Improved SEGAN

Tricks to improve SEGAN performance. Eveything is re-implemented into Keras with Tensorflow backend.

Supporting document with evaluation results and other details can be found here.

Deepak Baby, iSEGAN: Improved Speech Enhancement Generative Adversarial Networks, Arxiv preprint, 2020.

Pre-requisites

Install tensorflow and keras
Install tqdm for profiling the training progress
The experiments are conducted on a dataset from Valentini et. al., and are downloaded from here. The following script can be used to download the dataset. Requires sox for converting to 16kHz.
```
$ ./download_dataset.sh
```

Running the model

Prepare data for training and testing the various models. The folder path may be edited if you keep the database in a different folder. This script is to be executed only once and the all the models reads from the same location.
```
python prepare_data.py
```
Running the models. The training and evaluation of the various segan models are implemented in run_isegan.py. which offers several cGAN configurations. Edit the opts variable for choosing the cofiguration. The results will be automatically saved to different folders. The folder name is generated from files_ops.py and the foldername automatically includes different configuration options.

The options are:

Different normalizations
- Instance Normalization
- Batch Normalization
- Batch Renormalization
- Group Normalization
- Spectral Normalization
One Sided Label Smoothing: Encouranging the discriminator to estimate soft probabilities (0.8, 0.9, etc.) on the real samples.
Trainable Auditory filter-bank layer: The first layer is initialized using a gammatone filterbank and use it as a trainable layer.
Pre-emphasis Layer : Incorporating the pre-emphasis operation as a trainable layer.

Evaluation on testset is also done together with training. Set TEST_SEGAN = False for disabling testing.

Misc

This code loads all the data into memory for speeding up training. But if you dont have enough memory, it is possible to read the mini-batches from the disk using HDF5 read. In run_<xxx>.py
```
clean_train_data = np.array(fclean['feat_data'])
noisy_train_data = np.array(fnoisy['feat_data'])
```
change the above lines to
```
clean_train_data = fclean['feat_data']
noisy_train_data = fnoisy['feat_data']
```
But this can lead to a slow-down of about 20 times (on the test machine) as the mini-batches are to be read from the disk over several epochs.

References

[1] S. Pascual, A. Bonafonte, and J. Serra, SEGAN: speech enhancement generative adversarial network, in INTERSPEECH., ISCA, Aug 2017, pp. 3642–3646.

Credits

The keras implementation of cGAN is based on the following repos

saurabh-kataria / isegan