support for large scale data training(millions or even tens of millions of hours)?

[brainbpe]

The K2 framework does not support large-scale data (millions or even tens of millions of hours) of ASR training, such as the efficiency of multi-machine and multi-card GPU training is not perfect. Are there any plans to improve this in the future?

[kobenaxie]

Actually we can use lhotse like Nemo-lhotse-dataset to load large scale data.