Audio based similarity search

We discuss how to use parts of musical audio to search for clips of a similar feel and timbre. The solution includes self supervised learning and a triplet loss function. We provide practical insights from the work.

• Self supervision circumvents the need for huge labeled data sets
• Musical perception is subjective and user testing is necessary 
• Practically, deep learning is error-prone and it pays to have a solid setup

————————————————————————————————————————————————————

Johan Bjurgert – Head of Machine Learning | Epidemic Sound

Johan Bjurgert was born in 1982. He received the Master of Science in Industrial Engineering and Management from Linköping University, Sweden, in 2008, and the Master of Science in Systems Control and Robotics from KTH Royal Institute of Technology, Stockholm, Sweden, in 2015. His research interests lie in identification and control of dynamical systems as well as in machine learning applications. His work experience includes roles in quantitative finance developing algorithms targeting statistical arbitrage as well as roles in data science and machine learning engineering. Previously, he was the head of data science at LeoVegas. Currently, he is the head of machine learning at Epidemic Sound.

Oscar Utterbäck – Machine Learning Engineer | Epidemic Sound

Oscar graduated with an M.Sc. in Engineering Physics from the department of numerical analysis at Lund University, LTH, in 2017, following a machine learning-based thesis project at Expektra AB focusing on feature selection methods for short term load prediction. Since then, he has worked with machine learning-based ventures ranging from computer vision on edge devices to ML based methods for analysis and control of power systems. After testing the waters in academia, he joined Epidemic Sound as a machine learning engineer in 2020, longing for a return to more practical work. Having a life-long music interest, the cross-section of machine learning and music that Epidemic Sound provides was the perfect match. His research interests encompass among other things the futuristic possibilities of manipulating and analyzing audio data with machine learning-based tools.