IEEE Access (Jan 2017)

On Losses, Pauses, Jumps, and the Wideband E-Model

  • Muhammad Adil Raja,
  • Anna Jagodzinska,
  • Vincent Barriac

DOI
https://doi.org/10.1109/ACCESS.2017.2705428
Journal volume & issue
Vol. 5
pp. 16130 – 16148

Abstract

Read online

There is an increasing interest in upgrading the E-Model, a parametric tool for speech quality estimation, to the wideband and super-wideband contexts. The main motivation behind this has been to quantify the quality gain lent by various new codecs and communication situations. There have been numerous such contributions, and all of them have been more or less successful. This paper reports on an extension of the E-Model to the mixed narrowband/wideband (NB/WB) context. More specifically, we take a novel approach toward deriving effective equipment impairment factors (Ie,WB,eff ) by considering additional impairments related to the underlying communications network. These additional impairments are the pause and jump temporal discontinuities along with network-related loss and pure codec-related impairments. While the effect of loss is a thoroughly studied topic and has been integrated into the E-Model, pauses and jumps have been given little attention. Pauses and jumps manifest themselves as temporal dilation and contraction, respectively, in the resulting speech signal that is presented to the listener and are normally caused by jitter and jitter buffer interaction. In this paper, we initially present a fourstate Markov model to characterize, and also emulate, loss, pause, and jump impairments. Then, we present alternative models for computing effective equipment impairment models. A large number of test stimuli were generated using different NB and WB codecs. WB-PESQ was used to evaluate the stimuli. Genetic programming was employed to derive equipment impairment factors. The proposed models have a high correlation with WB-PESQ. We claim that the models proposed by us outperform the existing E-Model by a factor of approximately 29% while using WB-PESQ as a reference model. The models also outperform the E-Model against results from auditory tests. It is also shown that the models outperform the results of multiple linear regressions.

Keywords