Note: Our original audio corpora are not stored in this repository (due to their size), but we will supply it upon request.
Our trained neural network models can be found in models/.
Source code can be found in accent-classifier/.
Experiment notebooks can be found in experiments/ (though we do not recommend re-running these).
Under supplementaryMaterial/, we include 1_0_20sec.wav, a 20 second file from Librit (our British audio corpus). In our report, we specify that our neural networks identify the following:
| Phoneme | Word of the phoneme | Time Window in Audio (seconds) |
|---|---|---|
| eu | Feuerbach | 1:49 to 2:19 |
| or | org | 10:86 to 11:39 |
| eu | Feuerbach | 17:65 to 18:41 |
You can play the audio using any open source or commercial audio player to skip to the following timestamps and verify our results. Additionally, the image files showing the visualization of the weights learned by the neural network are available here in higher resolution.
First, clone the repository:
git clone https://github.com/violetguos/project-ift6390.git
cd project-ift6390
We recommend running any code in a virtual environment. This allows for easy installation of dependencies:
sudo pip install virtualenv # This may already be installed
virtualenv .env # Create a virtual environment
source .env/bin/activate # Activate the virtual environment
pip install -r requirements.txt # Install dependencies
# Run any code you need...
deactivate # Exit the virtual environment
A special dependency needed (for formant feature extraction) is the acoustic software Praat, which needs to be added to your system's path (Mac/Windows/Unix).