Abstract:
The durations of phonemes varies for different speakers. To this end, the correlations between phonemes across different speakers are studied and a novel approach to predict unknown phoneme durations from the values of known phoneme durations for a particular speaker are presented, based on the maximum likelihood criterion. Several interesting patterns are observed. Phonemes from the same broad phonetic class tend to covary most strongly (and therefore intra-class predictions of unknown phoneme durations are most accurate), but significant cross-class correlations are also present. Consequently, knowledge of only a few highly-correlated phonemes’ durations is necessary to make a good duration prediction.