The thesis deals with various aspects of Automatic Musical Instrument Recognition (AMIR). AMIR means, intuitively speaking, that given a musical recording, the computer attempts to identify which parts of the music are performed by which musical instruments. AMIR research has gained popularity over the last 10 years especially due to its applicability as a component inside "Intelligent" music search-engines, which can allow searching the Internet or mass-storage devices in personal "MP3" players for music using "intelligent" criteria such as musical style or composition - as opposed to searches involving only textual information provided with the musical files. Other usages of AMIR include integration and improvement of other Musical Information Retrieval tasks such as Automatic Transcription and Score Alignment, and as a tool in applications for composers and recording studios. AMIR is a compound process involving many challenging stages. The various stages of the AMIR process as presented in this thesis include obtaining and formatting of Learning and Test sound databases, computing feature descriptors on the sounds, automatic purging of the databases, feature weighting and dimension reduction of the feature descriptor space and finally, classification of the sounds as belonging to different instruments. Performing informative evaluation of the AMIR process is also important and non-trivial. This work deals in detail with the different stages of the AMIR process and while "filling holes" in the theory it introduces new techniques and methods for performing many of the tasks, accomplishing AMIR of separate tones, Solo performances and polyphonic, multi-instrumental music.
