Attach call recordings to Langfuse traces

[markbackman]

⚠️ Requires Pipecat 1.8.0 before it can be merged / used.

Summary

Adds call recording to the Langfuse example. Each call's audio is attached to its trace, giving two levels of playback in the Langfuse UI:

• The whole call on the trace root — stereo, user left and bot right, so barge-in stays audible.
• Each turn on that turn's turn span — the user's speech as the span's input, the bot's reply as its output.

Langfuse media travels over the REST media API rather than OTLP and is linked to a trace by id, so everything uploads after the pipeline has shut down and the call itself pays nothing for it.

langfuse_recording.py holds the integration and uses only public Pipecat API, so it can be lifted into another bot as a single file. Wiring it up is an AudioBufferProcessor in the pipeline plus three calls:

recorder = LangfuseRecorder.from_env()
if recorder:
    recorder.attach(audiobuffer, worker)

# on disconnect, while the pipeline is still alive
await recorder.stop_and_collect()

# after runner.run()
await recorder.upload()

Credentials are read from OTEL_EXPORTER_OTLP_HEADERS and the host from OTEL_EXPORTER_OTLP_ENDPOINT, since Langfuse authenticates its OTLP endpoint with the same key pair the REST API wants. Media and spans therefore reach the same project by construction, with no second pair of environment variables to keep in sync. LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY / LANGFUSE_HOST override it for a self-hosted media deployment.

Depends on unreleased pipecat

The recording needs pcm_to_wav() (pipecat-ai/pipecat#5326, merged) and per-turn turn audio events (pipecat-ai/pipecat#5329, open). Until both are in a release, pyproject.toml pins pipecat-ai to a local checkout through [tool.uv.sources]; that section and its comment come out when they ship.

This replaces #240, which took the same feature through a helper in pipecat core.

Testing

• Driven end to end against a real pipeline with the Langfuse HTTP calls stubbed: 8/8 clips attached, one whole-call recording and seven per-turn clips across four turns, each on its own turn span and all under one trace id.
• Checked against a barge-in with the mic open through the bot's reply: the interrupted turn keeps the audio it was cut off in the middle of, and open-mic audio does not leak into it.
• Run it for real with uv run bot.py and open the trace in Langfuse; the players appear on the trace root and on each turn span.

[jamsea]

Works great, I can hear the full call audio WAV and the turn by turn splits 🎉