A recent development in feature extraction is the use of neural network feature extractors, where the parameterized signal is passed through a neural network trained to discriminate between targets representing e.g. different phone classes or speakers. While the transformed feature representation often enhances class discriminability and thereby overall performance, the transformation performed by the network cannot directly be interpreted by human experts. However, explicit knowledge about this transformation could lead to the definition of a simpler function on the input features which might eventually be incorporated into the basic parameterization method. In this paper we investigate a rule extraction technique for transforming the trained network into a set of if-then rules capable of representing the transformation in a more transparent way, and apply it to the problem of distinguishing between the English fricative classes /f,v/ and /s,z/ from the TIMIT and OGI Numbers95 speech corpora.
Cite as: Kirchhoff, K. (2000) Speech analysis by rule extraction from trained artificial neural networks. Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000), vol. 2, 1077-1080, doi: 10.21437/ICSLP.2000-460
@inproceedings{kirchhoff00_icslp,
author={Katrin Kirchhoff},
title={{Speech analysis by rule extraction from trained artificial neural networks}},
year=2000,
booktitle={Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000)},
pages={vol. 2, 1077-1080},
doi={10.21437/ICSLP.2000-460}
}