irasl2018

Code for the paper "Investigating the effect of residual and highway connections in speech enhancement models"

For generating the dataset:

• Download the IEEE dataset and save it to data/IEEE_dataset. Copy sentence lists 01 to 67 to IEEE_dataset/train and 68 to 72 to IEEE_dataset/test.
• Download the DEMAND dataset and save it to data/DEMAND.
• Run the script gen_noisy_dataset.py.
• Use the script create_desc_json_noisy.py to create a JSON file that defines the dataset for the training script.

For training the models, use main.py. For testing the model and generating visualizations, use test_model_vis.py. Both scripts have a list of parameters that define the model type, number of blocks, GRU layers per block, and other parameters.

This code was tested using PyTorch 0.4.1. For generating the dataset, you will need my library maracas.