Computational Auditory Scene Analysys and Sound Separation


Multi-channel based Sound Separation

Time-Frequency Representation of Sound

Time-frequency representation of a sound obtained as the output of a set of band-pass filters. To match the sensitive level of human, the level of each frequency band is adjusted by the MAF (minimum audible field) curve of human.


Local Peak Map


Echo-Free Onset Map based on the Model of the Precedence Effect

A map of estimated sound to echo ratio in time-frequency domain.

Detected echo-free onsets by the model of the precedence effect.


Histogram of Onset ATDs

The Arrival temporal disparity (ATD) of each sound component (between different channels) is calculated at the onset of component. A histogram of onset ATD is calculated in which each significant peak probably corresponds to an existing sound source. Sound components are then separated according to their assigned groups.


Grouping Components by Onset ATDs

Sound component group 1 (/tyusin/).

Sound component group 2 (/anoseikatsuo/).


Results of Separation

The following figures show the waveforms of sound before and after the separation.

(Experiment in an Anechoic Chamber)

(Experiment in an empty room enclosed with concrete walls)


[Return to home page]
last modified: 19 August 2000 by Jie Huang (