Signal representation for speech separation

Thumbnail Image



Huang, Ying

Journal Title

Journal ISSN

Volume Title


University of Guelph


The development of next-generation speech enhancement systems depends on new forms of speech representation that can capture the non-stationarity of speech signal in a way amiable to element-based signal separation processing. Research in this thesis focuses on developing new structure-revealing signal representations that can meet this challenge. The idea is first investigated in developing a Gabor-atom based time-frequency component signal representation. Problems encountered in this stage of investigation lead to the main focal point of the research: the development of a new chirp-decomposition method suitable for speech-like multi-component signal analysis. Research starts from polynomial phase signal (PPS) analysis, followed by studying the complex ambiguity function (CAF) time-frequency distribution for chirp detection. Two new chirp detection algorithms are proposed. One is the principal-chirp Wigner-Hough (PC-WH) transform and another one is the ambiguity auto-term ridge (AAR) detection algorithm. The potential for using the resulting chirp-graph for speech separation is demonstrated in a new speech enhancement system.



speech enhancement systems, speech representation, |speech signal, signal separation processing, signal representation