Towards a speaker localization from spontaneous speech: north-south classification for speakers of contemporary German

Authors: Thomas Kisler, Florian Schiel

Abstract:

Geographical regression analysis based on phonetic features aims to locate the origin of a speaker by relating phonetic features derived from a small speech sample to longitude/latitude coordinates. In this paper we present results from a preliminary experiment in which using Random Forests, we classify speakers into two geographical regions North/South based on openSMILE [1] features derived from the German “German Today” [2] corpus. The aims are to test the feasibility of a data-driven approach to geolocalization with Random Forests, to evaluate the German phone classes that carry geoinformation, and to evaluate which phonetic features contribute to such a binary classification of speakers from the North or South. Results show that based on the voiced fricative /z/ alone it is possible to correctly classify 81.72% of the speakers, which confirms the often reported North/South voicing contrast in Germany of /z/ in most positions [3]. We identify a number of features that do not contribute to the geolocalization and will therefore be discarded in future experiments.