In this paper, we propose a novel acoustic-to-articulatory mapping model based on mixture of probabilistic canonical correlation analysis (mPCCA). In PCCA, it is assumed that two different kinds of data are observed as results from different linear transforms of a common latent variable. It is expected that this variable represents a common factor which is inherent in the different domains, such as acoustic and articulatory feature spaces. mPCCA is an expansion of PCCA and it can model a much more complex structure. In mPCCA, covariance matrices of a joint probabilistic distribution of acoustic-articulatory data are structuralized reasonably by using transformation coefficients of the linear transforms. Even if the number of components in mPCCA increases, the structuralized covariance matrices can be expected to avoid over-fitting. Training and mapping processes of the mPCCA-based mapping model are reasonably derived by using the EM algorithm. Experiments using MOCHA-TIMIT show that the proposed mapping method has achieved better mapping performance than the conventional GMM-based mapping.
Cite as: Uchida, H., Saito, D., Minematsu, N. (2017) Acoustic-to-Articulatory Mapping Based on Mixture of Probabilistic Canonical Correlation Analysis. Proc. Interspeech 2017, 989-993, doi: 10.21437/Interspeech.2017-1010
@inproceedings{uchida17_interspeech,
author={Hidetsugu Uchida and Daisuke Saito and Nobuaki Minematsu},
title={{Acoustic-to-Articulatory Mapping Based on Mixture of Probabilistic Canonical Correlation Analysis}},
year=2017,
booktitle={Proc. Interspeech 2017},
pages={989--993},
doi={10.21437/Interspeech.2017-1010}
}