Efficient coding of natural sounds [Elektronische Ressource] / Stefan Strahl

Efficient Coding of Natural SoundsVon der Fakult at fur Mathematik und Naturwissenschaftender Carl-von-Ossietzky-Universit at Oldenburgzur Erlangung des Grades und Titels einesDoktors der Naturwissenschaften (Dr. rer. nat.)angenommene DissertationDipl.-Math. M.Sc. Stefan Strahlgeboren am 6. M arz 1975in AugsburgGutachter: Prof. Dr.-Ing. Alfred MertinsZweitgutachter: PD Dr. Volker HohmannTag der Disputation: 3.7.2009AbstractIn this dissertation, methods for an e cient coding of natural sounds are proposed thatare based on the concept of \sparse coding" which has been identi ed as one majormechanism involved in the neurosensory information processing. In the rst chapter itis demonstrated that in the MPEG-2/4 AAC audio coding scheme the modi ed discretecosine transform (MDCT) can be replaced with a shift-invariant sparse signal modelrealized by a matching pursuit algorithm. Thereby an improved perceived audio qualitywas possible, especially at low bitrates. The second chapter addresses the non-trivialproblem to select the optimal degree of overcompleteness of a sparse signal model. Byusing a frame theoretic approach, it is shown that the number M 100 of gammatone lters (2.4 lters per ERB) leads to a near-perfect reconstruction of the signal space ofnatural sounds. In the last two chapters it is demonstrated how a set of signi cancetrees can be used to achieve an e ective encoding of sparse coe cients.