Modeling of Speech Parameter Sequence Considering Global Variance for HMM-Based Speech Synthesis

Abstract

Speech technologies such as speech recognition and speech synthesis have many potential applications since speech is the main way in which most people communicate. Various linguistic sounds are produced by controlling the configuration of oral cavities to convey a message in speech communication. The produced speech sounds temporally vary and are significantly affected by coarticulation effects. Thus, it is not straightforward to segment speech signals into corresponding linguistic symbols. Moreover, the acoustics of speech vary even if the same words are uttered by the same speaker due to differences in the manner of speaking and articulatory organs. Therefore, it is essential to stochastically model them in speech processing. The hidden Markov model (HMM) is an effective framework for modeling the acoustics of speech. Its introduction has enabled significant progress in speech and language technologies. In particular, there have been numerous efforts to develop HMM-based acoustic modeling techniques in speech recognition, and continuous density HMMs have been widely used in modern continuous speech recognition systems (Gales & Young (2008)). Moreover, several approaches have been proposed for applying the HMM-based acoustic modeling techniques to speech synthesis technologies (Donovan & Woodland (1995); Huang et al. (1996)) such as Text-to-Speech (TTS), which is ... from a given text. Recently, HMM-based speech synthesis has been proposed (Yoshimura et al. (1999)) and has generated interest owing to its various attractive features such as completely data-driven voice building, flexible voice quality control, speaker adaptation, small footprint, and so forth (Zen et al. (2009)). A basic framework of HMM-based speech synthesis consists of training and synthesis processes. In the training process, speech parameters such as spectral envelope and fundamental frequency (F0) are extracted from speech waveforms and then their time sequences are modeled by context-dependent phoneme HMMs. To model the dynamic characteristics of speech acoustics with HMMs, which assume piecewise constant statistics within an HMM state and conditional independence, a joint vector of static and dynamic features is usually used as an observation vector. In the synthesis process, a smoothly varying speech parameter trajectory is generated by maximizing the likelihood of a composite sentence HMM subject to a constraint between static and dynamic features with respect to not the observation vector sequence including both static and dynamic features but the static feature vector sequence (Tokuda et al. (2000)). Finally, a vocoding technique is employed Modeling of Speech Parameter Sequence Considering Global Variance for HMM-Based Speech Synthesis 6