Improved phoneme segmentation of German-accented English by means of lexicon and acoustic model adaptation

In the present study, a German ASR system was used to perform phoneme segmentation of German-accented English speech. The phoneme models were created on German training data and the used lexicon consisted of English words whose pronunciation was represented by means of the German phoneme inventory. The production of accurate segmentation is significantly affected by the language mismatch between the German training data and the German-accented English test data. In order to reduce this mismatch, enhancement of the lexicon and of the phoneme models was performed. The lexicon was enhanced by means of pronunciation rules for German-accented English and according to recognition results analysis. Acoustic model adaptation was carried out to reduce mismatch regarding language and recording differences between training and test data. Lexicon enhancement and acoustic model adaptation improved recognition accuracy providing a reliable phoneme and word segmentation framework.