Basic LiveKit Voice Agent

This guide shows you how to trace a LiveKit voice agent using Noveum Trace. You'll learn how to wrap STT/TTS providers and monitor voice interactions.

🎯 Use Case

Drive-Thru Voice Agent: A voice-powered ordering agent that takes customer orders, uses tools to process items, and responds naturally through speech.

🚀 Complete Working Example

import os
import noveum_trace
from noveum_trace.integrations.livekit import (
    LiveKitSTTWrapper,
    LiveKitTTSWrapper,
    setup_livekit_tracing,
    extract_job_context
)
from livekit.agents import Agent, AgentSession, JobContext, function_tool
from livekit.plugins import deepgram, cartesia, openai
 
# Initialize Noveum Trace
noveum_trace.init(
    project="drive-thru-agent",
    api_key=os.getenv("NOVEUM_API_KEY"),
    environment="production"
)
 
# Define a tool for the agent
@function_tool
async def add_item_to_order(item: str, quantity: int = 1) -> str:
    """Add an item to the customer's order."""
    return f"Added {quantity}x {item} to your order"
 
# Create agent with tools
class DriveThruAgent(Agent):
    def __init__(self):
        super().__init__(
            instructions="You are a friendly drive-thru order taker...",
            tools=[add_item_to_order]
        )
 
# Server entrypoint
async def entrypoint(ctx: JobContext):
    session_id = ctx.job.id
    
    # Extract and enrich trace with JobContext metadata
    # This adds room info, participant details, and session context to traces
    job_metadata = extract_job_context(ctx)
    
    # Wrap STT provider
    traced_stt = LiveKitSTTWrapper(
        stt=deepgram.STT(model="nova-2", language="en-US"),
        session_id=session_id
    )
    
    # Wrap TTS provider
    traced_tts = LiveKitTTSWrapper(
        tts=cartesia.TTS(
            model="sonic-english",
            voice="friendly-voice-id"
        ),
        session_id=session_id
    )
    
    # Create session
    session = AgentSession(
        stt=traced_stt,
        llm=openai.LLM(model="gpt-4o-mini"),
        tts=traced_tts
    )
    
    # Setup tracing with enriched context
    setup_livekit_tracing(session, metadata=job_metadata)
    
    print(f"🍔 Agent connected to room: {ctx.room.name}")
    
    # Start agent
    await session.start(agent=DriveThruAgent(), room=ctx.room)

📋 Prerequisites

pip install noveum-trace livekit livekit-agents
pip install livekit-plugins-deepgram livekit-plugins-cartesia livekit-plugins-openai

Set your environment variables:

export NOVEUM_API_KEY="your-noveum-api-key"
export DEEPGRAM_API_KEY="your-deepgram-api-key"
export CARTESIA_API_KEY="your-cartesia-api-key"
export OPENAI_API_KEY="your-openai-api-key"
export LIVEKIT_URL="your-livekit-url"
export LIVEKIT_API_KEY="your-livekit-api-key"
export LIVEKIT_API_SECRET="your-livekit-api-secret"

🔧 How It Works

1. STT Wrapper

Wraps your speech-to-text provider to trace:

• Audio transcriptions: Full text of what was spoken
• Audio input duration and format
• Processing latency and performance
• Confidence scores and accuracy metrics
• Provider details (model, language, etc.)

2. TTS Wrapper

Wraps your text-to-speech provider to trace:

• Text-to-speech input: Exact text sent for audio generation
• Generated audio metadata: Duration, format, and quality details
• Audio generation time and latency
• Voice and model configuration
• Provider details and parameters

3. Session Tracing

setup_livekit_tracing() automatically traces:

• Agent lifecycle events
• User speech inputs
• Agent responses
• Tool executions

4. Context Enrichment

extract_job_context() enriches traces with LiveKit session metadata:

• Job Information: Job ID, agent name, and dispatch information
• Room Details: Room name, SID, and metadata
• Participant Info: Participant identity and connection details
• Session Context: Timestamps, permissions, and custom metadata

What gets added to your traces:

{
    "livekit.job.id": "job_123",
    "livekit.room.name": "customer-session-456",
    "livekit.room.sid": "RM_abc123",
    "livekit.participant.identity": "user_789",
    "livekit.agent.name": "drive-thru-agent",
    "session.metadata": {...}  # Custom metadata from room
}

Benefits:

• Better filtering: Search traces by room, participant, or job ID
• Context awareness: Understand which user session each trace belongs to
• Debugging: Quickly identify issues for specific rooms or participants
• Analytics: Aggregate metrics by room, user, or agent

📊 What You'll See in the Dashboard

After running the agent, check your Noveum dashboard:

Trace View

• Complete conversation flow
• STT transcriptions with original audio recordings
• Agent LLM calls
• Tool executions
• TTS generations with synthesized audio

Span Details

• Audio recordings: Play back actual STT input audio and TTS output audio
• Full transcription text and synthesis text
• Audio processing times and latency
• Transcription accuracy and confidence scores
• Response timing and quality metrics
• Tool call details

Metrics

• Session duration
• Turn-by-turn timing
• Audio quality metrics
• Cost tracking per operation

🔍 Troubleshooting

No audio traces?

• Verify STT/TTS wrappers are applied
• Check that setup_livekit_tracing() is called
• Ensure session is started with wrapped providers

Missing tool executions?

• Verify tools are defined with @function_tool
• Check that agent has tools in its configuration
• Ensure LLM has access to tool definitions

💡 Pro Tips

1. Use session IDs: Tie traces to user sessions for better context
2. Monitor latency: Track STT/TTS processing times for optimization
3. Extract job context: Always use extract_job_context() to enrich traces with room and participant metadata
4. Add custom metadata: Include user context and business-specific data in trace attributes

🚀 Next Steps

• Explore LangChain integration for chaining
• Learn about LangGraph agents for complex workflows