Patricia: um sintetizador de canto em tempo real para o português brasileiro

Abstract

Singing voice synthesis involves the artificial generation of a song given its lyrics and musical notes through computational methods. This paper describes the development of PATRICIA, a system that performs real-time singing synthesis for Brazilian Portuguese. A systematic mapping was conducted to provide guidelines for the design and implementation of PATRICIA. Additionally, experiments were carried out to demonstrate the musical capabilities of the system and perform a comparative analysis of its performance on a personal computer and a Raspberry Pi device. The mapping resulted in the choice of a concatenative synthesis approach based on samples and text files providing the song lyrics in advance. An initial version of the synthesizer was implemented in the SuperCollider language as a proof of concept. The conducted experiments validated this proof of concept, demonstrating the feasibility of the idea, particularly in terms of performance, which showed excellent indicators on both chosen platforms. Some improvements are pointed out with the aim of making the synthesizer overcome its current limitations in future versions.