For downloading data, please refer to download.py.
For generating training data, please refer to generate_training_data.py.
For loading data in shar format, you can refer to this notebook.
The training set consists of all data except the dev/test sets and the doreco dataset. The test set consists of the test sets from librispeech, aishell, mls, and the whole doreco dataset.
Note that the license of magicdata does not allow us to redistribute the data and does not permit commercial use.