Questions about the soundstorm paper

[olup]

If I understand correctly, they apply the maskGIT technic to audio. MaskGIT predicts tokens parallelly in multiple pass. It works because in image you have a predefine number of token. How does soundstorm do ? Conceptually, as a sequence to sequence multimodal transformer, it's hard to get free from the autoregressive approach. Do they arbitrarily decide on an audio output length ? In that case, what happen with too long/short text prompt (in the case of tts) ?