Speech:Exps 0294 004

Description
Author: Jon Shallow

Date: 4-28-2016

Purpose: Re-run 0288/003 (team cap baseline). We expect to see 29.1% WER on seen data. This is to confirm: A.) 17.5% WER on 0294/001 is due to parameter changes and not any bugs B.) Confirm multithreading decoding does not alter scores. Two decodes will be ran: 1.) Single core using run_decode.pl prior to multithreading additions 2.) 8 cores using run_decode.pl after multithreading additions. We expect to see the same WER as 0288/003 (29.1%) on both decodes. Train will be ran using makeTrain (not makeTest).

Details:
 * Training
 * Orgininal Experiment: 0288/003
 * Server: Idefix
 * Corpus: switchboard/30hr
 * sphinx_train.cfg variable alterations
 * $CFG_NPART = 8 default 2 (0288/003 was ran with 2, this should be the only difference between the two)
 * $CFG_N_TIED_STATES = 4000 default 1000
 * $CFG_FINAL_NUM_DENSITIES = 16 default 8

Decoding
 * Decoding on: /mnt/main/corpus/switchboard/30hr/test/trans/train.trans (seen data)
 * Decoding at: 4000 senones to match the senone count in the train configuration

Results:
 * WER Seen Single Core = 29.0%
 * WER Seen Multi Core = 29.0%