Mengyuan Zhao 2016-05-16

• sino 8k training:

1. Finish 1000h mpe training for the model with noise phone. result looks ok (significant deletion error decrease, and similar WER compared with old model).
2. Start adaptation using PingAn-All data.

• 16k embedded model training:

1. 5-500 mpe model training done (without svd), delivered to sino and to be tested.
2. 5-500 mpe model with svd, on training.