Novel alignment method for DNN TTS training using HMM synthesis models

Abstract

In order to train neural networks (NN) for text-to-speech synthesis (TTS), phonetic segmentation must be performed. The most accurate segmentation is performed manually, but the process of creating manual alignments is costly and time-consuming, so automatic procedures are preferable. In this paper, a simple alignment method based on models trained during hidden Markov Model (HMM) based TTS system training is presented. It is shown that this approach slightly outperforms standard alignment procedures based on monophone models. Both objective measurements, as well as listening tests, show that NN trained with alignments obtained with the proposed method, can produce speech of higher quality compared to NN trained with monophone alignments.