This is Tensorflow 2.0 verison of Temporal FiLM for Speech Super Resolution.
Note that this is unofficial implementation.
https://github.com/kuleshov/audio-super-res
https://arxiv.org/abs/1909.06628
tensorflow 2.0+
pip install numpy h5py tqdm scipy librosa soundfile
Download the preprocessed VCTK Single Speaker dataset from the following link and run h5_data.py to generate dataset in 'h5' format.
https://pan.baidu.com/s/1Q8uPLtaJXZ9Odx17Itawtg extraction code: 3tl6
https://drive.google.com/file/d/123NN9H1tx2lNnwl3eikvn0Ay-a52untB/view?usp=sharing
Before running the code, please set paths to the dataset.
# the folder of HR audios and LR audios
in_dir_hr_train = "/path_to/train_hr/"
in_dir_lr_train = "/path_to/train_lr/"
in_dir_hr_test = "/path_to/test_hr/"
in_dir_lr_test = "/path_to/test_lr/"
# the path of output .h5 file
out_dir_train = "./train.h5"
out_dir_test = "./test.h5"Run train.py for training.
Run test.py for evaluation.
(Single Speaker ratio=4)
| SNR (dB) | LSD (dB) | |
|---|---|---|
| paper | 16.8 | 3.5 |
| my results | 17.37 | 3.425 |