One thing I’ve noticed with voice agents is that the issue usually isn’t the voice itself, it’s the conversation state changing mid-call. If the transcript confidence drops, the AI misreads intent, gets interrupted, or loses context for a second, the TTS can suddenly sound over-energetic or unnatural trying to “recover” the interaction. A few things that helped me: • tightening the system prompt tone instructions • reducing unnecessary enthusiasm words/exclamation patterns • improving interruption handling • checking transcripts for misheard phrases before the weird moment happens • keeping responses shorter during live calls Most times the strange voice moment actually starts earlier in the pipeline than people expect.