Comparison of Fujisaki-model parameters between German Learners and native speakers of Mandarin

Abstract:

The current study reports on the continued activities to develop a computer-aided phonetic learning system for German learners of Mandarin. Two corpora were used in the current study. The first one consists of recordings from German learners of Mandarin. It was used to adapt the Hidden Markov Models (HMM) of phone and tone recognizers. The second corpus consists of recordings from German students and native speakers of Mandarin. The phone and syllable labels as well as the F0 contour and Fujisaki-model parameters were manually corrected for the second corpus. The differences in the amplitude and duration of tone commands of the Fujisaki-model at the beginning and at the end of sentences between German and Chinese data for the four Mandarin tones were compared. The correctness of the phone and tone recognition systems with adapted models is better than with the original models. The duration of tone commands at the end of sentences is greater than at the beginning of sentences for both German and Chinese data and for all tones. The amplitude of tone commands of German data is greater than in the Chinese data. It was also found that for most utterances phrase commands of magnitude greater zero occurred, indicating that the phrase component should be taken into account when analyzing and synthesizing F0 contour of Mandarin.


Year: 2011
In session: Poster zu verschiedenen Themenbereichen
Pages: 146 to 153