Approximation Neural Network for Phoneme Synthesis

摘要

The paper presents a dynamic method for phoneme synthesis using an elemental-based concatenation approach. The vocal sound waveform can be decomposed into elemental patterns that have slight modifications of the shape as they chain one after another in time but keep the same dynamics which is specific to each phoneme. An approximation or RBF network is used to generate elementals in time with the possibility of controlling the characteristics of the sound signals. Based on this technique a quite realistic mimic of a natural sound was obtained.