GOP-LSTM
This repo is the code base of "Improving the Goodness of Pronunciation score by using Deep Neural Networks: Single-input Classification and Sequence-to-Sequence Classification" (2018).
The OHSU Child Corpus is not included.
GOP-LSTM is organized into the following folders:
1. Results
2. features_assembly
3. force_alignment
4. result_recording
5. src
6. utils
7. External: HoldDir
Results
holds AE_phones(AutoEncoder) and CDNN_phones(Convolutional & Deep Neural Networks). These two folders hold the saved results from training and testing DNN in the form of Pickled Python Dictionaries.
features_assembly
contains scripts for creating train and test folders from the Childs' corpus.
forced_alignment
holds the Montreal Forced Aligner(MFA) scripts, custom feature_assembly scripts for MFA use, and alignment evaluation.
result_recording
uses Results to sort and print results, plot results, or create sorted csv results for LaTex tables.
src
contains scripts for creating, training, and testing Neural Network architectures on specified features and sbatch scripts for running on slurm cluster. NOTE: to save Train and Test directories you'll need to setup a "HoldDir" (Hold directory) somewhere on you computer and link in the script file appropriately.
utils
contains additional scripts, such as generators for training, model saving/loading/weight transfer, and attention decoder
HoldDir
holds Train & Test directories to prevent having to . them with git repeatedly