Untitled in english

Author(s)

Roberto Roschel Belli

• Document type: Master's Dissertation
• Press: São Paulo.
• Institution: Universidade de São Paulo (USP). Escola Politécnica (EP/BC)
• Defense date: 2001-02-01

Examining board members

Phillip Mark Seymour Burt; Max Gerken; Joao Marcos Travassos Romano

• Advisor: Phillip Mark Seymour Burt

Abstract

Nowadays, there is a lot of interest in hands-free communication terminals. This kind of terminals are employed, for example, in telepresence systems. One basic problem in these terminals is multichannel acoustic echo cancellation (AEC). In this work an introduction to the problem of multichannel acoustic echo cancellation is initially presented. It is shown how the problem of misalignment appears, as a result, mainly, of the great correlation between the input signals of the adaptive filter responsible for the echo cancellation. In multichannel acoustic echo cancellation systems an attempt is made to reduce misalignment as much as possible, since a large misalignment leads to great variations in the echo reduction. Presently proposed methods for misalignment reduction try to decorrelated the input signals by using non-linearities or adding uncorrelated white noise to the input signals. As a result, this methods introduce a degradation in the audio quality. In the following, results of acoustic measurements to be used in simulations are presented. Measurements consisted mainly of impulse responses of rooms in which a telepresence system could be used. Efficient Matlab programs were developed with the purpose of making simulations and other calculations concerning multichannel AEC. The FTF and NLMS adaptive algorithms were implemented. The effect of changes in the position of the speaker was verified by simulations using the implemented programs. It was possible to verify in practice the misalignment problem in multichannel AEC and that for the monochannel case misalignment is of little importance. It was also verified that sucessive changes in the position of the speaker leads to misalignment reduction. A theoretical analysis of this behaviour is presented for algorithms of the NLMS and RLS families. In order to reduce misalignment, a method that involves changing the cost function with the introduction of a term proportional to the norm of the adaptive filter coefficients was investigated. To implement this change, noise was added to the input signals of the algorithm. Three alternatives were considered for noise addition, with different effects on audio quality, ERLE reduction after changes in the position of the speaker and computational cost. The leaky-FTF algorithm was also investigated. It is a stabilized version of the FTF algorithm, with a leakage factor that changes the cost function. The change in the cost function helps reduce misalignment. Another method for misalignment reduction is proposed in this work: microphone switching, which is equivalent to periodic changes in the position of the speaker. The result obtained with this method is better or at least equivalent to the results obtained in the literature, with the advantage that no distortion is introduced in the speech signals. (AU)