Digital singing voice synthesis using a new alternating reflection model

Abstract

Many models for computer generated singing voices have been proposed in the past and have been shown to produce a wide variety of synthesized voices. While many of these models are capable of synthesizing a particular singing voice with high musical quality, they typically are challenged with respect to naturalness, range, the ability to synthesize both male and female voices, as well as the ability to capture the identity of the singer. The analysis-by-synthesis/overlap-add (ABS/OLA) sinusoidal model has proven to be effective in producing high quality voices with manageable computational cost. It is based on the combination of a block overlap-add sinusoidal representation and an analysis-by-synthesis parameter estimation technique. ABS/OLA is flexible enough to allow for modifications such as time and pitch scaling; however, it can suffer from quality degradation under such conditions. This paper presents an analysis/synthesis model that incorporates new methods to improve synthesis. These improvements add to the naturalness and flexibility in controlling perceptually important musical characteristics.