The majority of pronunciation tutoring systems is addressing foreign learners of a language, while some applications, such as the rhetoric or articulation training of actors and managers, are also dedicated to native speakers. In particular, accent-reduction tutoring requires reliable methods of Automatic Dialect Classification (ADC) on both, learners' or reference speech. Beyond that, ADC can support further applications of language and speech technology, e.g. the localization of call-center talks or a forensic analysis. Our contribution describes ADC experiments on different corpora of read and spontaneous German speech, which are not based on prior corpus transcriptions. We started with selected feature combinations and classification methods, which have been already studied on English, Mandarin or Arabic. Based on this, we implemented, trained and evaluated classifiers on German dialect varieties in the corpora "Regional Variants of German 1" (RVG) and "Deutsch Heute" (DH). Our test design is focused on differences among the read and spontaneous speech data. The evaluation indicates, that a three or nine-class dialect discrimination and classification on read and spontaneous speech are utilizing the same basic principles, although the overall results, purely using spectral or correlating features, are less sophisticated and call for further clarifications.
Cite as: Dobbriner, J., Jokisch, O. (2019) Implementing and Evaluating Methods of Dialect Classification on Read and Spontaneous German Speech. Proc. 8th ISCA Workshop on Speech and Language Technology in Education (SLaTE 2019), 53-58, doi: 10.21437/SLaTE.2019-10
@inproceedings{dobbriner19_slate,
author={Johanna Dobbriner and Oliver Jokisch},
title={{Implementing and Evaluating Methods of Dialect Classification on Read and Spontaneous German Speech}},
year=2019,
booktitle={Proc. 8th ISCA Workshop on Speech and Language Technology in Education (SLaTE 2019)},
pages={53--58},
doi={10.21437/SLaTE.2019-10}
}