Singer Identification and Transformation through Dynamic Modeling of Vocal Fold and Vocal Tract Parameters

Oftentimes when we listen to a familiar singer, the unique qualities of that performer’s voice allow us to establish the singer’s identity with relative ease. It is believed that the unique acoustic qualities of an individual singer’s voice arise from a combination of innate physical factors (e.g. vocal tract and vocal fold physiology) and individual characteristics of performance and expression (e.g. pronunciation and accent). In this research, we jointly estimate pole-zero filter and LF glottal waveform model parameters to model the shape of the vocal tract and the glottal excitation, respectively over short time periods. These time-varying parameters, corresponding to the physical characteristics of the singer, are used to train a Hidden Markov Model (HMM). The HMM is used to model the dynamic behavior of the source-filter parameters, corresponding to some of the expressive characteristics of the singer. We propose a system that is able to identify singers based upon the model of greatest likelihood among the individually trained HMMs. We also explore the use of individual HMM states as a method of mapping from the parameters of one singer to another in a preliminary attempt at singing voice transformation. The data used in this analysis was recorded from four conservatory-level classically trained singers.