This recipe is trained on LDC2013S08 (text transcripts from LDC2013T20) which is
 Gale Phase 2 Chinese Broadcast News speech: 126 hours of of Mandarin Chinese
 broadcast news speech collected in 2006 and 2007 by LDC and HKUST.

 There is no separate test set; we just use 6 hours held out from the training
 data, to test on.

 The recipe is in s5/.