Speech to Singing

The current project is a follow-up work to a project who resulted in a system that automatically converts speech segments into singing segments. Our aim is to improve the performance of the previous work in terms of the running time, the quality of speech segmentation into syllables and the naturalness of the result. We mainly used a paper that deals with converting speech to singing by decomposing the signal into parametric components using a vocoder and synchronizing the speech signal to the melody of the singing signal. In this work we are using a faster vocoder that significantly shortens the running time compared to the previous project. We tried using a different syllable segmentation algorithm in order to segment the speech more accurately into syllables, but without success. In addition, several attempts have been made to use datasets of speech and singing recordings to identify features that will help to achieve more natural singing, however we haven’t drawn any substantial conclusions in this regard. We also supply a MATLAB standalone user interface that enables to modify a large number of parameters related to the conversion process, thus allowing some control over the final result of the system.


Elad Keidan