Towards a Dialect Classification in German Speech Samples

Abstract

The automatic classification of a speaker’s dialect can enrich many applications, e.g. in the human-machine interaction (HMI) or natural language processing (NLP) but also in specific areas such as pronunciation tutoring, forensic analysis or personalization of call-center talks. Although a lot of HMI/NLP-related research has been dedicated to different tasks in affective computing, emotion recognition, semantic understanding and other advanced topics, there seems to be a lack of methods for an automated dialect analysis that is not based on transcriptions, in particular for some languages like German. For other languages such as English, Mandarin or Arabic, a multitude of feature combinations and classification methods has been tried already, which provides a starting point for our study. We describe selected experiments to train suitable classifiers on German dialect varieties in the corpus “Regional Variants of German 1” (RVG1). Our article starts with a systematic choice of appropriate spectral features. In a second step, these features are post-processed with different methods and used to train one Gaussian Mixture Model (GMM) per feature combination as a Universal Background Model (UBM). The resulting UBMs are then adapted to a varied selection of dialects by maximum-a-posteriori (MAP) adaptation. Our preliminary results on German show, that a dialect discrimination and classification is possible. The unweighted recognition accuracy ranges from 32.4 to 54.9% in a 3-dialects test and from 19.6 to 31.4% in a classification of 9-dialects. Some dialects are easier distinguishable, purely using spectral features, while others require a different feature set or more sophisticated classification methods, which we will explore in future experiments.