Architectures for Speech Synthesis from Human Voice Audio Database

This paper describes some of the ongoing work carried out within the NLP group of Otago University for speech synthesis from diphone audio databases which are prepared from human voice recording. A speech synthesis system based on such databases claims to have human level performance provided enough descriptions are given to extract and manipulate voice data. Based on the software provided by the same project developing the databases [1][2][3], many control parameters are accommodated for preparing for a text file as a directive for speech synthesis. Thus, for a speech system with detailed phonetic descriptions stored in dictionary entries, synthesizing speech for words is a fairly straightforward task. Our work however aims at building a text-to-speech transducer without a dictionary, so that no limitations can be put to the kind of words it can process. Instead, we use a neural network as a learning system for storing text-to-speech correspondence knowledge for the task. In this paper we will introduce the training parameters and encoding method used for high-quality speech synthesizing and a text-to-speech system as a whole. In particular, we will introduce work currently being carried out for synthesizing speech of French words, based on one of the voice databases provided by the MBROLA project [3].