ADR 0003 Speech on Ada

ADR 0003 — Speech compute on Ada via Reach

Status: Accepted
Date: 2026-08-03
Milestone: M6 (voice path)

---

Context

Phase 1 originally kept STT/TTS on the Pi (comstar-stt / comstar-tts) so the voice path did not depend on LAN bandwidth and PCM never left the box until a transcript reached AO. AO v1.28 and Reach v0.2 now advertise OpenAI-compatible speech sidecars on WebSocket hello.speech, discovered as SessionBridge.speechClient.

The Pi is already a thin I/O terminal for vision (frames → CPAI). Moving speech compute to the Ada host matches that split while capture, VAD, wake, playback, and the kiosk stay on the Pi.

---

Decision

1. Prefer Reach speech after SessionBridge.start when speechClient != null.
2. Fall back to COMSTAR_STT_URL / COMSTAR_TTS_URL (local Pi units or Mac bring-up) when speech is absent, disabled, or the session is not open.
3. Do not route utterance turns through AO MCP/planner solely for STT, and do not ferry PCM on the Reach session WebSocket.
4. Optional ReachConnectionConfig.speechToken from COMSTAR_SPEECH_TOKEN / AGENTIC_SPEECH_TOKEN when Ada sets a sidecar bearer.
5. Overlay / direct_agent must keep working with speechClient == null.

---

Consequences

• PCM for utterances leaves the Pi over LAN to Ada-advertised STT/TTS URLs — same trust boundary as CPAI and AO (LAN-only, Phase 1). Privacy docs reflect this deliberately.
• Production can stop requiring always-on comstar-stt / comstar-tts once Ada speech is proven; keep those units for Mac/dev and offline fallback.
• Re-bench live bridge fixtures (mic → audio → bridge → remote STT) against the <15s / ideally <6s turn budget.
• Requires AO ≥ 1.28 with AGENTIC_SPEECH_ENABLED=1 and advertised sidecar URLs.

---

Related

• Speech
• Privacy Model
• Contracts — AO Reach
• ADR Index