Working with audio sounds easier than it is: a deep learning perspective

Convolutional neural networks emerged as a natural model for solving computer vision problems and are now accountable for many state-of-the-art results. Perhaps surprisingly, they have also been shown to be very effective in music applications. This talk describes some of the work we have done in this domain at Peltarion during the past year in collaboration with our partner Epidemic Sound. We also dive deeper into the representations commonly used for audio, both waveform and spectral. We compare audio with natural images, describing desirable properties, such as translational equivariance, that no longer persist in audio representations. Along the way, we will look at examples of how researchers have adapted models and representations to introduce similar properties in other applications. Music information retrieval is a large domain very well suited for deep learning and I hope to inspire more work in it from both academia and industry.