How does Vapi add stuttering and natural pauses to voice output? I've been looking into what makes Vapi sound so natural in voice agent use cases, and one thing that really stands out is how it adds stuttering and hesitation to speech. For example: - Input (LLM text):

Do you have a few minutes to chat about how we might be able to help your business?

- Output (transcribed audio):

Do you have a few minutes to *to* chat about how we might be able to help your business?

It literally repeats words mid-sentence to simulate natural human disfluency. Is this implemented as: 1. A rule-based text transformation applied to the LLM output before TTS? 2. An additional LLM layer that rewrites the text to add disfluencies? 3. Something handled at the TTS level directly?