ADR 0004 — Pipecat, not LiveKit, for the voice pipeline

2026-08-29 · Status: accepted** · Full workings: strategy/31

Context

Both frameworks were built as working POCs and driven with live calls. The choice needed to be made on evidence rather than reputation.

Pipecat. Reasons in order of weight:

Licence. LiveKit's turn detector is Apache-2.0 with a §3(i) restriction that forbids standalone use outside their SDK. We had provisionally planned exactly that. Verified against the licence text.
Frames are the right abstraction here. Money safety needs a processor sitting immediately before the voice, seeing final text after substitution and transliteration. Pipecat's pipeline makes that a ten-line processor.
Native OpenTelemetry, conversation/turn spans, and a behavioural eval harness we now run as a regression suite.

LiveKit's POC is archived at poc/_archive/lk-livekit-poc, not deleted.
We inherit Pipecat's defaults, and its defaults have bitten us three times (text-mode eval bypassing our processor, English TTS/STT in a Hindi suite, float32 Whisper on CPU). Treat every framework default as unset.
An earlier version of this decision cited a latency advantage that turned out to be a measurement artifact. Pipecat's real median (~1,184 ms) is close to LiveKit's 939 ms; latency is not a reason for this choice.

Telephony at 8 kHz is tested end to end, or concurrency requirements exceed what one process per call can serve.