End-to-end training in Kaldi

Hi
I read the 'TOWARDS DISCRIMINATIVELY-TRAINED HMM-BASED END-TO-END MODELS FOR AUTOMATIC SPEECH RECOGNITION' recently, and I am glad to find that the example scripts have been merged into kaldi-master today. This paper proposes a simple yet effective approach to model left biphones(bichars) by creating a trivial full biphone tree, I have question regard to this approach. When training EE-LF-MMI for an language whose number of phonemes is 280 or more, the number of context dependent phones is 280\*280\*2=156800, it is too large. Is there any approach to reduce this number?
@hhadian