Speech:Exps 0294 006

Description
Author: Matthew Heyner

Date: 5-3-2016

Purpose: Re-run 0288/003 decode on 30hr/test/train/dev.trans with the LW = 13 rather than 25 due to the WER being 74.1%.

Details:
 * Training
 * Orgininal Experiment: 0288/003
 * Server: Idefix
 * Corpus: switchboard/30hr
 * sphinx_train.cfg variable alterations
 * $CFG_FINAL_NUM_DENSITIES = 16 (density)
 * $CFG_N_TIED_STATES = 4000(senones)


 * Decoding
 * Server: Caesar
 * Decoding on: /mnt/main/corpus/switchboard/30hr/test/trans/dev.trans
 * Decoding at: 4000 senones to match the senone count in the train configuration
 * LW = 13 (Language Weight)

Results:
 * WER Seen = 29.1% (Baseline seen from 0288/003
 * WER Unseen 1 = 54.1% (Baseline unseen from 0288/003)
 * WER Unseen 2 = 45.4% (8 core decode WTF?)
 * 0294/005 WER Unseen 3 = 74.1% (LW configuration setup)
 * 0294/006 WER Unseen 4 = 53.1% (LW #2 configuration setup)


 * A 1% improvement on unseen. Next I will create 0294/007 with weight insertion penalty set to 0.06 as well as LW = 13.