LiveKit Omni-Modal Coaching

LiveKit Omni-Modal Coaching

Overview

Real-time multi-modal coaching via LiveKit with:

• Continuous listening: Always-on voice input from participants
• Continuous coaching: Persistent guidance via "The Queen" voice persona
• Stick-breaking modality selection: Poisson-Dirichlet weights determine which modality gets attention
• Dynamic sufficiency gating: ε-machine prevents action without verified skills
• Symbolic expression output: All observations become s-expressions for categorical processing

Architecture

┌─────────────────────────────────────────────────────────────────────────────┐
│  OMNI-MODAL LIVEKIT COACHING SYSTEM                                        │
└─────────────────────────────────────────────────────────────────────────────┘

                        ┌─────────────────────┐
                        │    LiveKit Room     │
                        │  (WebRTC SFU)       │
                        └──────────┬──────────┘
                                   │
         ┌─────────────────────────┼─────────────────────────┐
         ▼                         ▼                         ▼
┌─────────────────┐    ┌─────────────────────┐    ┌─────────────────┐
│  Audio Stream   │    │   Video Stream      │    │  Data Track     │
│  (continuous)   │    │   (screenshare)     │    │  (CRDT sync)    │
└────────┬────────┘    └──────────┬──────────┘    └────────┬────────┘
         │                        │                        │
         ▼                        ▼                        ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                    STICK-BREAKING MODALITY SELECTOR                         │
│                                                                             │
│   ├────────────────┤←────────────┤←────────────────────────────────────────┤│
│       w₁ = 0.45         w₂ = 0.30          w₃ = 0.25                       │
│       (audio)           (video)             (data)                          │
│       SELECTED          fallback            fallback                        │
│                                                                             │
│   Max fraction color: #E12A4E (audio segment wins)                          │
└─────────────────────────────────────────────────────────────────────────────┘
                                   │
                                   ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                      DYNAMIC SUFFICIENCY GATE                               │
│                                                                             │
│   Task: "process audio for coaching feedback"                               │
│   Causal State: (domain=audio, operation=transcribe, tools=(whisper,))     │
│                                                                             │
│   Required Skills:                                                          │
│   ┌──────────────┐  ┌──────────────┐  ┌──────────────┐                     │
│   │ say-narration│  │ signal-msg   │  │ whitehole    │                     │
│   │   (-1)       │  │    (0)       │  │   (+1)       │                     │
│   └──────────────┘  └──────────────┘  └──────────────┘                     │
│                                                                             │
│   ε-Machine: PROCEED (coverage=1.0, missing=0)                              │
└─────────────────────────────────────────────────────────────────────────────┘
                                   │
                                   ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                     THE QUEEN'S VOICE OUTPUT                                │
│                                                                             │
│   Voice: Serena (Premium) - English UK - "Bertha Swirles" persona           │
│   Trit: Computed from stick-breaking max-fraction color                     │
│                                                                             │
│   Output: S-expression for categorical processing:                          │
│                                                                             │
│   (coaching-event                                                           │
│     :timestamp 1735689600                                                   │
│     :modality :audio                                                        │
│     :weight 0.45                                                            │
│     :color "#E12A4E"                                                        │
│     :trit +1                                                                │
│     :observation "participant mentioned confusion about types"              │
│     :guidance "Consider explaining the relationship...")                    │
└─────────────────────────────────────────────────────────────────────────────┘

GF(3) Conservation

say-narration (-1) ⊗ livekit-omnimodal (0) ⊗ whitehole-audio (+1) = 0 ✓

The Queen's Voice

"The Queen" is Serena (Premium) - a British English voice representing Bertha Swirles (quantum physicist). She provides continuous coaching guidance with gravitas and clarity.

# The Queen speaks
say -v "Serena (Premium)" "I observe you're struggling with the type system. Consider that types are propositions and programs are proofs."

