VALL-E 2: Neural Codec Language Models are Human Parity Zero-Shot Text to Speech Synthesizers