Why not gpt voice directly instead of elevenlabs for voice and sonnet for intelligence?