Thymia Sentinel

Voice AI safety monitoring through multimodal biomarker analysis.

Sentinel streams voice conversations to Thymia's Lyra server for real-time extraction of clinical speech biomarkers, combined with policy-based safety reasoning to detect mental health concerns that text-only systems miss.

Why Multimodal?

Text-only safety moderation has two fundamental failure modes:

1. False negatives (minimization): Users experiencing mental health crises frequently downplay their distress verbally. "I'm fine, just tired" may be spoken with voice biomarkers indicating severe depression.

2. False positives (alarm fatigue): Phrases like "I'm dying of embarrassment" trigger crisis pathways despite no clinical concern, leading to desensitization and wasted resources.

Both failures stem from the same limitation: reliance on semantic content without physiological ground truth. Sentinel addresses this by combining speech biomarkers with conversation analysis through explicit concordance checking.

Installation

pip install thymia-sentinel

Quick Start

from thymia_sentinel import SentinelClient

async def handle_result(result):
    policy = result["policy"]
    policy_name = result.get("policy_name", policy)
    if policy_name == "demo_wellbeing_awareness":
        classification = result["result"]["classification"]
        level = classification["level"]  # 0-3
        alert = classification["alert"]  # none, aware, supportive, mindful

        if level >= 2:
            print(f"Risk level {level}: {alert}")
            print(f"Recommended: {result['result']['recommended_actions']['for_agent']}")

sentinel = SentinelClient(
    user_label="user-123",
    policies=["demo_wellbeing_awareness"],
    biomarkers=["helios", "apollo"],
    on_policy_result=handle_result,
)

await sentinel.connect()

# In your voice AI audio loop:
await sentinel.send_user_audio(user_audio_bytes)      # PCM16 @ 16kHz
await sentinel.send_agent_audio(agent_audio_bytes)
await sentinel.send_user_transcript("I've been feeling okay lately")
await sentinel.send_agent_transcript("That's good to hear. How has your sleep been?")

# When done:
await sentinel.close()

Configuration

Demo Policies

• demo_wellbeing_awareness: Wellbeing awareness analysis with risk classification and recommended actions
• demo_field_extraction: Extracts basic user fields (name, age) from conversation

Biomarkers

• helios: Wellness indicators (distress, stress, burnout, fatigue, low self-esteem)
• apollo: Clinical disorder probabilities (depression, anxiety) and symptom-level severities
• psyche: Real-time affect detection (happy, sad, angry, fearful, etc.)

Awareness Levels

The wellbeing awareness policy returns a 4-level classification:

Framework Integrations

For framework-specific examples, see the examples directory:

• LiveKit Agents - Automatic audio capture from LiveKit rooms
• Pipecat - Integration with Pipecat pipelines
• VAPI - WebSocket integration for VAPI
• Gemini Live - Google Gemini Live API integration

License

MIT License - see LICENSE for details.