Modeling of speech parameter sequence considering global variance for HMM-based speech synthesis