A new model of speech quality under delay is presented that includes conversational interactivity. It is based on two previously reported narrowband telephony conversation tests involving different delays, with subject-pairs judging overall quality after each conversation. The tests were conducted with different conversation scenarios targeting different levels of interactivity. The instructions given prior to the tests were varied in their emphasis on speed of task completion. Based on the test results, the paper proposes an extension of a widely used conversational speech quality model, the so-called E-model (ITUT Rec. G.107), to cover the joint effect of interactivity and delay. To this aim, two new parameters are introduced, one of which represents the minimum perceivable delay, and the other expresses in how far users will attribute the delay-effect to the conversational quality of the line. Based on the analysis of the recorded test conversations in terms of its surface structure (turns, speaker activities, etc.), prominent differences and delay-dependencies of a number of conversation parameters were found that characterize the impact of delay on the conversational flow and on perceived quality.

In this paper we demonstrate how speech signal features can be used to detect and explain differences in human to human conversation tests. To this end, we compare the results of two conversational VoIP quality experiments designed to quantify the impact of network delay on perceived speech quality. Both studies followed the same procedures and used the same scenarios, but were conducted in two different labs. Our comparison shows that the two studies, despite having been executed correctly using the same test design, still can produce surprisingly different results regarding the users quality perception on a MOS scale. In this respect, speech signal features extracted from conversation recordings help identifying divergent participant behavior as plausible cause for such differences. Our in-depth analysis reveals how novel parameters developed by us like Intended and Unintended Interruption Rate (IIR, UIR) and the corrected Speaker Alternation Rate SARcorrcan be used to successfully determine the extent to which the results of different conversational speech quality studies are directly comparable and thus eligible for pooling, or not.

In this paper, we present ongoing experiments and insights regarding automatic assessment of perceived personality. While within the INTERSPEECH Speaker Trait Challenge participants will train systems in order to recognize binary targets along the Big 5 personality trait, we will analyze and discuss properties of the data, the labeling scheme and the predictive quality. Conducting factor analyses, estimating reliability, and building regression models capturing dimensions of personality we compare all results to our former and current work and introduce a new extension of our personality database. Eventually, this paper contributes in methodology and understanding on how to asses the perceived personality from an unknown speaker by humans and machines.