Confidence measures for tandem connectionist feature extraction

This report proposes and compares a number of tandem-like feature extraction schemes. The proposed schemes use relative phone posteriors as condence measures estimated from the MLP outputs directly or using Gamma function. The analysis of variances shows that the proposed tandem-like features discriminate better between phone classes than the conventional tandem features. But these capabilities are lost when the complexity of the model (number of gaussians) increases. Evaluations are conducted on TIMIT database.