In recent years we have experienced a big increase in music services - and the number of songs they offer. Music streaming services lure us now with access to millions of songs. But how to find your music in such catalogues? By genre? By mood? By occasion? And how to predict these characteristics from music? How to interpret a performance and how to transcribe it to notes?
In this Meetup we focus on the complex research domain of Music Information Retrieval and how Deep Learning currently transforms this field.
Talk 1:
Deep Learning for Music & Audio Analysis
Thomas Lidy, Musimap and Alexander Schindler, AIT
Thomas and Alexander will show us how music is analysed with Deep Learning, based on audio spectrograms as the input. They will start with well-known Convolutional Neural Networks, pointing out however the major differences to analyzing images with CNNs. They will cover various tasks in the audio domain, such as music/speech separation, instrumental/vocal detection, genre recognition or detecting moods and emotions in music. This talk will also feature more advanced topics, such as Siamese Networks for Music Similarity, and LSTM Recurrent Neural Networks for Beat Detection. Finally they’ll point to their Getting Started Tutorial for Music Analysis, provided freely on Github.
Talk 2:
Drum Transcription via Joint Beat and Drum Modeling using Convolutional Recurrent Neural Networks - Richard Vogl, TU Wien & CP, JKU Linz
When working with Deep Learning, we sometimes face the criticism of using arbitrarily designed black-box systems. However, incorporating domain knowledge into problem modeling and network architecture can improve the performance of trained models for complex tasks. This talk demonstrates this principle in the context of automatic drum transcription, where the goal is to extract a symbolic representation of drum note onsets from an audio signal.
Hot Topics:
This year’s main conference of the International Society of Music Information Retrieval (ISMIR 2018, http://ismir2018.ircam.fr) showed that Deep Learning is also advancing heavily the music analysis domain. As three of our hosts were at that conference, they are going to show a quick recap of novel approaches to various different tasks in the music analysis domain. Jan is going to present (remotely) his novel approach of Zero-Mean Convolutions for Level-Invariant Singing Voice Detection. Alex and Tom will follow with an overview of other “hot” new approaches.
Then, Stephan Wöber from A1 will report from the recent Spark & AI Summit in London.
Finally, Rene Donner, Head of Machine Learning & Engineering at Contextflow will report on other recent advances in Deep Learning.
Join us for networking & discussions in the break and after the talks.
We wish to express special thanks to Anexia IT who is sponsoring not only the venue and the catering of this meetup, but also produced the brand-new official Vienna Deep Learning Meetup T-Shirt! There was a voting to decide between two different designs and we will proudly announce the winning design this evening! Please come to the meetup to get your free T-Shirt (they are limited edition! ;-).
Looking forward to welcoming you,
Tom, Alex, Rene and Jan
