Speech:Exps 0294 007

From Openitware
Jump to: navigation, search

Description

Author: Matthew Heyner

Date: 5-4-2016

Purpose: Re-run 0288/003 decode on 30hr/test/train/dev.trans with the LW = 6 rather than 13 from 0294/006. This will give us a comparison between the values that CMU recommended both the low end and the high end.

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: Idefix
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 = 6 (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 #1 at 25)
0294/006 WER Unseen 4 = 53.1% (LW #2 at 13)
0294/007 WER Unseen 5 = 59.8% (LW #3 at 6)
The score went down by 6.7% to me this indicates that altering the LW has significant effects on the WER. With more experimentation we could find the optimal configuration of this variable.