An Objective Measure of Quality for Time-Scale Modification of Audio

Deep Learning-Based Single-Ended Objective Quality Measures for Time-Scale Modified Audio

“…As the prediction of opinion scores for novel TSM methods is the use case, early stopping based on validation loss was not used. The optimal epoch minimised the distance measure of Roberts and Paliwal (2020a), where the minimum overall distance (D), is calculated by where ρ and L are calculated by…”

“…An evaluation set of 6000 files, published as part of Roberts and Paliwal (2020a), was generated from the reference files in the test set. 20 new time-scales in the range of 0.22 < β < 2.2, with all TSM methods listed in Section I used to process the reference files.…”

“…9 and 11 show TSM methods for which differences in mean are statistically significant. As in Roberts and Paliwal (2020a), all results for β = 1 and β < 0.25 were excluded from averaging calculations forming Tables II and III, as time-scaling is applied at β = 1, and the minimum β available for EL is 0.25. Common trends are presented, followed by CNN and then BGRU-FT analysis.…”

“…The total published research is found in the following papers. Fierro and Välimäki (2020) published preliminary work towards an objective measure, initial measures were published by Roberts and Paliwal (2020b) and formalised in Roberts and Paliwal (2020a), which extended PEAQ with additional features and explored synchronization of reference and time-scale processed signals.…”

“…The DE method of Roberts and Paliwal (2020a), referred to as OMOQDE from this point, considered six methods of signal alignment before calculation of PEAQ features -in addition to hand-crafted features specific to the artefacts of time-scaled signals. Formulated as a regression problem, an FCNN was used to predict the MOS targets of the TSMDB.…”