Two-Stage Duration Model for Indian Languages Using Neural Networks

In this paper we propose a two-stage duration model using neural networks for predicting the duration of syllables in Indian languages. The proposed model consists of three feedforward neural networks for predicting the duration of syllable in specific intervals and a syllable classifier, which has to predict the probability that a given syllable falls into an interval. Autoassociative neural network models and support vector machines are explored for syllable classification. Syllable duration prediction and analysis is performed on broadcast news data in Hindi, Telugu and Tamil. The input to the neural network consists of a set of phonological, positional and contextual features extracted from the text. From the studies it is found that about 80% of the syllable durations are predicted within a deviation of 25%. The performance of the duration model is evaluated using objective measures such as mean absolute error (μ), standard deviation (σ) and correlation coefficient (γ).