Sample-based singing voice synthesizer by spectral concatenation

The singing synthesis system we present generates a performance of an artificial singer out of the musical score and the phonetic transcription of a song using a frame-based frequency domain technique. This performance mimics the real singing of a singer that has been previously recorded, analyzed and stored in a database. To synthesize such performance the systems concatenates a set of elemental synthesis units. These units are obtained by transposing and time-scaling the database samples. The concatenation of these transformed samples is performed by spreading out the spectral shape and phase discontinuities of the boundaries along a set of transition frames that surround the joint frames. The expression of the singing is applied through a Voice Model built up on top of a Spectral Peak Processing (SPP) technique.