Toward invariant functional representations of variable surface fundamental frequency contours: Synthesizing speech melody via model-based stochastic learning

• Authors: Yi Xu;Santitham Prom-On
• Venue: Speech Communication
• Year: 2014

Variability has been one of the major challenges for both theoretical understanding and computer synthesis of speech prosody. In this paper we show that economical representation of variability is the key to effective modeling of prosody. Specifically, we report the development of PENTAtrainer-A trainable yet deterministic prosody synthesizer based on an articulatory-functional view of speech. We show with testing results on Thai, Mandarin and English that it is possible to achieve high-accuracy predictive synthesis of fundamental frequency contours with very small sets of parameters obtained through stochastic learning from real speech data. The first key component of this system is syllable-synchronized sequential target approximation-implemented as the qTA model, which is designed to simulate, for each tonal unit, a wide range of contextual variability with a single invariant target. The second key component is the automatic learning of function-specific targets through stochastic global optimization, guided by a layered pseudo-hierarchical functional annotation scheme, which requires the manual labeling of only the temporal domains of the functional units. The results in terms of synthesis accuracy demonstrate that effective modeling of the contextual variability is the key also to effective modeling of function-related variability. Additionally, we show that, being both theory-based and trainable (hence data-driven), computational systems like PENTAtrainer can serve as an effective modeling tool in basic research, with which the level of falsifiability in theory testing can be raised, and also a closer link between basic and applied research in speech science can be developed.