What's the recommended way to capture full call transcripts + per-turn latencies from a pipecat pipeline?

[HamzaFouad]

pipecat version

No response

Python version

No response

Operating System

No response

Question

hey everyone,

i am building an analytics layer on top of pipecat voice bot and trying to figure out the cleanest approach to capture:

1. full transcript — user utterances and agent responses, each with a start_ms / end_ms relative to the call start
2. tool calls — which function was called, with what args, what it returned, and when
3. per-turn latencies — things like stt latency, llm-ttfb, tts synthesis time, and e2e turn latency (user stops speaking → bot starts speaking)

What I've tried

i first tried adding a custom Observer that oversees the whole pipeline, but I didn’t make much progress and it became very complicated. then implemented a custom ProcessorFrame that intercepts other frames to capture the generated llm transcription and so on, but I ran into many issues and it got complicated very quickly — especially with race conditions between frames and similar problems.

Context

i am looking for any recommendations for examples, patterns, or existing integrations
that would be really appreciated..

[markbackman]

• Transcripts: https://docs.pipecat.ai/guides/fundamentals/saving-transcripts
• User to bot latency: https://docs.pipecat.ai/server/utilities/observers/user-bot-latency-observer
• Tool calls: You have a lot of options. You can use an LLM tracing library or you can use Pipecat's tracing: https://docs.pipecat.ai/server/utilities/opentelemetry#opentelemetry-tracing