andong-li-speech / cost_aware_enhancement

Communication Aware Microphone Selection For Neural Speech Enhancement with Ad-hoc Microphone Arrays

Communication Aware Microphone Selection For Neural Speech Enhancement with Ad-hoc Microphone Arrays

This repo contains the code to reproduce our results and experiments. Our dataset is based off of the microsoft scalable noisy speech dataset.

Dataset Setup

First, you will need to clone the microsoft scalable noisy speech dataset with:

git clone https://github.com/microsoft/MS-SNSD.git

Then, you will need to edit the path to the microsoft dataset in the desired config within the data_gen_config.py file. Once this is done, you can run the data generation script with:

python data_gen.py --cfg <name of config>

This can take a while to run. Once it has finished generating then you need to edit the paths in your config in train_config.py to point to this new dataset.

Model Training

The model can then be trained with:

python train.py --cfg <name of config> --gpu <which gpu to run on>

To edit the model paramters you can change the config in train_config.py. Make sure that the config points to the dataset you generated earlier.

The website associated with the repo can be found here.