ABSTRACT
Millions of years of genetic evolution have shaped our auditory system, raising our way of listening to an art form. Despite a somewhat limited frequency spectrum, we are able to perform excellent and flexible discrimination of events. These unique capacities originate from our multidimensional perception of sounds and their temporal structures. By gaining insights from these mechanisms to drive our choices of algorithms, we first studied the problem of generating orchestral sound mixtures that would closely approximate any given audio signal. While performing this reconstruction, we avoid mixing the similarity into a single measure, but rather use multi-objective optimization concepts. This allows us to obtain a set of efficient solutions that provide various compromises among spectral objectives. We show that trying to solve these artistic problems generates a flexible informatics framework called HyperVolume-MultiObjective Time Series (HV-MOTS) matching, which has applications in several scientific fields. We will show that gaining this multi-objective and temporal flexibility produces a classification system that outperforms state-of-the-art methods on a wide range of scientific problems such as meteorological prediction, handwriting analysis, generic audio retrieval, and so forth. Finally, we will present our latest finding in which this method allows us to construct a biometric system for heartbeat identification, i.e., we show that we can accurately and uniquely identify someone through the sounds their heart produces.
ABOUT PHILIPPE ESLING
Philippe Esling has an M.Sc in acoustics, informatics and mathematics and is currently a third-year PhD student in computer music and signal processing on the topic of time series analysis at IRCAM in Paris. He is also a graduate from the National Conservatory of France after 12 years of studying piano. During his PhD, he worked 6 months at the Japanese-French Laboratory of Informatics at Tokyo University. In 2010, he won the Young Researcher prize from the Assocation Française d'Informatique Musicale. In 2011, he was selected for the Tokyo Wonder Site festival where he presented an experimental musical piece. He has authored papers in ACM Computing Surveys, PNAS and IEEE.