Third International Conference on Spoken Language Processing (ICSLP 94)
In this paper we describe our experiments to evaluate the performance of a Time-Delay Neural Network recognizing binary phonetic features. We show that the error is dependent on the number of occurrence of the features in the test set and therefore must be normalized by the frequencies of the features. To get a more objective measure of the network performance, we propose the normalized mutual information calculated between the targets and the network outputs and we show that these two measures are equivalent. By evaluating the mutual information we can compare the different error rates of the features and we show that the network is a good classificator for the features with an error rate between 1% and 10%. Furthermore we observe, that the phonetic features which describe the kind of articulation are easier to recognize by the network than the features which describe the place of articulation.
Bibliographic reference. Okawa, Shigeki / Windheuser, Christoph / Bimbot, Frédéric / Shirai, Katsuhiko (1994): "Evaluation of phonetic feature recognition with a time-delay neural network", In ICSLP-1994, 1531-1534.