We introduce GACELA, a generative adversarial network (GAN) designed to restore missing musical audio data with a duration ranging between hundreds of milliseconds to a few seconds, i.e., to perform long-gap audio inpainting. While previous work either addressed shorter gaps or relied on exemplars by copying available information from other signal parts, GACELA addresses the inpainting of long gaps in two aspects. First, it considers various time scales of audio information by relying on five parallel discriminators with increasing resolution of receptive fields. Second, it is conditioned not only on the available information surrounding the gap, i.e., the context, but also on the latent variable of the conditional GAN. This addresses the inherent multi-modality of audio inpainting at such long gaps and provides the option of user-defined inpainting. GACELA was tested in listening tests on music signals of varying complexity and gap durations ranging from 375ms to 1500ms. While our subjects were often able to detect the inpaintings, the severity of the artifacts decreased from unacceptable to mildly disturbing. GACELA represents a framework capable to integrate future improvements such as processing of more auditory-related features or more explicit musical features.
Install the requirements with pip install -r requirements.txt. Since ltfatpy is not available on windows, the code can only be used with Linux or Mac.
The datasets used for the experiments are available:
| Dataset | Type | Details |
|---|---|---|
| Lakh | Midi | Used LMD-matched |
| Maestro | Midi & piano | Use full dataset |
| Free Music Archive | General music | Used only rock song fom fma-small |
On the folder 'train paper networks' you can find the python scripts we used to train the networks in the paper. To retrain those, just change the dataset folder to where your data is stored.
To train new networks, just run python train.py --experiment_name awesome_name --data_folder /path/to/your/mp3/or/wav/data/. We trained GACELA for 7 days on a NVIDIA Titan pascal X, and the trained models occupy 250Mb. At inference time, GACELA's generator produces a batch of 64 gaps in 14ms.
To test GACELA, either use the Test GACELA notebook in the main folder, or use one of the examples provided in the notebooks folder.
-
The pre-print can be read on arxiv.
-
The paper was published by IEEE Journal of Selected Topics in Signal Processing.
-
To hear examples please go to the accompanying website.
-
The checkpoints used for the evaluation of the paper can be downloaded here. Please extract the archives in the folder
saved_results. Alternatively, you can simply run the scriptdownload_checkpoints.py:
python download_checkpoints.py
This project accompanies the research work on audio inpainting of large gaps done at the Acoustics Research Institute in Vienna collaborating with the Swiss Data Science Center. The paper was published by IEEE Journal of Selected Topics in Signal Processing.
We specially thank Michael Mihocic for running the experiments at the Acoustics Research Institute's laboratory during the coronavirus pandemic as well as the subjects for their participation.