I succeed. I setup pronunciation dictionary on vapi.ai. You need to use eleven_turbo_v2 model. And this solution is not working at First Message. But works great for rest of conversation.