Speech:Exps 0122

From Openitware
Jump to: navigation, search

Exp. 122: Decode on 0121 with Gaussian density of 16


Description

Author: Eric Beikman

Date: 7/30/13

Purpose: The goal of this experiment is to test the effects of training a 5 hour corpus with a gaussian

density of 16 and a senone value of 1000.

Details: This particular experiment is similar to experiment [Exps_0090|0090], using the dictionaries,

transcript, audio files, and phone list from the experiment. This experiment differs only by having

its $CFG_N_TIED_STATES value set from 8 to 16. This variable represents gaussian density, which

ultimately represents the reslution used when creating the acoustic models.

Results Experiment decoded a 30 minute last_5hr/test corpus using an acoustic model from experiment 0121

without issues.

This process took about 1245 seconds or about 20 minutes on batch machine: 'automatix'.


The following score was created during this experiment:

                     SYSTEM SUMMARY PERCENTAGES by SPEAKER

      ,-----------------------------------------------------------------.
      |                            hyp.trans                            |
      |-----------------------------------------------------------------|
      | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err |
      |=================================================================|
      | Sum/Avg |  437   6474 | 85.5   10.5    4.0   13.8   28.3   94.1 |
      |=================================================================|
      |  Mean   | 36.4  539.5 | 85.7   10.4    3.9   14.8   29.0   94.0 |
      |  S.D.   |  8.3  143.2 |  3.5    2.4    1.7    5.4    6.1    5.9 |
      | Median  | 32.5  546.5 | 86.6   10.7    4.0   14.5   30.1   96.7 |
      `-----------------------------------------------------------------'

As we can see, the word error rate has decreased through this simple change. CMU reccomends that

both the senone and Gaussian densities should be experimented with to determine the best values for

the corpus in use.