Subjective measurement and instrumental estimation of conversational speech quality based on perceptual dimensions

点击翻译按钮获取中文摘要

This research project aims at the identification, subjective measurement and instrumental prediction of quality dimensions associated with conversational transmitted speech. The idea is to describe conversational speech quality as a perceptual event in a multidimensional space, and to define a valid and reliable speech quality estimation model on this basis. Such a model estimates not only overall conversational speech quality, but also diagnoses causes and related perceptual effects of quality degradations.Following prior work accomplished for the listening-only phase, the first step is to identify perceptual dimensions in the talking-only and in the interactive (i.e. exchange of listening and talking) phase of the conversation. This is achieved via multidimensional psychoacoustic analyses, using an online simulation system which allows typical speech quality degradations of all conversational phases to be experienced by human test participants. In order to quantify the identified dimensions, a new subjective test method will be established which consists of a structured conversation test and efficiently measures all dimensions in all three phases of the conversation. This test method is used to collect databases of speech transmission settings and corresponding subjective ratings. The databases form the basis for deriving quality estimations of individual perceptual dimensions and - in a second step - of overall conversational speech quality. The derived model is finally validated against standard not-dimension-based models.