We tested ElevenLabs Scribe v2 and Scribe v2 Realtime as the transcriber on a Vapi assistant doing live NZ inbound reception. Ran them back to back on 5 real phone calls and every single one had turn taking issues that never happen on Deepgram Nova 3. The worst was Scribe v2 Realtime. The agent spoke its entire 4 sentence scripted intro as one TTS burst at t=0 with zero gaps for the caller to answer, then sat on about 47 seconds of dead air before the first tool call. The caller's whole reply arrived in the same second as the bot's whole intro. Looked like the realtime endpoint delivered the user turn as one batch and the LLM fired mid burst. On regular Scribe v2 the pattern was different but just as broken. Our customer.speech.timeout hook (the "no rush, take your time" filler) fired while the caller was actively speaking, twice in some calls, because the HTTP per chunk model has ~700ms round trips and the idle hook can't tell the difference between "caller silent" and "transcriber hasn't POSTed the next chunk yet." Callers ended up saying "hello? hello?" and hanging up.