- create a conda environment to setup the experiments
conda create --name comapre-env --file requirements.txt
- activate the environment
conda activate compare-env
We need the wav files and the corresponding labels to set up the experiments. We also need a hardcoded directory structure utilized by the training scripts. You can set up these requirements by running:
./recipes/common/setup.sh
The file recipes/common/feature_generator.py contains the feature generation module. You can run the feature generation process to generate training, developent and test set features by the following command.
./recipes/common/run_feature_generator.sh
For now the above scripts produces log-mel features which are preemphaised and apply a butterworth filter to the input. You will need to modify the generator script to create other type of features.
The file src/seqseq2d.py trains a simple neural network model over the generated features. The python script hardcodes the location of input features and labels. You will need to update to input your own features. A simple execution of this script is given in:
./recipes/common/run_seqseq2d.py
It also includes examples of training with other features.