Equalization by sub-band mint method for speech dereverberation

• Main Author: Xu, Guangyu.
• Other Authors: School of Electrical and Electronic Engineering
• Format: Theses and Dissertations
• Language: English
• Published: 2012
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/50955
• Institution: Nanyang Technological University

Equalization of room transfer functions (RTFs) is a crucial topic on inversion research topic with several applications in audio signal processing. For single-channel methods, perfect signal recovery is a challenging task so far due to approximate equalization. For multipleinput/ output inverse theorem (MINT), the non-minimum phase problem is eliminated and exact equalization can be achieved. However, MINT for RTFs equalization has high computational complexity and is very sensitive to the inexact estimation of room impulse responses. In most real scenarios, the length of RTFs are significantly and RTFs are unknown. The existing blind systerm identification techniques estimate RTFs inevitably up to a limited accuracy with some estimation errors, prior to inversion stage. This limits the practical applications of MINT algorithms. In this thesis, we propose a sub-band MINT equalization method with inexact RTFs estimates. In this sub-band MINT algorithm, the inverse filter is designed by using generalized discrete Fourier transformed (GDFT) filter-banks, where the full-band time-aligned RTFs are decomposed into equivalent sub-band filters prior to equalization. Simulations results show that sub-band MINT method is not only much less computationally complex, but also more robust to inaccurate RTFs estimation compared with full-band MINT method.