MARS: Sound Generation via Multi-Channel Autoregression on Spectrograms