Use this URL to cite or link to this record in EThOS:
Title: Blind identification of acoustic systems and enhancement of reverberant speech
Author: Gaubitch, Nikolay Dian
ISNI:       0000 0004 2743 4829
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2007
Availability of Full Text:
Access from EThOS:
Access from Institution:
Speech signals obtained in rooms with microphones positioned at a distance from the talker are degraded in quality due to additive noise and reverberation, where the latter arises from multiple reflections from the surrounding walls and objects. Reverberation causes speech to sound distant and spectrally distorted and can also reduce intelligibility. Therefore, dereverberation is an important speech enhancement process for hands-free terminals. In this thesis, dereverberation techniques are categorized into beamforming, speech enhancement and blind system identification/inversion. Two algorithms, one from each of the latter two categories, are proposed and evaluated. First, it is shown that the autoregressive coefficients of clean speech can be estimated accurately from reverberant multichannel observations and that reverberation mainly affects the predication residual. Consequently, a new method for processing the prediction residual of reverberant speech is derived, combining spatial averaging of the speech signals and temporal larynx cycle averaging. An enhanced speech signal with reduced reverberation is obtained with the processed residual. Second, an adaptive blind SIMO system identification (BSI) algorithm is introduced and an optimal adaptation step-size is derived, which results in faster convergence and increased robustness to noise, Then, an adaptive common roots identification algorithm is developed and employed to demonstrate the degrading effects of common zeros on the BSI algorithm. Adaptive BSI in subbands is proposed for reduced channel length and thus, increased efficiency. Furthermore, exact and approximate acoustic impulse response equalization is discussed, A new subband multichannel least squares inverse filtering method is derived, which significantly reduces the computational complexity and is less sensitive to inaccuracies in the channel estimates compared to its fullband counterpart. Substantial reduction in reverberation can be obtained using this method, even when the impulse response estimates contain errors. Finally, the thesis is concluded with a comparative discussion of the proposed methods and possible directions for prospective research in speech dereverberation.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available