Investigations into tandem features

This report proposes and evaluates a number of tandem feature extraction schemes. The proposed schemes use con dence measures estimated from the MLP outputs to derive tandem-like features. The analysis of variance shows that the proposed features discriminate better between phone classes than conventional tandem features. But they become less discriminant as the HMM model become more complex in term of number of gaussians. This report investigates also the use of contextual knowledge and its bene t to tandem based HMM system. We evaluate the use of context-dependent modeling techniques and the use of language model. Experimental results on TIMIT database show that, while tandem features, compared to standard MFCCs improve signi cantly the performance with context-independent models, these improvements did not generalized to context-dependent models. The same conclusion, with less e ect, could be drawn for the language model. When both context-dependent and the language model are used, all features perform almost equally. This report investigates also the capacity of tandem features to handle intrinsic variabilities. Experiments are carried out using OLLO corpus.