This Small Grant for Exploratory Research is investigating a framework of learning and unlearning algorithms that can be used for identifying noise robust auditory features. Even though most speech based recognition systems deliver robust performance under controlled laboratory conditions, their performance degrades significantly in presence of noise primarily due to mismatch between training and deployment conditions. The proposed exploratory study investigates the possibility of using information embedded in higher-order spectral and temporal manifolds which could remain intact even in the presence of ambient noise. Estimation of these non-linear manifolds in presence of noise, however, poses a significant challenge and is the focus of this study. We are investigating proof-of-concept features based on cooperative learning-unlearning (CLU) algorithms that estimates manifold parameters in a reproducing kernel Hilbert space (RKHS) spanned by speech signals. We are evaluating the robustness of these features in presence of room acoustics and background noise. The broader impact of this exploratory study will be development of enabling technology that can be used in the area of voice based biometrics, where seamless authentication can be performed over the internet, cell phones or other voice based media. The educational impact of the study includes graduate student training and development of public domain software tools for CLU algorithms which will be available to the scientific community.
