|
From: Lahiru S. <lah...@gm...> - 2015-02-13 03:47:04
|
Hi All, I noticed that the current TIMIT s5 setup has a degraded performance in comparison to earlier results. For example, Karel's DNN with CE now gives 18.5% but earlier it was 17.3%. I think this is due to a change in graph creation or decoding/scoring because when I decoded using the previous tri3 model which gave better results in the current setup, there is a considerable increase in WER. The details of the file with better results are given below. Lahiru svn info RESULTS Path: RESULTS Name: RESULTS URL: https://svn.code.sf.net/p/kaldi/code/trunk/egs/timit/s5/RESULTS Repository Root: https://svn.code.sf.net/p/kaldi/code Repository UUID: 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8 Revision: 4392 Node Kind: file Schedule: normal Last Changed Author: vesis84 Last Changed Rev: 4101 Last Changed Date: 2014-07-02 19:20:57 +0800 (Wed, 02 Jul 2014) Text Last Updated: 2014-09-09 12:24:22 +0800 (Tue, 09 Sep 2014) Checksum: 9a9fd1c9cd921b1fcdea5f847f0bba78 |