Voice Mode

Voice Mode Design

Optional voice interface for HAL. Off by default. Local-only by design.

---

Principles

• Local-only: no cloud TTS or STT dependency by default
• Explicit: voice input triggers the same router and confirm flow as text — no automatic execution
• Off by default: requires an explicit toggle to activate
• Thin layer: voice is input/output only — the safety model, allowlist, and intent contract are unchanged

---

Non-goals

• No always-on microphone
• No hidden recording or logging of audio
• No cloud voice API calls by default
• No command execution from voice without passing through the router
• No special permissions for voice that text does not have

---

Proposed design (future)

Toggle

mqlaunch hal bridget-on    # enable voice output for HAL responses
mqlaunch hal bridget-off   # disable voice output
mq-hal --voice "kör doctor"  # one-off voice input

Input pipeline

microphone
  ↓
local STT (e.g. whisper.cpp or macOS Speech framework)
  ↓
text prompt
  ↓
mq-hal router (same intent contract as text)
  ↓
confirm step (same as --confirm)
  ↓
execute allowlisted command

Output pipeline

HAL text response
  ↓
local TTS (e.g. macOS say, piper, or coqui-tts)
  ↓
audio output

---

Bridget voice toggle

Bridget is the HAL persona that can respond via voice. The toggle controls whether HAL speaks its responses — it does not change what HAL is allowed to do.

# future mqlaunch menu options
mqlaunch hal bridget-on    # HAL responses are spoken via local TTS
mqlaunch hal bridget-off   # HAL responses are text-only (default)

The Bridget voice toggle is a presentation preference stored in ~/.mq-hal/state.json. It has no effect on routing, allowlists, confirmation requirements, or session memory.

---

Safety constraints

Any voice implementation must satisfy:

• Voice input is processed as text before routing
• The intent contract and allowlist apply identically to voice input
• High-risk intents still require explicit confirmation
• The refuse intent applies to voice the same way as text
• Audio is never stored without explicit user opt-in

---

Implementation status

Not yet implemented. This document records the design intent so future work starts from the right constraints.