Hyper-parameters are not tuned to optimal
DDAE is reproduced with PyTorch in this repo.
pip install -r v2_py39_torch19/requirements.txt
python v2_py39_torch19/v2_spectrum.py data.h5 list_noisy list_clean
python v2_py39_torch19/v2_train_DNN.py data.h5
python v2_py39_torch19/v2_test_gen_spec.py DDAE.pt list_noisy
conda create --name py27 python=2.7
conda activate py27
pip install tensorflow=1.2
pip install keras=1.2.2
pip install librosa==0.3.1
pip install scikit-learn==0.16.1
- then run commands below ("Getting Started")
- Keras 1.2
- Tensorflow 1.x as backend
- h5py
- librosa
- scipy
Extract spectrogram features:
python spectrum.py data.h5 list_noisy list_clean
Train DDAE using Keras:
python train_DNN.py data.h5
Enhance test wave files using trained model:
python test_gen_spec.py model.hdf5 list_noisy