Statistical Signal Segmentation:
Detect the the points in an auditory stream where the statistical properties of the signal change. Using a dynamic Bayesian network model we can fuse information from multiple sources for the purpose of audio event detection. Our model explicitly accounts for: (1) rapid, but non-instantaneous signal change points, (2) not all information sources will be responsive to all events, and (3) delays among change points between information sources. We have also expanded our signal segmentation model to multiple audio channels for use with microphone arrays.
Audio Retrieval:
Facilitate the search of large audio databases using the query-by-example paradigm, i.e., the employed query is a sound file or a sound made by the user during the search. By considering imperfections in the query we can make the model distortion aware, specifically we can make it aware to common distortions in the human voice.
Audio Feature Extraction:
Define an audio feature set for use with natural and environmental sounds, i.e., features that are not specifically tailored for use with speech and music. We use time domain features, short-time Fourier transform (STFT) spectral features, and cepstral features.
Atmospheric Acoustics:
Develop methods to accurately and rapidly predict how the atmosphere influences outdoor sound propagation. Applications for wireless acoustic arrays used for motor vehicle tracking and weather prediction.