Speech:Exps 0120

From Openitware
Jump to: navigation, search

Test on train 0119 using Automatic Gain Control


Description

Author: Eric Beikman

Date: 7/30/2013

Purpose: This experiment is to determine the effect of Automatic Gain control for both training and decoding.

Details: Automatic gain control is used to ensure that an incoming audio waveform is equalized to a consistent volume level. By default, it is disabled in sphinx. We wish to test this functionality though repeating our current reference experiments, Experiments [Speech:Exps_0089|0089] and [Speech:Exps_0090|0090]. This particular experiment replicates the decode experiment for 0090 for experiment 0118.

Results

The following score was created during this experiment:

      ,-----------------------------------------------------------------.
      |                            hyp.trans                            |
      |-----------------------------------------------------------------|
      | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err |
      | Sum/Avg |  437   6474 | 82.8   13.1    4.1   14.7   31.9   95.9 |
      |=================================================================|
      |  Mean   | 36.4  539.5 | 82.9   13.1    4.0   15.8   32.9   96.0 |
      |  S.D.   |  8.3  143.2 |  4.7    3.8    1.5    5.7    7.7    3.7 |
      | Median  | 32.5  546.5 | 84.4   12.6    4.1   15.6   34.7   95.9 |
      `-----------------------------------------------------------------'

A score of 32.9 is the same as the reference experiment. This proves that AGC has already been applied during recording, or otherwise has no effect on the results.