PhD Qualifying Examination


"Feature Extraction in Content-Based Musical Instrument Recognition :
A Survey"

By Miss Yongzhen Zhuang


Abstract:

Content-based instrument recognition is one of the fundamental components
in music recognition. It is very useful in many applications such as music
retrieval on the internet, automatic music transcription, music indexing
and annotation, and video segmentation and retrieval. The development of a
robust, large-scale instrument recognition system relies on both reliable
feature extraction and accurate classification. Much work has been done to
obtain an optimal classifier. However, there has been little improvement
in feature extraction methods in recent years probably due to the relative
success of perceptual features such as Mel-Frequency Cepstral
Coefficients. This survey reviews existing content-based automatic
instrument recognition systems, both the monophonic and polyphonic
systems. We focus on feature extraction methods, describing a wide range
of features, including temporal, spectral, ceptral, LPC, and wavelet
features. Since human hearing system is the idea recognition system, it is
worthwhile to investigate human perception and find perceptually salient
features. We also review previous studies about human recognition ability
and discuss a method to evaluate perceptual salience of a feature through
human discrimination experiments. We present an approach to calculate the
correspondences between features and human discrimination in a set of
interpolated sounds, and expect to find a robust feature evaluation method
in the high dimensional timbre space.


Date:     		Tuesday, 23 December 2003

Time:                   10:00a.m.-12:00noon

Venue:                  Room 2302
			lifts 17-18

Committee Members:      Dr. Andrew Horner (Supervisor)
                        Dr. Brian Mak (Chairperson)
			Dr. Lydia Ayers
			Dr. David Rossiter


**** ALL are Welcome ****