14th International Congress of Phonetic Sciences (ICPhS-14)

San Francisco, CA, USA
August 1-7, 1999

Acoustic-to-Articulatory Neural Mapping under Different Statistical Characteristics of Articulatory Pattern Vectors

H. Altun (1), K. M. Curtis (2)

RISN Group, University of Nottingham, UK
(1) Department of Electrical and Electronic Engineering, Niğde University, Turkey
(2) Department of Electrical and Electronic Engineering, University of Nottingham, UK

This paper describes a mapping problem that tests and validates the findings from our analytical analysis of neural learning [1]. In this analysis different statistical characteristics of the target pattern vectors were investigated as to their effect on learning and generalisation. The problem reported on is a difficult function approximation problem, where the parameters of an articulatory speech synthesiser are estimated. The estimation of the articulatory parameters from the acoustic domain is difficult, due to the non-linear and ill-posed nature of the relationship between the acoustic and articulatory parameters. Despite proposals employing neural networks for this task in the past, neural network mapping has not been shown to be superior to other techniques that have been used to try and solve this inversion problem. Using the theoretical results of [1] we show that NN mapping can be used successfully to improve the solution to this problem.

Reference

  1. Altun, H. and Curtis K. M. 1998. Evaluation of neural learning in a MLP NN under different statistical characteristics of target pattern vectors, (Submitted to Neural Processing Letters).

Full Paper

Bibliographic reference.  Altun, H. / Curtis, K. M. (1999): "Acoustic-to-articulatory neural mapping under different statistical characteristics of articulatory pattern vectors", In ICPhS-14, 2017-2020.