Pitch-scaled spectrum based excitation model for HMM-based Speech Synthesis

The quality of speech generated from Hidden Markov Model (HMM)-based Speech Synthesis System (HTS) is suffered from ‘buzzing’ problem which is due to oversimplified vocoding technique. This paper proposed an excitation model to improve the parametric representation of speech in HTS. Residual got from inverse filtering keeps some detailed harmonic structure of speech which has not be included in linear prediction (LP) spectrum. Pitch-scaled spectrum can be used as a supplement of LP spectrum in speech reconstruction. This spectrum is compressed by principal component analysis (PCA) and eigenvalues are indicated as periodic parameter. Then, an aperiodic measure is also extracted from pitch-scaled spectrum and a sigmoid function is fitted to this measure as aperiodic parameter. These two parameters are integrated into HTS training as excitation parameter. Listening tests showed that this proposed technique could generate better sound than pulse train excitation model and take a comparable result with STRAIGHT.