Contribution
Multilingual Listening Effort Prediction - LEAP goes international
* Presenting author
Abstract:
The non-intrusive (i.e., reference-free) model for Listening Effort prediction from Acoustic Parameters, LEAP, has been proven to correlate well with subjective ratings of perceived listening effort in a number of studies. It is increasingly in demand from international interested parties. To serve these parties, the model has to be adapted to language-specific versions, since its core consists of a phoneme classifier trained on German data. Hence, languages other than German lead to somewhat different model output values compared to German speech. An earlier study with English data showed that re-mapping of LEAP’s direct output measure M to the subjective listening effort scale LE achieved similar prediction accuracy for English data as for German data.This contribution presents the test of two alternative approaches to estimate the language-specific mapping function f: M -> LE from speech samples without having subjective ratings. The two estimates of f are compared with a third mapping function derived from new experimental data from a listening test with Turkish speech rated by native Turkish speakers. The suitability of the tested estimation approaches as a substitute for formal listening tests will be determined by a comparison of the LE prediction accuracies achieved with the respective mapping functions.