Speech:Exps 0283 020

Description
Authors: James Schumacher

Date: 5/8/16

Purpose: Create a baseline to compare against 019. This experiment didn't remove the s tags while 019 did. If the scores are equivalent, then the s tags don't matter for training, only for scoring.

Details:
 * Train configuration
 * Corpus: /mnt/main/corpus/switchboard/30hr
 * Default values, except npart is set to 4, not the default 2
 * Train start: 12:42 PM 8 May 16
 * Train end: 1:29 PM 8 May 16
 * Decode configuration
 * Decoding on: /mnt/main/corpus/switchboard/30hr/test/trans/dev.trans
 * Decoding at: 1000 senones to match the senone count in the train configuration
 * Decode start: 1:30 PM 8 May 16
 * Decode end: 2:59 PM 8 May 16

Results: With s tags: SYSTEM SUMMARY PERCENTAGES by SPEAKER ,---.    |                             hyp.trans                             | |---|    | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins     Err   S.Err | |-+-+---|    |===================================================================|     | Sum/Avg | 3912  55254 | 50.3   39.2   10.6    9.5    59.2    91.9 | |===================================================================|    |  Mean   |  1.3   18.1 | 59.2   33.4    7.4   18.2    59.1    91.9 | | S.D.   |  0.5   16.2 | 22.5   19.3    9.4   38.8    40.4    25.4 | | Median |  1.0   13.0 | 56.6   33.3    3.6    6.1    58.3   100.0 | `---'

Without s tags: SYSTEM SUMMARY PERCENTAGES by SPEAKER ,---.    |                             hyp.trans                             | |---|    | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins     Err   S.Err | |-+-+---|    |===================================================================|     | Sum/Avg | 3912  47430 | 42.1   45.6   12.3   11.0    68.9    91.9 | |===================================================================|    |  Mean   |  1.3   15.5 | 43.8   47.1    9.1   40.9    97.1    91.9 | | S.D.   |  0.5   15.8 | 31.0   30.0   12.8  111.8   116.0    25.4 | | Median |  1.0   10.0 | 40.7   46.7    4.0    6.8    74.1   100.0 | `---'

Conclusions: The s tags are definitely needed for training. If you compare these values, respectively, to 019, you will see that they are lower. In addition, I tried running an experiment without the _train.trans file out of curiosity and it failed immediately, so the file itself is absolutely necessary to have.