Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.513424
Title: System identification with applications in speech enhancement
Author: Lin, Xiang
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2009
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
As the increasing popularity of integrating hands-free telephony on mobile portable devices and the rapid development of voice over internet protocol, identification of acoustic systems has become desirable for compensating distortions introduced to speech signals during transmission, and hence enhancing the speech quality. The objective of this research is to develop system identification algorithms for speech enhancement applications including network echo cancellation and speech dereverberation. A supervised adaptive algorithm for sparse system identification is developed for network echo cancellation. Based on the framework of selective-tap updating scheme on the normalized least mean squares algorithm, the MMax and sparse partial update tap-selection strategies are exploited in the frequency domain to achieve fast convergence performance with low computational complexity. Through demonstrating how the sparseness of the network impulse response varies in the transformed domain, the multidelay filtering structure is incorporated to reduce the algorithmic delay. Blind identification of SIMO acoustic systems for speech dereverberation in the presence of common zeros is then investigated. First, the problem of common zeros is defined and extended to include the presence of near-common zeros. Two clustering algorithms are developed to quantify the number of these zeros so as to facilitate the study of their effect on blind system identification and speech dereverberation. To mitigate such effect, two algorithms are developed where the two-stage algorithm based on channel decomposition identifies common and non-common zeros sequentially; and the forced spectral diversity approach combines spectral shaping filters and channel undermodelling for deriving a modified system that leads to an improved dereverberation performance. Additionally, a solution to the scale factor ambiguity problem in subband-based blind system identification is developed, which motivates further research on subbandbased dereverberation techniques. Comprehensive simulations and discussions demonstrate the effectiveness of the aforementioned algorithms. A discussion on possible directions of prospective research on system identification techniques concludes this thesis.
Supervisor: Naylor, Patrick Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.513424  DOI: Not available
Share: