Speech:Exps 0283 016

Description
Authors: Jon Shallow James Schumacher

Date: 3/27/16

Purpose: Continue testing new utt files with 145hr corpus. Changed $CFG_N_TIED_STATES to 8000 and $CFG_FINAL_NUM_DENSITIES to 32

Details:
 * Train configuration
 * Corpus: /mnt/main/corpus/switchboard/145hr
 * $CFG_VARNORM = 'no' (variance normalization)
 * $CFG_FINAL_NUM_DENSITIES = 32 (density)
 * $CFG_N_TIED_STATES = 8000 (senones)
 * $CFG_CONVERGENCE_RATIO = 0.04 (convergence ratio)
 * Train start: 12:45 PM 3/27/16
 * Train end: 1:28 AM 3/28/16
 * Decode configuration
 * Decoding on: /mnt/main/corpus/switchboard/145hr/test/trans/train.trans
 * Decoding at: 8000 senones to match the senone count in the train configuration
 * Decode start: 10:13 AM 3/28/16
 * Decode end: 7:50 PM 3/28/16

Results: ,   ---.     |                             hyp.trans                             | |---|    | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins     Err   S.Err | |===================================================================|    | Sum/Avg | 3898  57265 | 77.5   16.2    6.3    7.6    30.2    85.2 | |===================================================================|    |  Mean   |  1.3   18.6 | 79.2   15.8    4.9   15.7    36.5    85.4 | | S.D.   |  0.5   15.9 | 17.2   14.8    7.6   46.2    48.7    33.1 | | Median |  1.0   14.0 | 81.3   13.3    0.0    3.3    29.4   100.0 | `---'

30.2 Is out best result yet. One thing of interest is while we had many very low WER utternaces (some 0%), there also are some greater then 100%. It is unclear how an utterance can have a greater than 100% Error, definitely something that needs looking into.