Using an Onset-based Representation for Sound Segmentation

Citation

Smith L (1995) Using an Onset-based Representation for Sound Segmentation. NEURAP'95 - 8th International Conference on Neural Networks and their Applications, Marseille, France, 13.12.1995-15.12.1995.

Abstract
We present a technique for using pre-processing based on mammalian early auditory processing to produce a segmentation of sound based on onsets and offsets. The sound signal is bandpassed and each band processed to enhance onsets and offsets. The onset and offset signals are compressed, then clustered both in time and across frequency channels using a network of integrate-and-fire neurons. A spike-based representation of onsets and offsets is produced, and the timing of these spikes used to segment the sound. By considering spikes in varying number of bands, a multi-level segmentation tree can be built. This tree is a purely data-driven representation of the segmental structure of the sound.