Spoken language variety recognition in Dutch
Year: 2014
Type: article
Abstract: We used the Duch Spoken Corpus CGN Corpus Gesproken Nederlands to train a spoken language variety recognition system, that is capable of discriminating between Dutch spoken in The Netherlands and Belgium. We use a Support Vector Machine classifier to recognise utterances, with a linear kernel in a high-dimensional (50176) supervector space. The supervectors are scaled shifts of the concatenated means of a Gaussian Mixture Model (GMM). The shifted means for an utterance are found by Maximum A Posteriori adaptation of a so-called Universal Background Model, a 1024-component GMM that is trained with all available training speech in both language varieties. The acoustic features used are Shifted Delta Cepstra, based on Mel Frequency Cepstral Coefficients obtained from short duration frames of the speech signal. The discrimination performance varies with test segment duration, from an Equal Error Rate of 16.1 % for 3 second segments to 3.4 % for 30 s segments.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W795082863
