M.Sc Thesis

M.Sc StudentBerkun Reuven
SubjectQuality Analysis and Enhancement of Reverberated Speech
Using Microphone Arrays
DepartmentDepartment of Electrical and Computer Engineering
Supervisor PROF. Israel Cohen
Full Thesis textFull thesis text - English Version


Speech signals in enclosed environments are often distorted by reverberation and noise. Reflections from the walls and other obstacles are received in the microphone as a superposition of delayed and uncorrelated versions of the original signal, creating the reverberation effect. These late reflections smear and damage the quality and the intelligibility of the obtained signal. In this dissertation, we introduce novel microphone-array based techniques to handle various problems that exist in reverberant speech signals.

In the first part of this study, we deal with the quality evaluation of reverberated speech signals. In a distant-speech communication system with several distributed microphones, it is a nontrivial challenge to monitor the perceived quality of each microphone and select the channel with the best reception. Most of the existing methods for quality estimation usually require prior information or a clean reference signal, which is seldom available. In this dissertation, a practical non-intrusive measure for quality assessment of reverberated speech is proposed. Based on a statistical model of the reverberation process, we examine the energies as measured by unidirectional elements in a microphone array.
By measuring the power ratio we obtain a reliable measure for the reverberation amount in the acoustic signals. This measure is then utilized to derive a blind estimation of the direct-to-reverberation energy ratio (DRR). The proposed approach attains a simple and robust quality measure, shown here through persuasive simulation results.

In the second part of the research we deal with the enhancement of reverberated speech by using beamforming algorithms in microphone arrays. Conventional superdirective beamforming is a well-known multi-microphone enhancement method with superior directivity factor. Nevertheless, it is extremely sensitive to uncorrelated noise, thermal noise, and slight errors in the array elements, resulting in an inferior white noise gain level. The delay-and-sum beamformer, on the other hand, manages to maximize the white noise gain, but suffers from a very low directivity factor.
In light of that, we would like to find an optimal array beamformer, that would have both high gain for reverberated signals, together with fine control on its white noise gain. In this part, we discuss the design of different beamformers which control both directivity factor and white noise gain. First, we combine modified versions of known conventional beamformers to create a robust regularized superdirective beamformer.
We then extract analytic closed-form expressions of the beamformer gain responses, and extend them to derive beamformers with full control of the desired gain and performance measures. Second, we introduce a different type of beamformer with tunable superdirective gain.
This approach plays a role of a regularized superdirective beamformer, where instead of solving a constrained optimization problem, we minimize both white noise and diffuse noise energy in the proposed solution. By using a tunable regularization parameter, we enable direct control of the tradeoff between the white noise gain and the directivity factor, together with an effective user-controlled performance.

All of the proposed approaches exhibit simple yet impressive beamformers, with fully-controllable characteristics. The performed simulations show highly-robust solutions with-robust solutions with very encouraging results.