Duration Modeling by Multi-Models based on Vowel Production characteristics

An accurate estimation of segmental durations is needed for natural sounding textto-speech (TTS) synthesis. This paper propose multi-models based on production aspects of vowels. In this work four multi-models are developed based on vowel length, vowel height, vowel frontness and vowel roundness. In each multimodel, syllables are divided into groups based on specific vowel articulation characteristics. In this study, (i) linguistic constraints represented by positional, contextual and phonological features and (ii) production constraints represented by articulatory features are used for predicting duration patterns. Feed-forward Neural Networks are used for developing duration models. From the results, it was observed that the average prediction error is reduced by 23.21% and correlation coefficient is improved by 9.64% using multi-model developed based on vowel length production characteristics, compared to single duration model.