Relative decrease in WER is not so significant for lower percentages. How about "we make 6 errors on 100 words but Kaldi makes 7".

It is cool anyone can use CNTK to produce something similar now

23% only