Comparison between rule-based and data-driven natural language processing algorithms for Brazilian Portuguese speech synthesis

Abstract

Due to the exponential growth in the use of computers, personal digital assistants and smartphones, the development of Text-to-Speech (TTS) systems have become highly demanded during the last years. An important part of these systems is the Text Analysis block, that converts the input text into linguistic specifications that are going to be used to generate the final speech waveform. The Natural Language Processing algorithms presented in this block are crucial to the quality of the speech generated by synthesizers. These algorithms are responsible for important tasks such as Grapheme-to-Phoneme Conversion, Syllabification and Stress Determination. For Brazilian Portuguese (BP), solutions for the algorithms presented in the Text Analysis block have been focused in rule-based approaches. These algorithms perform well for BP but have many disadvantages. On the other hand, there is still no research to evaluate and analyze the performance of data-driven approaches that reach state-of-the-art results for complex languages, such as English. So, in this work, we compare different data-driven approaches and rule-based approaches for NLP algorithms presented in a TTS system. Moreover, we propose, as a novel application, the use of Sequence-to-Sequence models as solution for the Syllabification and Stress Determination problems. As a brief summary of the results obtained, we show that data-driven algorithms can achieve state-of-the-art performance for the NLP algorithms presented in the Text Analysis block of a BP TTS system.