Speech:Exps 0294 003

Description
Author: Matthew Heyner

Date: 4-27-2016

Purpose: Re-run 0289/023 decode that using the AM of 0289/019. This is because we discovered the results on unseen data have been tainted.

Details:
 * Training
 * Orgininal Experiment: 0289/019
 * Server: Idefix
 * Corpus: switchboard/30hr
 * sphinx_train.cfg variable alterations
 * $CFG_NPART = 6 default 2
 * $CFG_N_TIED_STATES = 5000 default 1000
 * $CFG_CONVERGANCE_RATIO = 0.001 default 0.04
 * $CFG_FINAL_NUM_DENSITIES = 32 default 8

Decoding
 * Decoding on: /mnt/main/corpus/switchboard/30hr/test/trans/dev.trans
 * Decoding at: 5000 senones to match the senone count in the train configuration

Results:
 * WER Seen = 18.0
 * WER Unseen 1 = 40.8 (This was with having the LM generated off of dev.eval)
 * WER Unseen 2 = 52.8 (Using the proper LM from source exp directory i.e 0289/019)


 * It does appear the LM we were using for 0289/023 gave far more favorable results than what we see in 0294/003 Unfortunatly the results of using the LM in 0289/023 is un-usable because its not a true representation of unseen data.