Please see these calls: 019ead41-30c3-7000-96e3-8667da92999e 019ead3f-7d06-7000-96b9-c0fc568f78f0 The assistant's words constantly have gaps I've tried with a different LLM, like a faster LLM, a different TTS model, and a smaller prompt, and it's still the same issue

Hi Ethan, Thanks for sharing those calls. I looked into both of them. Your pipeline is actually healthy. TTFT is under 1 second and Cartesia is returning first audio in under 250ms. The gaps are not caused by the LLM, TTS model, or prompt size themselves. What I found: the LLM is outputting paragraph breaks (newlines) in the middle of responses. For example, the model generates something like: "...make it more specific than that. [newline] What is the thing no one else is really doing the way you are?" Vapi splits the response at those newline boundaries and sends each piece to Cartesia as a separate audio chunk. Even within the same Cartesia streaming context, there is a brief processing gap at each chunk boundary, and that is what you are hearing as gaps in the words. It happens consistently wherever the model inserts a line break. The fix is to tell the LLM not to format responses that way. In your system prompt, add an instruction along these lines: "Speak in continuous, natural sentences. Never use line breaks, paragraph breaks, bullet points, or any text formatting in your responses." This ensures each response is treated as a single continuous audio chunk, which removes the gaps entirely. Let me know if you need help adjusting the prompt or if the issue persists after making that change. Regards, Chiranjeet Vapi Support