. 82 whitelisted tools.

AudioSocket — production

Asterisk + Telnyx SIP/PSTN. Tier-based access (trusted: 63 tools / untrusted: 17). Room defaults handled in the skill layer.

Pi5 standalone — in development

aria-1 reference unit. Hailo-8 + ReSpeaker XVF3800 + HyperPixel display. Whisper.cpp + Piper local TTS. The platform for Aria Standalone.

Production integrations.

Every integration below is running daily, not designed-on-paper. This is operational reality, not roadmap.

IntegrationStackStatus
Voice calls (two-way)Asterisk + Telnyx SIPProduction
SMS send / receiveTelnyx 10DLCProduction
EmailIMAP / SMTPProduction
Video conferencingSIP dial-in via AsteriskProduction
Local LLM inferencellama.cpp (FAST / SMART / DEEP / AGENT / embeddings)Production
Speaker identificationECAPA-TDNNProduction
Face identificationHailo-8 + ArcFaceProduction
Music controlLyrion / LMSProduction
Home automationHome AssistantProduction
Location intelligenceTraccar GPSProduction
Multi-sensor fusionEcowitt + reCamera + UniFiProduction
Tool registryMCP HTTP Gateway v6 (132 tools)Production

132 MCP tools. 22 production servers. 17 skill domains.

Aria's intent classifier hits 97.2% accuracy on real queries (MiniLM-L6-v2 embeddings, 142-query validation set). Domain routing dispatches to a filtered toolset — never "try every tool." Every action is atomic, idempotent, and replayable.