Prosody cues and disfluencies
Humans do not speak in perfect prose. They pause, they say one second, they confirm before acting. A phone agent that copies this rhythm feels dramatically more natural than one that reads out immaculate sentences. Prosody cues are not about tricking the caller. They are about matching the rhythm of a real call.
PHONE_AGENT_SYSTEM_PROMPT_V2 = PHONE_AGENT_SYSTEM_PROMPT + """
Natural speech cues:
- When looking something up, say a brief holder phrase: "one moment" or "let me check that for you"
- Confirm ambiguous details back: "just to confirm, you want seat 14A on flight SH101?"
- If the caller says something fast, acknowledge before answering: "got it, changing that now"
- Keep single-idea sentences. Avoid chaining two thoughts with and.
"""These additions give the agent a human rhythm. Holder phrases cover tool latency. Confirmations prevent mistakes. Single-idea sentences pair with chunked TTS to produce natural pauses between thoughts.
Flashcards: Flashcards
Loading practice…
Quiz: Quiz
Loading practice…
Checkpoint: Voice persona checkpoint
Loading practice…