examples/voice_agent_webrtc_langgraph/README.md

Voice Agent WebRTC + LangGraph (Quick Start)

This example launches a complete voice agent stack:

• LangGraph dev server for local agents
• Pipecat-based speech pipeline (WebRTC, ASR, LLM adapter, TTS)
• Static UI you can open in a browser

1) Mandatory environment variables

Create .env next to this README (or copy from env.example) and set at least:

• NVIDIA_API_KEY or RIVA_API_KEY: required for NVIDIA NIM-hosted Riva ASR/TTS
• USE_LANGGRAPH=true: enable LangGraph-backed LLM
• LANGGRAPH_BASE_URL (default http://127.0.0.1:2024)
• LANGGRAPH_ASSISTANT (default ace-base-agent)
• USER_EMAIL (any email for routing, e.g. test@example.com)
• LANGGRAPH_STREAM_MODE (default values)
• LANGGRAPH_DEBUG_STREAM (default true)

Optional but commonly used:

• RIVA_ASR_LANGUAGE (default en-US)
• RIVA_TTS_LANGUAGE (default en-US)
• RIVA_TTS_VOICE_ID (e.g. Magpie-ZeroShot.Female-1)
• RIVA_TTS_MODEL (e.g. magpie_tts_ensemble-Magpie-ZeroShot)
• ZERO_SHOT_AUDIO_PROMPT if using Magpie Zero‑shot and a custom voice prompt
• ZERO_SHOT_AUDIO_PROMPT_URL to auto-download prompt on startup
• ENABLE_SPECULATIVE_SPEECH (default true)
• TURN/Twilio for WebRTC if needed: TWILIO_ACCOUNT_SID, TWILIO_AUTH_TOKEN, or TURN_SERVER_URL, TURN_USERNAME, TURN_PASSWORD

2) What it does

• Starts LangGraph dev server to serve local agents from agents/.
• Starts the Pipecat pipeline (pipeline.py) exposing:
  • HTTP: http://<host>:7860 (health and RTC config)
  • WebSocket: ws://<host>:7860/ws for audio and transcripts
• Serves the built UI at http://<host>:9000/ (via the container).

By default it uses:

• ASR: NVIDIA Riva (NIM) with RIVA_API_KEY and NVIDIA_ASR_FUNCTION_ID
• LLM: LangGraph adapter streaming from the selected assistant
• TTS: NVIDIA Riva Magpie (NIM) with RIVA_API_KEY and NVIDIA_TTS_FUNCTION_ID

3) Run

Option A: Docker (recommended)

From this directory:

docker compose up --build -d

Then open http://<machine-ip>:9000/.

Chrome on http origins: enable “Insecure origins treated as secure” at chrome://flags/ and add http://<machine-ip>:9000.

Option B: Python (local)

Requires Python 3.12 and uv.

uv run pipeline.py

Then start the UI from ui/ (see ui/README.md).

4) Swap TTS providers (Magpie ⇄ ElevenLabs)

The default TTS in pipeline.py is NVIDIA Riva Magpie via NIM:

tts = RivaTTSService(
    api_key=os.getenv("RIVA_API_KEY"),
    function_id=os.getenv("NVIDIA_TTS_FUNCTION_ID", "4e813649-d5e4-4020-b2be-2b918396d19d"),
    voice_id=os.getenv("RIVA_TTS_VOICE_ID", "Magpie-ZeroShot.Female-1"),
    model=os.getenv("RIVA_TTS_MODEL", "magpie_tts_ensemble-Magpie-ZeroShot"),
    language=os.getenv("RIVA_TTS_LANGUAGE", "en-US"),
    zero_shot_audio_prompt_file=(
        Path(os.getenv("ZERO_SHOT_AUDIO_PROMPT")) if os.getenv("ZERO_SHOT_AUDIO_PROMPT") else None
    ),
)

To use ElevenLabs instead:

1. Ensure pipecat ElevenLabs dependency is available (already included via project deps).
2. Set environment:
   • ELEVENLABS_API_KEY
   • Optionally ELEVENLABS_VOICE_ID and model settings supported by ElevenLabs
3. Change the TTS construction in pipeline.py to use ElevenLabsTTSServiceWithEndOfSpeech:

from nvidia_pipecat.services.elevenlabs import ElevenLabsTTSServiceWithEndOfSpeech

# Replace RivaTTSService(...) with:
tts = ElevenLabsTTSServiceWithEndOfSpeech(
    api_key=os.getenv("ELEVENLABS_API_KEY"),
    voice_id=os.getenv("ELEVENLABS_VOICE_ID", "Rachel"),
    sample_rate=16000,
    channels=1,
)

That’s it. No other pipeline changes are required. The transcript synchronization already supports ElevenLabs end‑of‑speech events.

Notes for Magpie Zero‑shot:

• Provide RIVA_TTS_VOICE_ID like Magpie-ZeroShot.Female-1 and RIVA_TTS_MODEL like magpie_tts_ensemble-Magpie-ZeroShot.
• If using a custom voice prompt, mount it via docker-compose.yml and set ZERO_SHOT_AUDIO_PROMPT. You can also set ZERO_SHOT_AUDIO_PROMPT_URL to auto-download at startup.

5) Troubleshooting

• Healthcheck: curl -f http://localhost:7860/get_prompt
• If UI can’t access mic on http, use Chrome flag above or host UI via HTTPS.
• For NAT/firewall issues, configure TURN or Twilio credentials.