Live Audio Tracing Demos

This repo hosts two LangSmith-integrated audio agents:

Demo	Stack	Entry Script	Span Processor
livekit-test/	LiveKit Agents	livekit_test/livekitagents_langsmith.py	livekit-test/langsmith_processor.py
pipecat-test/	Pipecat	pipecat-test/pipecat_langsmith.py	pipecat-test/langsmith_processor.py

Both pipelines stream STT → LLM → TTS and export OpenTelemetry spans enriched with prompts, completions, and conversation summaries so LangSmith shows a clean trace tree.

Prerequisites

• Python 3.13
• Virtualenv or similar (recommended)
• Provider API keys (AssemblyAI, OpenAI, Cartesia, etc.)
• LiveKit RTC + API keys for the LiveKit demo
• LangSmith API key (used by both demos)

Create a .env.local file at the repo root with the following required environment variables:

OPENAI_API_KEY=<api-key>
# other API keys, depending on providers used
LANGSMITH_OTEL_ENABLED=true
OTEL_EXPORTER_OTLP_ENDPOINT=https://api.smith.langchain.com/otel
OTEL_EXPORTER_OTLP_HEADERS=x-api-key=<api-key>,Langsmith-Project=pipecat-voice
# if using LiveKit
LIVEKIT_API_KEY=<api-key>
LIVEKIT_API_SECRET=<secret>
LIVEKIT_URL=wss://demo-xxxxx.livekit.cloud

Install Dependencies

The repo includes a fully pinned environment snapshot. After cloning:

git clone <repo-url>
cd audio

python3.13 -m venv .venv
source .venv/bin/activate

pip install -r requirements.txt

If you prefer to install from pyproject.toml, run pip install -e ., but the requirements file guarantees you match the exact environment used to build the demos.

LiveKit Demo (livekit-test/)

source .venv/bin/activate
LANGSMITH_PROCESSOR_DEBUG=true python livekit-test/livekitagents_langsmith.py

• Forces console mode so you can talk via your microphone in the terminal.
• Uses AssemblyAI universal streaming STT, OpenAI gpt-4.1-mini, Cartesia Sonic TTS, Silero VAD, and the multilingual turn detector.
• The custom span processor aggregates:
  • Full prompts/completions on every llm_node, tts_node, etc.
  • Conversation summaries on both agent_session and job_entrypoint spans (the root is deferred until the flow completes).

Open your LangSmith workspace → latest trace to see the entire hierarchy. Set LANGSMITH_PROCESSOR_DEBUG=false if you don’t want console logs.

Pipecat Demo (pipecat-test/)

source .venv/bin/activate
python pipecat-test/pipecat_langsmith.py

• Demonstrates the same LangSmith instrumentation pattern inside a Pipecat pipeline.
• The Pipecat span processor captures STT prompts, LLM calls, TTS synthesis, and per-turn summaries.
• Recordings saved under pipecat-test/recordings/ are automatically attached to the root span when available.

Troubleshooting

• No spans in LangSmith – check OTEL_EXPORTER_OTLP_* env vars; the scripts warn if tracing is disabled.
• Console exits immediately – LiveKit’s CLI needs a TTY; run inside Terminal/iTerm, not inside certain IDE consoles. Grant microphone access if prompted.
• Missing dependencies – regenerate the environment: rm -rf .venv && python3.13 -m venv .venv && pip install -r requirements.txt.
• Different API vendors – swap STT/LLM/TTS providers in the scripts; the span processors rely only on the OpenTelemetry attributes, so they’ll adapt as long as the spans include lk.* or the Pipecat equivalents.

Project Structure

audio/
├── livekit/
│   ├── langsmith_processor.py
│   └── livekitagents_langsmith.py
├── pipecat/
│   ├── langsmith_processor.py
│   ├── pipecat_langsmith.py
│   └── recordings/
├── requirements.txt
├── pyproject.toml
└── README.md

Feel free to extend either processor (attachments, extra metadata, custom evaluators) and re-run the demos. Everything flows through LangSmith with consistent span semantics, so you can compare the two stacks side by side.