NSF-CAREER: The Listening Machine IIS-0238301 2003–2008

This six-year project started with the idea of applying sound recognition and sepa-ration techniques that had originated in speech recognition to a broader domain of environmental sound mixtures. As it proceeded, the work diversified into several distinct areas, reflecting the different directions of the graduate students primarily supported by the project: Manuel Reyes and Keansub Lee worked on environ-mental sound separation and classification; Graham Poliner and Michael Mandel worked on music audio transcription and classification, and Xanadu Halkias ap-plied many of the same ideas and techniques to the analysis of underwater record-ings of dolphin sounds. 1.1.1 Sound Separation A key problem in analysis and classification of sound mixtures is that the sounds emanating from individual sources cannot, in general, be directly observed or 1 measured, only the combined superposition of all simultaneous sources. To ad-dress this, we looked at a number of techniques for separating or describing indi