Phoneme classification project using a Multi-Headed Convolutional Neural Network.
Project research paper in LaTeX
- Pytorch (1.7.0)
- Sklearn (0.22.1)
- h5py (2.10.0)
- Numpy (1.18.1)
- SoundFile (0.10.3)
- Librosa (0.8.0)
- tqdm (4.51.0)
* note: The above package version were used for development and testing, but exact versions shouldn't be required.
wget https://data.deepai.org/timit.zip
-
build_timit.py
Takes location of TIMIT train and test directories as argument and creates train and test h5py files used for the training script.
-
train.py
Takes location of train and test h5py files and trains/tests multi-headed CNN for phoneme classification.
-
model.py
Multi-Headed CNN class and helper functions.
-
prep_data.py
Preprocessing steps for preparing train and test data for training PyTorch model.
-
visualize_data.py
Plots examples of raw TIMIT data along with examples of the features used for training (MFCCs, MFCC Distances, MFCC Delta, MFCC Delta Delta).
