Structured encoding of the singing voice using prior knowledge of the musical score

Abstract

The human voice is the most difficult musical instrument to simulate convincingly. Yet a great deal of progress has been made in voice coding, the parameterization and re-synthesis of a source signal according to an assumed voice model. Source-filter models of the human voice, particularly linear predictive coding (LPC), are the basis of most low bit rate (speech) coding techniques in use today. This paper introduces a technique for coding the singing voice using LPC and prior knowledge of the musical score to aid in the process of encoding, reducing the amount of data required to represent the voice. This approach advances the singing voice closer towards a structured audio model in which musical parameters such as pitch, duration, and phonemes are represented orthogonally to the synthesis technique and can thus be modified prior to re-synthesis.