Audio segmentation, classification and visualization

Zhang, Xin
Whalley, Jacqueline
Brooks, Stephen
Macdonell, Stephen
Item type
Degree name
Doctor of Philosophy
Journal Title
Journal ISSN
Volume Title
Auckland University of Technology

This thesis presents a new approach to the visualization of audio files that simultaneously illustrates general audio properties and the component sounds that comprise a given input file. New audio segmentation and classification methods are reported that outperform existing methods. In order to visualize audio files, the audio is segmented (separated into component sounds) and then classified in order to select matching archetypal images or video that represent each audio segment and are used as templates for the visualization. Each segment's template image or video is then subjected to image processing filters that are driven by audio features. One visualization method reported represents heterogeneous audio files as a seamless image mosaic along a time axis where each component image in the mosaic maps directly to a discovered component sound. The second visualization method, video texture mosaics, builds on the ideas developed in time mosaics. A novel adaptive video texture generation method was created by using acoustic similarity detection to produce a resultant video texture that more accurately represents an audio file. Compared with existing visualization methods such as oscilloscopes and spectrograms, both approaches yield more accessible illustrations of audio files and are more suitable for casual and non expert users.

Audio , Segmentation , Classification , Visualization , Time mosaics , Video textures
Publisher's version
Rights statement