The basic sequence is this:
-
You need an
iss-wsj
working version! All the language models and the like are the same. Ideally it should be../iss-wsj
-
Then, run this sequence of scripts to set up the infrastructure:
CreateLinks.sh # Creates links to the database and to iss-wsj
CreateLists.sh # Creates the file lists in ./local
- Now models can be trained. Replace my-dir with whatever you like.
ExtractTrain.sh my-dir # Extract features for training
TrainGMM.sh my-dir # Train Gaussian mixture HMMs using HTS
- Test the models. Check the scripts to determine which tests (i.e.,
modify
acousticModel
inTestGMM.sh
to reflect whatever you chose asmy-dir
).
ExtractTest.sh my-dir # Extract all 14 test types
TestGMM.sh my-dir # Run all 14 tests
Score.sh my-dir # Score all 14 tests
ISS was originally written for aurora4! This is the paper that eventually resulted:
@Article{Garner2011,
author = "Garner, Philip N.",
title = "Cepstral normalisation and the signal to noise ratio
spectrum in automatic speech recognition",
journal = "Speech Communication",
year = 2011,
volume = 53,
number = 8,
month = "October",
pages = "991--1001",
pdf = "http://publications.idiap.ch/downloads/papers/2011/Garner_SPECOM_2011.pdf"
}
The paper is here.
Phil Garner, April 2012 ...but based on work from a couple of years earlier.