CosyVoice 2: Scalable Streaming Speech Synthesis with Large Language Models