Inhalt des Dokuments
Analysing Outliers of a Perceptual Speech Quality Estimator
LOCATION: TEL, Auditorium 3 (20th floor), Ernst-Reuter-Platz 7, 10587 Berlin
Date/Time: 14.05.2018, 15:00-15:45
SPEAKER: Tom Lord (TU Berlin)
In recent decades, innovations in speech communication technology have ensured improved bandwidths and greater flexibility within a telecommunication context. This has come at the expense of quality, with a wide range of new impairments impacting on the perceived quality of transmitted speech signals. Network providers motivated to ensure a high Quality of Experience (QoE) for their customers have become increasingly concerned with measuring speech quality instrumentally. This has fuelled the development of perceptual speech quality estimators, which attempt to predict the overall listening quality of a degraded speech signal as perceived by the end user. P.AMD, a work item of ITU-T SG12, offers a new approach to speech quality assessment by predicting four separate quality scores for the perceptual dimensions of noise, discontinuity, coloration and loudness. This thesis analyses the distortion types resulting in high prediction errors within P.AMD and investigates a correlation between noise activity during non-speech signal sections and a high noise prediction error.