| Abstract | Singing voice is one of the most challenging musical instruments to model. For decades,
much research and study has been performed in order to synthesize and imitate it. The
fruit of this labor was a large number of synthesizers. Among all these approaches, the
sequential concatenation of samples from a database is the one that had most success.
However, it has some weak points regard to the flexibility and expressivity. In addition,
current systems designed for voice samples transformation, such as: transposition, timescaling,
or resampling, among others, may cause a decrease in sound quality, and they do
not allow lyrics modification.
The goal of this thesis is to model, synthesize and transform Vocal Riffs: short rhythmic,
melodic, or harmonic vocal figures charged of personality that provides the singer taste
to a song. The modeling process extracts high level features from the target Vocal Riff
and stores them as a template in a database. This template can be directly synthesized or
transformed in order to create new Vocal Riffs. The software aspires to not lose sound
quality during the transformation process, since the transformation is applied to the features
in the template, instead of being applied to the sample. Furthermore, this fact allows
other singing voice aspects to be modified, such as the lyrics, and provides flexibility to
the usage of Vocal Riffs.
|
