The GlottHMM Entry for Blizzard Challenge 2011: Utilizing Source Unit Selection in HMM-Based Speech Synthesis for Improved Excitation Generation

This paper describes the GlottHMM speech synthesis system for Blizzard Challenge 2011. GlottHMM is a hidden Markov model (HMM) based speech synthesis system that utilizes glottal inverse filtering for separating the vocal tract and the glottal source from speech signal and models both components individually. In this year’s entry, stabilized weighted linear prediction (SWLP) is used to yield more robust estimates of the vocal tract filter of the high-pitched female voice. After the inverse filtering, the resulting source signal is parameterized into excitation features and a glottal flow pulse library, consisting of the variety of different glottal flow pulses. In the synthesis stage, a unit selection scheme is used for reconstructing the source signal: by minimizing the target and concatenation costs, best matching glottal flow pulses are selected from the pulse library in order to create a natural voice source. Finally, speech is synthesized by filtering the excitation signal by the vocal tract filter.