Speech:Exps 0278

From Openitware
Jump to: navigation, search

Description

Author: Zachery Boynton

Date: 4-20-2015

Purpose: Get more context as to how the decoding process works.

Details: There have been a few issues with subsequent decodes being run on a train that has already been decoded not producing different scoring results when the configurations are changed. Initially, some experimentation took place in 0273, but this directory will now be used.

  • Original tests
    • Original train produces 42.5% error rate on first 1,000 audio files
    • Second and third train also produce 42.5% error rate on first 1,000 audio files, with each train holding different decode configurations
    • Deleted sphinx_decode.cfg in third train and train still ran
    • Deleted sphinx_decode.cfg in first train and ran third train. Third train still ran

For these tests, 5hr trains will be used for the sake of saving time. Decodes will be run on only the first 500 audio files with adequate controls in place to ensure consistency.

The planned tests are as follows:

  • 001
    • Run train and decode with base parameters
    • Run train and decode with altered parameters
  • 002
    • Run train and decode with altered parameters
  • 003
    • Run train and decode after deleting sphinx_decode.cfg
  • 004
    • Run train
    • Alter parameters
    • Run decode

Results: