Diphone Embedding, a Step Towards Neural Network Speech Synthesis

Abstract

Aiming for Neural Network Speech Synthesis system. The work of this thesis consists of building a Diphone embedding based on the auto-encoder model. To achieve that, we first prepared a database of Diphone sounds and their linguistic features. Then we trained the encoder part of the model. In which we took the Diphone sounds and their acoustic characteristics as inputs to generate an embedded form linked to the linguistic features. The evaluation of the trained model and it tests gave acceptable results with a correct prediction rate that resch 80%.