Time and place

The public defense will take place 12:00—16:00 in room 3.107 at CREATE, Aalborg University, Rendsburggade 14, 9000 Aalborg. For more information contact Kristina Wagner Røjen, phone: +45 9940 9926, email: kwro@create.aau.dk.

Assessment committee

• Professor Stefania Serafin (Chairwoman), Aalborg University (DK)
• Professor Gaël Richard, TELECOM Paris (FR)
• Professor Patrick Naylor, Imperial College (UK)

Program for the defense

• 12:00 – 12:15  Welcome by the moderator Dean Henrik Pedersen
• 12:15 – 12:45  Presentation by Professor Mads Græsbøll Christensen
• 12:45 – 13:00  Break
• 13:00 – 16:00  Questions from the assessment committee / ex auditorio
• 16:00 – 18:00  Reception (sign up) 

Abstract

This thesis is concerned with model-based analysis and processing of speech and audio signals, to which a number of scientific contributions are made in the form of new mathematical models and new methods for the processing of such signals. The thesis demonstrates how a number of models can be used for modeling speech and audio signals in different ways and for different purposes. It is shown how the problem of estimating the parameters of these models can be solved in a number of principled ways using methods such as maximum likelihood, subspace methods, and sparse approximations, whereby both very accurate and robust estimators that explicitly take the properties of speech and audio signals and the presence of noise into account are obtained. Among the parameter estimation problems considered are those of fundamental frequency estimation, linear prediction, source localization, and order selection, problems that have many important applications in speech and audio processing, including the analysis, coding, and enhancement of such signals. It is then shown how such models can be integrated in filtering methods to solve both signal enhancement and parameter estimation problems, such as noise statistics estimation and fundamental frequency estimation, and it is shown how these principles can be extended to multiple channels to solve the problems of beamforming and source localization. The results of the thesis as a whole demonstrate the benefits of the model-based approach compared to the typically non-parametric methods used in speech and audio processing, not only in terms of obtaining new and better methods but also advancing our understanding of both speech and audio signals and the associated estimation problems.