Source Separation Project
Run main.py to train the model. Use TensorBoard to view results.
### Data Setup ###
# WSJ0
1) Download wsj0.tar.gz (https://catalog.ldc.upenn.edu/ldc93s6a)
2) tar -xf wsj0.tar.gz -C data/wsj0_sph/
2) Install sph2pipe (https://www.ldc.upenn.edu/language-resources/tools/sphere-conversion-tools)
3) python preprocess_wsj0.py (requires sph2pipe)
# LibriSpeech
1) Download test-clean.tar.gz and train-clean-100.tar.gz (http://www.openslr.org/12/)
2) tar -xf *clean*.tar.gz -C data/
3) python prepocess_libri.py
# RealTalkLibri (RTL)
1) Download rtl.tar.gz
2) tar -xf rtl.tar.gc -C data/
### Code Overview ###
# Hyperparameters
* hyperparameters - General Hyperparameters
* CNNparameters - CNN specific parameters
* RNNparameters - RNN specific parameters
# Core
* main - Construct the graph and train it
* graph - Build the neural network model
* train - Train the network
# Data
* loader.py - Make batches and prepare labels
* rtl_loader.py - Make batches and prepare labels for RTL data
* data_lib - transform between waveform, spectrogram, and neural network
input representations
* bss_eval - Metric for calculating proxy_SDR
* mir_bss_eval - Metric for calculating SDR
# Model Results & Visualizations
* summaries - image, audio, scalar summary plots
* embedding_summary - Visualizing embeddings in PCA space (in TensorBoard)
# Misc
* kmeans - kmeans implementation
* helper - various useful functions
* utilities - save hparams, track experiment number
# Versions:
Python: 3.6.1
TensorFlow: 1.6.0-dev20180116
CUDA: 9.0, V9.0.176
cudNN: 8.0