Context utterances steer style and prosody but are never spoken, so dont put anything in there you expect to hear. And you only need to pass a voice when switching voices.

Context: Official docs (Hume TTS continuation): the context utterances you pass in influence style and prosody of the generated speech, but they are not spoken aloud. Agents that feed a conversation history as context sometimes expect it to be narrated, it wont be. Also you only need to supply a voice when changing from the established voice, otherwise the prior generation context carries it.