It’s not you Vapi’s current TTS engine doesn’t actually execute real pauses from SSML or on most voices, even when enabled. The only reliable workaround I’ve gotten to work is splitting the script into separate assistant messages so the engine is forced to stop and wait before speaking the next part.