Speech:Exps 0083

From Openitware
Jump to: navigation, search

Test on train for Experiment 0081 using Last_5hr/test corpus


Description

Author: Group B&C

Date: April 10th 2013

Purpose: To score the acoustic model generated in Experiment 0081 using the Last_5hr/test corpus.

Details: The goal for this experiment was to test the best possible way to remove non-word entries in the transcript. We found in experiment 0081 that by encapsulating each non-word or intelligible word in the transcript with double-plusses ‘++’, we could force the trainer to not take these into account when training. We then added each word that appears encapsulated as such in the phone list, mapping them each to the fake phones +NOISE+”, “+UNINTELLIGIBLE+” and “+LAUGHTER+.

In the previous decode experiment for experiment 0081, the incorrect corpus was utilized; making any comparisons to previous decodes using the last_5hr/test corpus impossible.

Results The experiment dictionary was derived from Experiment 0075. The filler and phone-list was derived from experiment 0081.

The following score was created during this experiment:


                     SYSTEM SUMMARY PERCENTAGES by SPEAKER

      ,-----------------------------------------------------------------.
      |                            hyp.trans                            |
      |-----------------------------------------------------------------|
      | Sum/Avg |  437   6571 | 78.0   16.3    5.8   11.7   33.7   97.3 |
      |=================================================================|
      |  Mean   | 36.4  547.6 | 78.0   16.4    5.6   12.5   34.5   97.6 |
      |  S.D.   |  8.3  142.4 |  4.7    3.5    2.2    4.7    6.9    2.5 |
      | Median  | 32.5  557.0 | 78.0   16.0    5.3   11.4   36.9   97.8 |
      `-----------------------------------------------------------------'

The above score is slightly worse than what we found in experiment 0075. We can determine that either the process or the theory behind the mapping non-word or unintelligible vocalizations to fake phones to be flawed.