Stick-Breaking Modality Selection

Each incoming modality stream is assigned a segment weight via Poisson-Dirichlet:

using Gay.WorldStickBreaking

# Each modality gets a stick segment
modalities = [:audio, :video, :screenshare, :data, :chat]
pd = world_stick_breaking(alpha=1.0, n_segments=length(modalities), seed=session_seed)

# Select dominant modality
selected = world_max_fraction_color(pd)
# => (color="#E12A4E", weight=0.45, index=1, trit=1)

# Audio wins with 45% of attention weight
dominant_modality = modalities[selected.index]  # :audio

Dynamic Sufficiency Integration

Before any coaching action, verify skill coverage via ε-machine:

from sufficiency import EpsilonMachine, Action, CoverageResult

# Create action representing coaching intent
action = Action(
    operation="coach",
    domain="audio",
    language="natural",
    tool="whisper"
)

# Check sufficiency
epsilon_machine = EpsilonMachine()
state = epsilon_machine.infer_state(action)
coverage = epsilon_machine.check_coverage(action, loaded_skills)

if coverage.is_sufficient:
    # Proceed with coaching
    emit_sexp(coaching_event)
else:
    # Load missing skills first
    for skill in coverage.missing:
        load_skill(skill)

S-Expression Output Format

All observations and coaching events are emitted as s-expressions for categorical processing:

;; Coaching event structure
(coaching-event
  :id "CE-2026-01-01-001"
  :timestamp 1735689600
  :session-id "room-xyz"
  
  ;; Modality selection (from stick-breaking)
  :modality :audio
  :weight 0.45
  :color "#E12A4E"
  :trit +1
  
  ;; Dynamic sufficiency result
  :causal-state (audio transcribe (whisper))
  :coverage 1.0
  :sufficient t
  
  ;; Observation from modality
  :observation "participant expressed confusion about monads"
  :observation-embedding #<vector 1024>
  
  ;; Queen's guidance
  :guidance "A monad is simply a monoid in the category of endofunctors."
  :voice "Serena (Premium)"
  :confidence 0.92)

;; GF(3) conservation record
(gf3-triplet
  :minus (say-narration -1)
  :ergodic (livekit-omnimodal 0)
  :plus (whitehole-audio +1)
  :sum 0
  :conserved t)

Required Skills (Dependency Analysis)

Currently Have ✓

Skills to Acquire ✗

Skill Gap S-Expression

(skill-gap-analysis
  :task "livekit-omnimodal-coaching"
  :have (say-narration whitehole-audio dynamic-sufficiency gay-mcp)
  :need (whisper-transcribe livekit-spectral vision-llm crdt-livekit)
  :coverage 0.55
  :sufficient nil
  :action :load-skills
  :priority-order (whisper-transcribe livekit-spectral vision-llm crdt-livekit prosody-analyzer))

Usage

Start Coaching Session

import asyncio
from livekit import api, rtc
from livekit_omnimodal import OmnimodalCoach

async def main():
    # Connect to LiveKit room
    room = rtc.Room()
    await room.connect(LIVEKIT_URL, token)
    
    # Initialize coach with Queen voice
    coach = OmnimodalCoach(
        room=room,
        voice="Serena (Premium)",
        stick_alpha=1.0,  # Poisson-Dirichlet concentration
        sufficiency_threshold=0.8
    )
    
    # Start continuous listening + coaching
    await coach.start()
    
    # Coach emits s-expressions for each observation
    async for sexp in coach.events():
        print(sexp)
        # (coaching-event :modality :audio :guidance "...")

asyncio.run(main())

Emit S-Expression

def emit_coaching_sexp(event: CoachingEvent) -> str:
    """Convert coaching event to s-expression."""
    return f"""(coaching-event
  :id "{event.id}"
  :timestamp {event.timestamp}
  :modality :{event.modality}
  :weight {event.weight:.3f}
  :color "{event.color}"
  :trit {event.trit:+d}
  :observation "{event.observation}"
  :guidance "{event.guidance}"
  :voice "{event.voice}")"""

Local Sand / Italian Woman Mystery

The "Italian woman always talking about local sand" is Emma (Enhanced) or Federica (Enhanced) from the say-narration skill. These Italian voices speaking English are used for agent announcements.

Resolution: The Queen (Serena Premium) replaces Italian voices for coaching output. To stop Italian voices:

# Kill any running say processes
pkill say

# Check which skills trigger announcements
grep -r "say -v" ~/​.claude/​skills/​*/​SKILL.md

The "local sand" might be mishearing "locale" + "sans" (French) from multi-locale voice announcements.

Files

Related Skills

• say-narration - Voice output personas
• whitehole-audio - Audio routing
• dynamic-sufficiency - ε-machine gating
• gay-mcp - Color generation
• iroh-p2p - P2P data sync
• signal-messaging - Fallback messaging