Towards Time-Aware Music Interpolation in Latent Space

Demo page with audio samples

In this preliminary work, we interpolate between two music samples in the embedding space of Music2Latent. We compare two classical schemes that allow for averaging two feature sequences with arbitrary lengths: symmetric averaging (by Kruskal and Liberman) and DTW Barycenter Averaging. The former method results in stable samples that are coherent with temporal alignment.

Although I failed to visit the conference in person, I attach the paper and the poster, in case anyone is interested. I intend to expand the work in the future, so do not miss any updated in the repository.

How to use

1. Clone the repository
2. Install the required packaged via pip install -r requirements.txt in your virtual or conda environment
3. Follow the notebooks barycenters.ipynb and interpolation.ipynb for the examples of usage