Hmm-based Classification of Glottalization Phenomena in German-accented English

The present paper investigates the automatic detection of word-initial glottalization phenomena (glottal stops and creaky voice) in German-accented English by means of HMMs. Glottalization of word-initial vowels can be very frequent in German-accented English, as well as in German. Detection and classification of glottalization phenomena is useful in order to obtain a pre-segmentation of speech for basic phonetic research. Moreover, the inclusion of acoustic models for glottalization, which are good indicators of word and morpheme boundaries, can improve the overall performance of an Automatic Speech Recognition (ASR) system. Since creaky voiced vowels have different spectro-temporal acoustic characteristics than modal voiced vowels, it is possible to train HMM models for creaky voice. A corpus of BBC news bulletins read by German speakers was recorded and a subset of the corpus with manually labeled glottalization at word-initial vowels was used to initialize HMM models. These models were used to label the training and the test set in the corpus. To analyze detection accuracy, the automatically derived labels were compared with the manual labels. Results show that the proposed framework can be used for the identification of glottalization phenomena, providing a reliable automatic pre-segmentation for the purpose of phonetic basic research.