Speech-to-speech APIs collapse the ASR-LLM-TTS pipeline into one model — cutting latency but concentrating risk in a single dependency worth benchmarking carefully.