Description
Date depot: 12 septembre 2019
Titre: Meaningful audio synthesis and musical interactions through representation learning of sound sample databases
Directeur de thèse:
Carlos AGON (STMS)
Directeur de thèse:
Philippe Joseph Rene ESLING (STMS)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
Since the earliest developments of computers, new musical concepts and practices have been supported by technological and computational advances. Modern music production techniques extensively rely on computer assisted tools and sound sample libraries, both for electroacoustic compositions and for purely electronic music pieces. However principled and scalable techniques to interact with large sound databases are still lacking despite the ever-increasing amount of digital audio contents and instruments available for composers, sound designers and hobbyists. Over the last decades, pattern recognition and computer vision have greatly advanced, becoming a major research topic in artificial intelligence. Novel tasks have been addressed such as automatic object detection and segmentation as well as image generation with meaningful control over visual features (eg. facial expressions) [1]. These features are modeled onto unsupervised representations of lower dimensionality which dimensions can in turn be used as high-level generative variables [2]. The latent representation can support down-stream analysis tasks and sampling to produce new data and variations that are consistent with the training domains. Experiments in music have mainly focused on information retrieval tasks such as classification and recommandation based on learned audio features. Efficient solutions for sound generation are still lacking in comparison with the powerful results achieved for images as diverse as faces, complex visual scenes or paintings. The thesis investigates such models with the goal of learning meaningful sound synthesis and adapted interactions for musical purpose.
Doctorant.e: Bitton Adrien