Repository for MS Thesis

I submitted my thesis in June 2023. My research question was whether transfer learning methods can be used to transfer knowledge of L2 speech to improve dysarthric speech recognition. I found that knowledge can be transferred from L2 speech to dysarthric speech recognition, improve word error rates. Multitask learning proved to be the best paradigm to facilitate knowledge transfer.

License

Shield:

This work is licensed under a Creative Commons Attribution 4.0 International License.

Datasets

The thesis uses three datasets: L2Arctic, TORGO, and UA-Speech.

Getting started

To download and use Conda:

1. wget https://repo.anaconda.com/archive/Anaconda3-2022.10-Linux-x86_64.sh

2. sh Anaconda3-2022.10-Linux-x86_64.sh

Note: The scripts to train and submit to HTCondor use source ~/.bash_profile. If sh Anaconda3-2022.10-Linux-x86_64.sh creates or modifies a .bashrc file, then move it's created contents to source ~/.bash_profile.

A sample .bash_profile will look like:

# >>> conda initialize >>>
# !! Contents within this block are managed by 'conda init' !!
__conda_setup="$('/home2/hsteinm/anaconda3/bin/conda' 'shell.bash' 'hook' 2> /dev/null)"
if [ $? -eq 0 ]; then
    eval "$__conda_setup"
else
    if [ -f "/home2/hsteinm/anaconda3/etc/profile.d/conda.sh" ]; then
        . "/home2/hsteinm/anaconda3/etc/profile.d/conda.sh"
    else
        export PATH="/home2/hsteinm/anaconda3/bin:$PATH"
    fi
fi
unset __conda_setup
# <<< conda initialize <<<

3. rm Anaconda3-2022.10-Linux-x86_64.sh

4. source ~/.bash_profile

5. conda env create -f environment-linux.yml