This thesis discusses the assessment of speech quality transmitted through telecommunication networks. The aim is to produce a model able to estimate the overall listening quality of speech signals as measured by subjective tests. Objective models for speech quality assessment have been developed for the last twenty years and the most widely adopted is PESO. the currently in-force ITU-T Recommendation P .862. PESQ shows inaccuracy when assessing signals recorded from modern t elecommunication networks that exhibit highly va riable delay such as Voice over lP. This issue is investigated and addressed in this thesis. Objective models for qua lity assessment are generally designed to predict subjective tests, on which they are trained and verified. The behaviour of the model and its accu racy are therefore high ly dependent on the relia bility and the resolution of the subjective t ests. Some aspects of subjective test methodologies are disc ussed in this thes is. The most reliable speech quality. assessment models are perceptual algorithms that compare shortterm representations of the input and the o utput signals of the system under test. This type of model relies on an accurate estimation of the time re lat io nship between the two signa ls. This t hesis shows that the inaccuracy of PESQ for quality assessment of modern telecommunication networks is due to its t ime alignment. Previous time alignment methods for objective models are not suited to frequent delay variation. A new t ime-alignment technique based on correlation of frequency domain representations and sho rtterm delay histograms is presented, allowing robust alignment in t he presence of highly varying delay. A new objective model built from the integration of the proposed time-a lignment with PESQ was verified on a very large number of subjective tests. Results show signifi cant improvements over PESQ in situations presenting frequent delay va riations while keeping the similar level of accuracy in the cases of occasional variations.
Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:601374 |
Date | January 2010 |
Creators | Malfait, Ludovic |
Publisher | University of Essex |
Source Sets | Ethos UK |
Detected Language | English |
Type | Electronic Thesis or Dissertation |
Page generated in 0.002 seconds