A Generative System for the Creation of New Songs from Portuguese Prosody

Depending on the prosodic choices of the reader, using various acoustic parameters to apply emphasis, the meaning of the text may change. In songs, where the text and music work together, the prosody ensures coherence between the intentions of both languages, reducing possible ambiguities. This paper presents a generative system for automatic composition of melodies from lyrics using the prosody of the Portuguese language. This approach is divided into two principal stages. First, the prosody information is extracted from the reading of the lyrics and the captured audio is aligned with the text. The inflection lines from the expressive intonation by the reader associated with the temporal alignment is used to generate a probabilistic set of note transitions. A chain of constraints is applied sequentially in order to define musical scale, harmonic field, prosody and stylistic properties. The second stage focuses on the rhythmical structure. A set of rules based on the Portuguese prosody is used to generate compatible rhythm structures along the text. These two stages generate songs ensuring the correct prosody, which is the main goal of the system. Tests were performed to evaluate this approach with aim to evaluate the capability of the algorithm to generate songs without prosody errors, and also to evaluate how pleasant the songs generated by the machine are, in comparison with songs generated by human composers using the same set of text. The results showed that the proposed system can create pleasant songs while keeping a low number of prosody errors.