Modified Senone Value (2500) Train on Last_5hr
Author: Eric Beikman
Purpose: To determine the effect of forcing the Sphinx trainer use the number of tied states, or Senones, from the default of 1000, to 2500.
Senones are the result of the trainer collecting all the HMM states of all Triphones and training it into a single, collective global state. Tied-States is a synonym of Senones. In other words, Senones are multiple phones which are related to each other.
According to CMU, Specifically [This table], The amount of Senones roughly corresponds to the amount of audio data. We are using a 5 hour corpus, as such, the default value of 1000 is inadequate. For this experiment, we shall utilize 2500 senone , as recommended by CMU for the amount of audio data.
The remaining aspects of the experiment was taken directly from experiment 0089. We are using its dictionaries, phone list, and using genTrans5.pl to generate the experiment's transcript.
This experiment is similar to Experiment 0102, except with a larger senone value.
Results The train was able to run successfully. This experiment is continued in Experiment 0108.