To create the dadaset was used public domain texts.Initially, the texts were extracted from Wikipedia articles displayed in the Highlights section. In a second phase, texts were also extracted from Chatterbot-corpus, a corpus originally created for the construction of chatbots. We also used 20 sets of phonetically balanced phrases, each set containing 10 phrases proposed by Seara (1994). The total number of words is 71,358, with 13,311 distinct words.
The dataset developed in this work has approximately 10 hours and 28 minutes of speech from a single speaker, recorded at 48Khz, containing a total of 3,632 audio files in Wave format. Audio files range from 0.67 to 50.08 seconds.
Since the audios were not recorded in an acoustic studio, there is noise present in the audio files, so we chose to use a noise suppression library. For this purpose we used the library RNNoise . It is based on Recurrent Neural Networks, more specifically Gated Recurrent Unit (GRU) and demonstrated good performance for noise suppression. The dataset is open source, and public available under the terms of the license Creative Commons Attribution 4.0 (CC BY 4.0).
Download this Dataset: TTS-Portuguese Corpus
Models trained in this corpus:
| Model | Colab Demo URL |
|---|---|
| DCTTS model (Experiment 1) | https://drive.google.com/file/d/1tlmk4DCtYgvUUcJgOPHXPw9o2ysnbVbg/view |
| DCTTS+WaveRNN model | https://colab.research.google.com/drive/160GFTlTUWBx8UiC_mzAy4HdBO2XU8Rn5 |
| Tacotron 1 model (Experiment 2) | https://drive.google.com/file/d/1tlmk4DCtYgvUUcJgOPHXPw9o2ysnbVbg/view |
| TTS Mozilla model without TL | https://drive.google.com/file/d/19r-IDQiG1FmcmSOlIStGl6u_H3kFJQKk/view |
| TTS Mozilla model with TL | https://drive.google.com/file/d/1OZ7T5fqR-QPyN5O6HFetkOeHNCKbdGsA/view |
| TTS Mozilla model with TL + Universal WaveRNN | https://drive.google.com/file/d/1OHKfmd7uqJ7FyMYkKBFCyT0sXkLm21zv/view |
| TTS Mozilla model with TL + Universal WaveRNN with TL (Experiment 3) (The Best Model) | https://colab.research.google.com/drive/1b5Yo0sghajCrjftolN0LbArZzar7fd01?usp=sharing |
Synthesized samples of the best model with Griffin–Lim (1.wav) and WaveRNN (1-wavernn.wav) vocoder : Link
Audio samples synthesized with best models: https://edresson.github.io/TTS-Portuguese-Corpus/
Cite this Dataset:
Preprint:
@article{casanova2020end,
title={End-To-End Speech Synthesis Applied to Brazilian Portuguese},
author={Casanova, Edresson and Junior, Arnaldo Candido and Shulby, Christopher and de Oliveira, Frederico Santos and Teixeira, Jo{\~a}o Paulo and Ponti, Moacir Antonelli and Aluisio, Sandra Maria},
journal={arXiv preprint arXiv:2005.05144},
year={2020}
}
Full Paper:
coming soon