טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentEliav Benisty
SubjectDereverberation and Noise Suppression of Multichannel
Speech Signals
DepartmentDepartment of Electrical Engineering
Supervisor Full Professor Cohen Israel
Full Thesis textFull thesis text - English Version


Abstract

Speech signals recorded in a real environment are often contaminated by nonstationary interfering signals (other speakers), as well as by stationary noise. Moreover, distortion imposed by the room impulse response (RIR) of the acoustic environment may severely degrade the fidelity and intelligibility of the speech signal. Extraction of desired speech signals recorded in such environments is required in many practical applications: Hands-free communication systems, cellular phones, teleconferencing, hearing aids, and man-machine interfaces among others.

In this thesis, beamforming techniques are utilized to address the problem of joint

dereverberation and noise reduction in the short-time Fourier transform (STFT) domain.

The term beamforming refers to the design of a spatio-temporal filter used to process signals received by a microphone array in order to extract a desired sound. Obviously such a spatial filter can be designed to optimize different cost functions, depending on the application's demands and constraints. The main goal of this study is to develop and analyze beamformers specifically designed to extract clean, early reflections from noisy, reverberating observations.

The study is divided into two parts. In the first part we show that beamformers originally developed for the purpose of noise reduction can be adjusted to resolve the problem of joint dereverberation and noise reduction. The analysis is based on the estimation of the full to early relative transfer function (RTF), defined as the ratio between the acoustical transfer function (ATF) of the early speech reflections, and the ATFs of the array's microphones. Using this estimate we express the beamformer observations vector as a function of the desired signal, from this perspective performance measures are defined and existing beamformers are deduced. A comprehensive experimental study, consisting of simulated environments proves the applicability of the proposed algorithm to the task of joint dereverberation and noise reduction (JDNR). Furthermore, the proposed algorithm outperforms the transfer function generalized sidelobe canceler (TF-GSC) algorithm in extracting the clean, early reflection.

In the second part we introduce a different outlook on the design of beamformers in the STFT domain. Specifically we use the full to early RTF, to divide the beamformer observation vector into three terms: the early reflection vector, the unwanted coherent interference vector (which includes the late reverberations), and the uncorrelated noise vector. This decomposition introduces new optimization problems, followed by beamformers with potentially high reverberation suppression capabilities. An experimental study consisting of simulated reverberant environments demonstrates the dereverberation and denoising capabilities of the suggested beamformers, compared to the minimum variance distortionless response (MVDR) beamformer.