Estimating music descriptions using convolutional neural networks

12/19/2019
Speaker(s) : Alice Cohen (Sciences et technologies de la musique et du son, Sorbonne Université)
In Music Information Retrieval (MIR) and voice processing, the use of machine learning tools has become in the last few years more and more standard. Especially, many state-of-the-art systems now rely on the use of Neural Network. More precisely, we will focus on convolutional neural networks (ConvNet), a class of neural networks designed for image analysis.
To apply such networks to sound and music, several problems can arise. We chose to study 3 of them: - What is the impact of input representation? Which can of transform can we use to inform the resolution of MIR problems? To answer those questions, we will present works don on structure estimation and singing voice separation. - How can we gather large amount of labeled data ? We will present two different strategies: one for singing voice detection using teacher student paradigm and one for singing voice separation, using signal processing tools for data augmentation. - How can we use ConvNet not only to solve problems, but also to validate solutions? To study this problem, we will present how we design and evaluate a voice anonymization method in urban sound recordings.
**Alice Cohen's bio**
Alice Cohen was a PhD student at Institut de Recherche et de Coordination Acoustique Musique (IRCAM). She defended her thesis, untitled "Estimating music and sound descriptions using deep learning" last October. Her research interests mostly focus on voice extraction and detection, with signal processing and machine learning tools.
_Plus d'information sur Alice Cohen :_