5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

Energy Contour Generation for a Sentence Using a Neural Network Learning Method

Jungchul Lee (1), Donggyu Kang (1), Sanghoon Kim (1), Koengmo Sung (2)

(1) ETRI, Korea
(2) Department of Electronic Engineering, Seoul National University, Korea

Energy contour in a sentence is one of major factors that affect the naturalness of synthetic speech. In this paper. we propose a method to control the energy contour for the enhancement in the naturalness of Korean synthetic speech. Our algorithm adopts syllable as a basic unit and predicts the peak amplitude for each syllable in a word using a neural network (NN). We utilize indirect linguistic features as well as acoustic features of phonemes as input data to the NN to accommodate the grammatical effects of words in a sentence. The simulation results show that prediction error is less than 10% and our algorithm is very effective for analysis/synthesis of the energy contour of a sentence.. and generates a fairly good declarative contour for TTS.

Full Paper

Bibliographic reference.  Lee, Jungchul / Kang, Donggyu / Kim, Sanghoon / Sung, Koengmo (1998): "Energy contour generation for a sentence using a neural network learning method", In ICSLP-1998, paper 0404.