Advanced LPC techniques of voice regeneration for “Virtual Dubbing”

Some recent voice conversion techniques consider models that make use of well-known paradigms of signal processing, such as Linear Predictive Coding and spectral modelling. We propose a voice converter based on Linear Predictive Coding, in which properly trained Gaussian Mixture Models transform the encoder coefficients, accounting for the glottal characteristics of a source voice, into new coefficients which provide the decoder with information about the glottal characteristics of a target voice. This voice conversion procedure results in a filter block diagram suitable for real time implementation, whose parameters can be accommodated depending on the performances of the DSP hardware at hand. A Simulink® model of the voice converter that can be directly translated into DSP code is presented. Listening experiments are shown, reporting that both non-expert and expert subjects rated the voice converter positively.