Conversão Texto-Fala para o Português Brasileiro Utilizando Tacotron 2 com Vocoder Griffin-Lim

Rodrigo K Rosa, Danilo Silva

DOI: 10.14209/sbrt.2021.1570727280

Evento: XXXIX Simpósio Brasileiro de Telecomunicações e Processamento de Sinais (SBrT2021)

Keywords: Redes neurais Síntese de voz Tacotron 2 Português brasileiro

Abstract

This paper presents the training of a state-of-the-art neural network model, Tacotron-2, using a open-source voice dataset from the Common Voice project. Results from training the model from scratch and by applying transfer learning of a pre-trained english model were evaluated. The results show that it is possible to train the model with limited data resources and the model trained from scratch had less synthesis errors.

Download