Voice surface
Hands stay on the work. Voice covers the rest.
Porcupine wake-word on device. GPT-realtime 2.0 for the conversation. EN to ES code-switching mid-sentence. BIPA-compliant by design.
How voice works
- →
Wake word
Porcupine on-device — zero network round-trip
- →
Voice Gateway
Authenticated WebSocket session
- →
GPT-realtime 2.0
Intent classification + tool routing
- →
Tool execution
clock_in / cost_code_switch / daily_log_append
Henry voice-ack
Spoken confirmation in detected language
Bilingual code-switching mid-sentence
Worker: "Hola ConstructOps, clock me into cost code 04-210, varilla placement."
// EN: "clock me into" → tool: clock_in_cost_code { code: "04-210" }
Henry: "Registrado en codigo 04-210, colocacion de varilla."
BIPA consent flow
- 1Worker opts in at onboarding
- 2Voice print stored encrypted on-device only
- 3Annual retention reminder email
- 4Auto-destroy on offboard
Why Picovoice + OpenAI, not always-on cloud streaming
| Aspect | ConstructOps | Always-on cloud |
|---|---|---|
| Wake word | Porcupine on-device | Always-on cloud streaming |
| Cost at scale | ~$0.003/session | $0.024/min streaming |
| Battery drain | Minimal (local inference) | High (constant upload) |
| Privacy | Voice never leaves device until utterance | Audio streamed continuously |
| BIPA exposure | Minimized (no cloud print) | High (cloud storage) |