Minimal classification error optimization for a speaker mapping neural network
Masashi Sugiyama, Kentaro Kurinami · 2003
The authors prepose a novel optimization technique for speaker mapping neural network training using the minimal classification error criterion. The conventional speaker mapping neural networks were trained under minimal distortion criteria. The minimal classification error optimization technique is applied to train the speaker mapping neural network. The authors describe the speaker mapping neural network and the minimal classification error optimization technique, and formulate and derive the minimal classification optimization technique in the speaker mapping neural network and a novel backpropagation algorithm. Vowel classification experiments are carried out, showing the effectiveness of the proposed algorithm. Experiments on speaker mapping with five vowels were performed and achieved a classification accuracy of 99.6% for training data and 97.4% for test data.>