Tracing Voice Agents in LangSmith
Key point
LangSmith has launched a feature that provides visibility into voice agents, supporting major frameworks such as Pipecat and LiveKit.
Details
LangSmith has launched Python integrations for four popular voice agent frameworks: Pipecat, LiveKit, OpenAI Realtime, and Gemini Live (Google ADK). As the voice agent market grows rapidly, the importance of observability for managing these systems is also increasing as model precision and naturalness improve.
Voice agents are broadly divided into two architectures. The 'Sandwich' architecture chains together STT, a text-based agent, and TTS, while the 'Speech-to-Speech' architecture processes audio directly through a multimodal model. LangSmith can trace the complex events that occur in both architectures.
With the new integrations, users can view the following key data in a single trace tree:
- Full conversation audio recordings overlaid on the trace
- Latency and metadata for STT and TTS inference
- Voice Activity Detection (VAD) and Interruption events
- Model inputs and outputs, tool calls, arguments, and results
- Timing analysis for each stage of the pipeline
This allows developers to debug voice agent behavior, evaluate performance, and improve voice agents within the same workflow used for text agents.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.