Author: Eric Beikman
Date: (Started) 3/28/2013
Purpose: To attempt to replicate training issues being encountered by others. The hypothesis that this experiment aims to test is that the training failures are caused by invalid phones being used in the dictionary and phone list.
Details: This experiment was created in response to issues which others been experiencing during their experiments. Such failures have been reported during making the feats and Modules 40 through 50 of the Sphinx trainer. After comparing the phone and dictionary list of a known good experiment with the phone and dictionary list of a failed experiment, the latter had phones within the phone list and dictionary which were invalid, specifically, they were vowel phones but had no stress indicators at the end. Using this information, we hypothesize that since the invalid phones are listed in the phone list, the checks for such inconsistencies in the first module of the Sphinx trainer is succeeding; but since the phone does technically exist, the sphinx trainer doesn't know what to do with it and errors out.
This experiment will a short train using the mini/eval corpus. The dictionary and phone list have been deliberately sabotaged by removing stress indicators on some of the vowels, then adding these new phones into the phone list.
As the goal of this experiment is failure, no models are expected to be created; furthermore, any models which are created should not be considered valid due to the sabotaged experiment dictionary.
Results Even with the sabotaged dictionary and phone list, neither the Sphinx Trainer or the Feats generation script displayed the behaviors experienced by others. The Trainer did complete and created an acoustic model. We conclude that the Sphinx Trainer will accept phones which do not have stress indicators. As the goal of this experiment was to replicate a failure, we will not be scoring the resulting model.