Previous research has shown that parameters of the quantitative Target Approximation model (qTA) proposed by Prom-on and Xu can be directly extracted from natural speech with high accuracy through analysis-by-synthesis implemented in PENTAtrainers. While this may raise the possibility that PENTA-trainers actually simulate natural acquisition of prosody production, it is questionable that the human brain actually replicates the full articulatory mechanics represented by qTA in order to learn and control prosody production. In this paper we explore if a much simpler function can be used to extract at least some of the qTA parameters. We first managed to reduce the number of qTA parameters from three to two by evaluating their relative sensitivity. We then tested a pursuit function that learns only pitch target height and slope. Using a corpus of Mandarin utterances varying in lexical tone and focus, we show that parameters learned by the pursuit function can be used in qTA synthesis to generate F0 contours closely resembling those generated with parameters learned with qTA-based analysis-by-synthesis, with the advantage of having a much simpler learning algorithm. These results suggest that it is possible to learn articulatory control parameters for prosody without fully replicating the mechanical process itself.
Cite as: Liu, H., Xu, Y. (2014) A Simplified Method of Learning Underlying Articulatory Pitch Target. Proc. Speech Prosody 2014, 1017-1021, doi: 10.21437/SpeechProsody.2014-193
@inproceedings{liu14d_speechprosody,
author={Hao Liu and Yi Xu},
title={{A Simplified Method of Learning Underlying Articulatory Pitch Target}},
year=2014,
booktitle={Proc. Speech Prosody 2014},
pages={1017--1021},
doi={10.21437/SpeechProsody.2014-193}
}