Ah, no, TS3-Codec was trained with 10-second audio segments, while BigCodec-S was trained with 2.5-second audio segments (Section 4.5). This was a somewhat tricky (and perhaps debatable) part of the configuration, and we did our best to tune the hyperparameters within the constraints of GPU memory.

Thanks! To the extent that we checked, yes. The important point is limiting the attention window.