# End-To-End Speech Synthesis Applied to Brazilian Portuguese

**Authors:** Edresson Casanova, Arnaldo Candido Junior, Frederico Santos de Oliveira, Christopher Shulby, João Paulo Teixeira, Moacir Antonelli Ponti, Sandra Maria Aluisio  
**arXiv:** [2005.05144](https://arxiv.org/abs/2005.05144)  
**Date:** 2020-05-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2005.05144v1/

## Key points

- Voice synthesis systems are popular in various applications
- Voice technology provides a natural way for human-computer interaction
- Not all languages have the same level of resources and systems available for voice synthesis
- A team of researchers created publicly available resources for Brazilian Portuguese language in the form of a dataset and deep learning models for end-to-end voice synthesis
- The dataset used in this study consists of 10.5 hours from a single speaker
- The researchers investigated three different architectures to perform end-to-end speech synthesis: Tacotron 1, DCTTS and Mozilla TTS
- They also analyzed the performance of models according to different vocoders (RTISI-LA, WaveRNN and Universal WaveRNN), phonetic transcriptions usage, transfer learning (from English) and denoising
- In the proposed scenario, a model based on Mozilla TTS and RTISI-LA vocoder presented the best performance with a 4.03 MOS value.
- Transfer learning from English data, phonetic transcriptions usage and denoising are useful techniques to train models over the presented dataset.
- The obtained results are comparable to related works covering English even though they used a smaller dataset.
- This research contributes significantly to improving voice technology's accessibility by providing publicly available resources for Brazilian Portuguese language that can be utilized in various applications such as personal assistants or accessibility tools.

## Summary

Voice synthesis systems have become increasingly popular in various applications, including personal assistants, GPS applications, screen readers and accessibility tools. Voice technology provides a natural way for human-computer interaction. However, not all languages have the same level of resources and systems available for voice synthesis. In this study, a team of researchers created publicly available resources for Brazilian Portuguese language in the form of a dataset and deep learning models for end-to-end voice synthesis. The dataset used in this study consists of 10.5 hours from a single speaker. The researchers investigated three different architectures to perform end-to-end speech synthesis: Tacotron 1, DCTTS and Mozilla TTS. They also analyzed the performance of models according to different vocoders (RTISI-LA, WaveRNN and Universal WaveRNN), phonetic transcriptions usage, transfer learning (from English) and denoising. In the proposed scenario, a model based on Mozilla TTS and RTISI-LA vocoder presented the best performance with a 4.03 MOS value. The researchers also found that transfer learning from English data, phonetic transcriptions usage and denoising are useful techniques to train models over the presented dataset. The obtained results are comparable to related works covering English even though they used a smaller dataset. This study's findings demonstrate that it is possible to create high-quality voice synthesis systems for Brazilian Portuguese using deep learning techniques with limited data availability. Overall, this research contributes significantly to improving voice technology's accessibility by providing publicly available resources for Brazilian Portuguese language that can be utilized in various applications such as personal assistants or accessibility tools.