The CPQD-Unicamp system for Blizzard Challenge 2021

This paper presents the CPQD-UNICAMP text-to-speech system for Blizzard Challenge 2021. The system consists of a bilingual linguistic front-end, an acoustic model based on Tacotron2 and a Parallel Wavegan neural vocoder. A multispeaker Brazilian Portuguese dataset was added to the Blizzard 2021 dataset in order to train a bilingual acoustic model. The system was later fine-tuned with the target speaker data. Sentences were classified according to the punctuation type and a specialized model was trained for each category to better model the intonation pattern of non-declarative sentences. The Blizzard Challenge evaluation for the hub task shows that the proposed strategy achieved high naturalness, intelligibility and similarity results.