Binaural Speech Intelligibility Prediction and Nonlinear Hearing Devices

Title: Binaural Speech Intelligibility Prediction and Nonlinear Hearing Devices
Authors: Ellaham, Nicolas
Date: 2014
Abstract: A new objective measurement system to predict speech intelligibility in binaural listening conditions is proposed for use with nonlinear hearing devices. Digital processing inside such devices often involves nonlinear operations such as clipping, compression, and noise reduction algorithms. Standard objective measures such as the Articulation Indeix (AI), the Speech Intelligibility Index (SII) and the Speech Transmission Index (STI) have been developed for monaural listening. Binaural extensions of these measures have been proposed in the literature, essentially consisting of a binaural pre-processing stage followed by monaural intelligibility prediction using the better ear or the binaurally enhanced signal. In this work, a three-stage extension of the binaural SII approach is proposed that deals with nonlinear acoustic input signals. The reference-based model operates as follows: (1) a stage to deal with nonlinear processing based on a signal-separation model to recover estimates of speech, noise and distortion signals at the output of hearing devices; (2) a binaural processing stage using the Equalization-Cancellation (EC) model; and (3) a stage for intelligibility prediction using the SII or the short-time Extended SII (ESII). Multiple versions of the model have been developed and tested for use with hearing devices. A software simulator is used to perform hearing-device processing under various binaural listening conditions. Details of the modeling procedure are discussed along with an experimental framework for collecting subjective intelligibility data. In the absence of hearing-device processing, the model successfully predicts speech intelligibility in all spatial configurations considered. Varying levels of success were obtained using two simple distortion modeling approaches with different distortion mechanisms. Future refinements to the model are proposed based on the results discussed in this work.
CollectionThèses, 2011 - // Theses, 2011 -