These proceedings contains refereed papers presented at the Fifteenth IEEE Workshop on Machine Learning for Signal Processing (MLSP’2005), held in Mystic, Connecticut, USA, September 28-30, 2005. This is a continuation of the IEEE Workshops on Neural Networks for Signal Processing (NNSP) organized by the NNSP Technical Committee of the IEEE Signal Processing Society. The name of the Technical Committee, hence of the Workshop, was changed to Machine Learning for Signal Processing in September 2003 to better reflect the areas represented by the Technical Committee. The conference is organized by the Machine Learning for Signal Processing Technical Committee with sponsorship of the IEEE Signal Processing Society. Following the practice started two years ago, the bound volume of the proceedings is going to be published by IEEE following the Workshop, and we are pleased to offer to conference attendees the proceeding in a CDROM electronic format, which maintains the same standard as the printed version and facilitates the reading and searching of the papers The field of machine learning has matured considerably in both methodology and real-world application domains and has become particularly important for solution of problems in signal processing. As reflected in this collection, machine learning for signal processing combines many ideas from adaptive signal/image processing, learning theory and models, and statistics in order to solve complex real-world signal processing applications. High quality across such topical diversity can only be maintained through a rigorous and selective review process. This year, 119 full papers (6 pages) were submitted, out of which 65 (resulting in an acceptance rate of 55%) were selected for oral or poster presentation, after reviews by three referees for each. We would like to thank the MLSP’2005 Technical Committee for taking the time to provide quality reviews. This year, the workshop featured research work in the areas of nonlinear signal processing, system identification, blind source separation, learning theory and models, neural networks, applications in image and video processing and speech processing, as well as implementation and other applications of machine learning. Two special sessions on bioinformatics and biomedical imaging and data fusion were included in the program as well as a tutorial on engineering aspects of fixed point theory. This was also the first year for a data competition which was chaired by Deniz Erdogmus. Our warmest, special thanks go to our plenary speakers: Prof. Andrew Barron of Yale University (USA), Prof. Barry Horwitz of the National Institutes of Health (USA) and Prof. Simon Haykin of McMaster University (Canada). Continuing the tradition of paperless and easy communication, many of the details of the MLSP’2005 Workshop were handled electronically through the workshop webpage (http://mlsp2005.conwiz.dk), which, among other features, included web-based submissions, review, and registration.