Speech:Exps 0053

From Openitware
Jump to: navigation, search

Description

Author: Eric Beikman

Date: (Started) 3/28/2013

Purpose: To verify the results in Experiment 0024 and 0025 in which models that were created with redundant transcript entries are slightly more accurate than models created with only unique transcript entries. This experiment is the second half of this experiment; the first part is located at Speech:Exps_0026.

Details: To begin this experiment, the experiment directory was set up like a typical training session. We will be decoding using the mini/eval corpus subset. Since we are using the acoustic model of Experiment 0025, an Train was not completed for this experiment. Due to its necessity for the decoding process, a Language model for the mini/eval corpus subset was created for this experiment.


Results The following score was created:

                     SYSTEM SUMMARY PERCENTAGES by SPEAKER

      ,-----------------------------------------------------------------.
      |                         hyp.trans                         |
      |-----------------------------------------------------------------|
      | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err |
      |=================================================================|
      | Sum/Avg |  518  11051 | 46.2   43.9    9.9   11.0   64.8  100.0 |
      |=================================================================|
      |  Mean   |  2.7   58.2 | 46.9   44.7    8.4   16.3   69.4  100.0 |
      |  S.D.   |  1.7   43.0 | 13.6   13.3    6.6   17.5   20.4    0.0 |
      | Median  |  2.0   52.5 | 46.5   44.4    7.7   11.5   67.6  100.0 |
      `-----------------------------------------------------------------'
Summary

Comparing the error rates from this experiment, which utilized the acoustic model from 0025 that had any redundant transcripts removed before training, to that of Experiment 0026, which utilized the acoustic model created in 0024 that had redundant transcripts left in during training, we reach an interesting conclusion. This experiment had a lower overall error rate than Experiment 0026; essentially, this is the reverse to our results from comparing the scores from experiments 0024 and 0025. With this information, we can either assume that there is no tangible benefit between leaving redundant transcripts during training or removing them before training.