Speech:Exps 0117

From Openitware
Jump to: navigation, search

Modified Senome Value (4000) test on train 0116 using Fedora


Description

Author: Eric Beikman

Date: 7/9/13

Purpose: Test Experiment 0117.

Details:

During the previous set of experiments, we've determined that setting the senone value higher provides for better models. We wish to confirm our findings using a different 5 hour corpus.

In this Decode, we will be decoding 0.5 hours worth of audio from /first_5hr/test. Using a Language model derived from /first_5hr/train and an acoustic model from experiment 0117.

Results The Decode completed successfully: The following score was created:

                     SYSTEM SUMMARY PERCENTAGES by SPEAKER

      ,-----------------------------------------------------------------.
      |                            hyp.trans                            |
      |-----------------------------------------------------------------|
      | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err |
      |=================================================================|
      | Sum/Avg |  505   7651 | 82.8   12.6    4.6   16.6   33.8   94.5 |
      |=================================================================|
      |  Mean   | 50.5  765.1 | 82.9   13.0    4.2   19.2   36.3   95.2 |
      |  S.D.   | 22.3  319.2 |  5.4    4.0    2.1    9.5   12.6    4.5 |
      | Median  | 41.5  723.5 | 82.5   12.7    3.5   19.0   33.1   96.3 |
      `-----------------------------------------------------------------'

Experiment 0092 was the last compatible decode for this corpus and acoustic model; it had a word error rate of about 44.3. Increasing the senone value appears to have increased the accuracy of the model by a little bit; but not enough for it to be considered acceptable, and certainly not near the level of accuracy of models created using last_5hr.