K-subspaces and time-delay autoassociators for phoneme recognition

This paper presents a new approach using time-delay autoassociators (TDAA) to perform phoneme recognition. The time-delay autoassociator combines the time-delay design for phoneme recognition and the technique of multilayer perceptron autoassociators. Each time-delay autoassociator is constructed and trained to model one and only one phoneme using data belonging to that phoneme category. This non-classification training procedure provides a method with high recognition performance to avoid the drawback encountered in most conventional speech recognition neural networks that the network output values do not represent candidate likelihoods. The approach with the proposed architecture, K-subspaces with linear time-delay autoassociators, in which each phoneme is modelled by K linear TDAAs, has yielded a high recognition performance compared to that of a time delay neural net and a shift-tolerant LVQ trained by classification learning procedures, over the three difficult phonemes "B", "D" and "G". It has also been observed that the nonlinear time-delay autoassociators could perform better than linear ones.