VioLA: Unified Codec Language Models for Speech Recognition, Synthesis, and Translation