Skip to content


Repository files navigation

Audio inpainting with a context encoder

This project accompanies the research work on audio inpainting of small gaps done at the Acoustics Research Institute in Vienna collaborating with the Swiss Data Science Center. The paper was published at IEEE TASLP available now:


Install the requirements with pip install -r requirements.txt. For windows users, the numpy version should be 1.14.0+mkl (find it here). For the FMA dataset, librosa requires ffmpeg as an mp3 backend.


The paper uses both google's Nsynth dataset and the FMA dataset. In order to recreate the used dataset, execute in the parent folder either python or python The output of the scripts are three tfrecord files for training, validating and testing the model.

The default parameters for the network come pickled in the file magnitude_network_parameters.pkl and complex_network_parameters.pkl. In order to make other architectures use

To train the network, execute in the parent folder python or python This will train the network for 600k steps with a learning rate of 1e-3. You can select on which tfrecords to train the network, the script assumes you have created the nsynth dataset.

Sound examples