Speech:Exps 0282 003

Description
Author: Matthew Heyner

Date: Started 2-10-2016

Purpose:
 * Generate a train without the use of the Capstone provided scripts to get a better understanding about the training and Decoding process

Details:
 * Training
 * Followed the prepareTrainExperiment.pl line by line running each command manually and altering the configuration files.
 * sphinx_train.cfg
 * - $CFG_DB_NAME = "003" This sets the database name to use.
 * - $CFG_BASE_DIR = "/mnt/main/Exp/0282/003" This sets the location where to find all the needed files.
 * - $CFG_LTSOOV yes Use letter-to-sound rules to guess pronunciations of unknown words (English, 40-phone specific)
 * - $CFG_HMM_TYPE = .cont Set the train to continuous rather than semi-continuous.
 * - $CFG_N_TIED_STATES = 5000 Changed the senone value to 5000 samples
 * - $CFG_NPART = 6 Utilize more CPUs' for possibly a faster train.
 * - $CFG_CONVERGENCE_RATIO = 0.004 The desired number of iterations to be performed, every iteration lowers the convergence ratio until it is between 0.1-0.001 where it will stop. This setting is important to prevent over-training the sample.
 * - $CFG_STATESPERHMM = 3
 * - $CFG_SKIPSTATE = no
 * - $CFG_FALIGN_CI_MGAU = "yes"
 * - $CFG_CI_MGAU = "yes"
 * - $CFG_FORCEDALIGN = "yes"
 * - $CFG_FORCE_ALIGN_BEAM = 1e-60 This is default but taking note for future changes
 * Altered some settings in the feat.params file
 * - -nfilt 31 from 40
 * - -lowerf 200 from 133.33332
 * - -upperf 3500 from 6888
 * - -nfft 256 from 512
 * Training Errors
 * MODULE: 02 Training Context Independent models for forced alignment and VTLN
 * 001.1.1-2.bw.log, 001.1.1-3.bw.log, 001.1.1-4.bw.log, 001.1.1-5.bw.log, 001.1.1-6.bw.log These and all the other log errors are in the Pointing to backward.c
 * MODULE: 03 Force-aligning transcripts - I turned forced alignment on
 * Skipped: No sphinx3_align(.exe) found in /mnt/main/home/sp16/mru567/0282/001/bin. If you wish to do force-alignment, please copy or link the sphinx3_align binary from Sphinx 3 to /mnt/main/home/sp16/mru567/0282/001/bin and either define $CFG_MODEL_DIR in sphinx_train.cfg or run context-independent training first.
 * Found file named s3_align.c withing the bin directory in my experiment. Naming it and its header file to sphinx3_align and running the train again.

Results:
 * Unknown the Model group is running a decode on the 125hr corpus utilizing the resources of Caesar, I will wait until they're finished.