Speech:Exps 0111

Description
Author: Eric Beikman

Date: 6-23-2013

Purpose: To determine the effect of forcing the Sphinx trainer use the number of tied states, or Senones, from the default of 1000, to 2000, using a different corpus.

Details: In this experiment, we aim to both replicate the results found within Experiments 0108-0110 on a different corpus. This experiment is based on [Speech:Exps_0024| Experiment 0024], and subsequently uses the same corpus (mini/train), dictionaries, GenTrans.pl version (genTrans2.pl), phone list, and language model. The main difference is that we set the senone value within the sphinx_train.cfg file to 2000 from the default of 1000.

Like Experiment 0024, this experiment will consist of Training, and a subsequent test on dev.

Results No major errors preventing the creation of the models or decoding were observed, other than any issues noted in the Experiment 0024 report.

The subsequent score was created:

,-.     |                            hyp.trans                            | |-|     | SPKR    | # Snt # Wrd | Corr    Sub    Del    Ins    Err  S.Err | | Sum/Avg | 549  10919 | 87.9    6.0    6.1    5.0   17.0   81.2 | |=================================================================|     |  Mean   |  2.9   57.8 | 87.9    7.0    5.1   10.3   22.4   84.4 | | S.D.   |  1.9   45.0 | 13.2   10.2    6.3   19.6   24.8   27.1 | | Median |  3.0   47.0 | 91.9    3.9    3.5    5.4   14.6  100.0 | `-'

Experiment 0024 had a word error rate of about 22.3; by increasing the senone value by 1000, we have observed an 5.3 decrease in total error rates. This roughly correlates to the results we've seen in previous experiments in which we experimented with the senone values.