Mengyuan Zhao 2015-10-12

• LSTM training, Training on large dataset nearly done, get some conclusions:

1. MPE cannot get significant performance improvement as DNN.
2. MPE is easy to diverge.
3. Training seems over-fitting with current network config (2*512). wer on tr-set is 2%(abs) lower than cv-set.
4. 4-layer LSTM is better than 2-layer. Still testing differnt network config on 120h dataset.

• Reproducing "Self-informed nnet structure" on the paper cvss-464.

1. Looking into kaldi code in detail and modifying it.