In this paper the weight decay concept known from neural network theory is applied to the two modules involved in prosody generation within our text-to-speech system Papageno. Both modules are based on neural networks (NN). Preprocessing layers are inserted connected to the inputs of specialized NN architectures via diagonal weight matrices. The weight decay concept is applied to the weights of these diagonal weight matrices. This allows an importance analysis of the used input parameters in the context of the ised NN architectures.
In the symbolic prosody module the importabnce for phrase break prediction of part-of-speech (POS) tags could be evaluated. Further, the necessary length of a POS context window could be analyzed and optimized.
For f0 generation for Mandarin language an importance analysis of the phonological information could be performed. The importance analysis led to an optimized input feature set, reducing the squared error of the used NN architecture.
Cite as: Müller, A.F., Tao, J., Hoffmann, R. (2000) Data-driven importance analysis of linguistic and phonetic information. Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000), vol. 1, 66-69, doi: 10.21437/ICSLP.2000-16
@inproceedings{muller00_icslp,
author={Achim F. Müller and Jianhua Tao and Rüdiger Hoffmann},
title={{Data-driven importance analysis of linguistic and phonetic information}},
year=2000,
booktitle={Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000)},
pages={vol. 1, 66-69},
doi={10.21437/ICSLP.2000-16}
}