It varied a lot. Some speakers (that sound great) only had about 45 minutes of audio. Other speakers had up to five hours.

The key to training is that all of the models were transfer learned from the Linda Johnson speech dataset (LJS).